diff --git a/tophub/arm_cpu_v0.05.log b/tophub/arm_cpu_v0.05.log
new file mode 100644
index 0000000..0e18361
--- /dev/null
+++ b/tophub/arm_cpu_v0.05.log
@@ -0,0 +1,1167 @@
+# This is the pre-tuned parameters for arm cpu backends
+# TVM downloaded this during compilation
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 58242, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0202839052], 0, 1.0239202976226807, 1535933156.928128], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2642, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001043853], 0, 0.9550290107727051, 1535867704.1470199], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3506, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010108273737373737], 0, 0.9982681274414062, 1535868029.556969], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3506, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000978110932038835], 0, 1.1417245864868164, 1535868222.7269862], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 722, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0009452437358490566], 0, 1.1576757431030273, 1535868447.4127173], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3474, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009117656090909091], 0, 0.9399948120117188, 1535868681.3287132], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3378, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008779413245614035], 0, 0.9823009967803955, 1535868930.7811167], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2802, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0008455779327731092], 0, 1.0109384059906006, 1535869158.901984], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2642, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000813743430894309], 0, 0.9680273532867432, 1535869373.5936017], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 658, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0007799986434108527], 0, 1.1193089485168457, 1535869576.8542712], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3442, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000746372125925926], 0, 0.9636001586914062, 1535869831.548129], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3346, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007106397801418439], 0, 0.9676427841186523, 1535870116.5811934], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2610, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006803197823129251], 0, 0.9989221096038818, 1535870320.5205925], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2578, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006468487419354839], 0, 0.9897785186767578, 1535870611.9659176], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1330, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006140555521472392], 0, 0.9766724109649658, 1535870819.4726229], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1330, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005779040689655172], 0, 0.9882245063781738, 1535871092.7511504], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1165, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.00039303106640625], 0, 2.5247223377227783, 1535871308.9138055], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 562, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0005475784153005464], 0, 0.9800069332122803, 1535871619.313344], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3733, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.014695421857142857], 0, 1.0570013523101807, 1535871985.0209813], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0037036578928571428], 0, 0.9887669086456299, 1535872234.0142796], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6059, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035723218928571427], 0, 0.9913361072540283, 1535872565.8719385], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0034493371379310345], 0, 1.0222351551055908, 1535872859.5420387], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2987, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003316104387096774], 0, 0.9976027011871338, 1535873104.9854105], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13739, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0031898938125], 0, 1.028520107269287, 1535873340.2650568], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2731, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0030660625151515152], 0, 1.043668508529663, 1535873633.1236413], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10283, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002946220411764706], 0, 1.0112240314483643, 1535873911.2069273], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14123, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028230205], 0, 1.0393104553222656, 1535874217.0312018], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2731, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002691451473684211], 0, 1.041308879852295, 1535874530.5183444], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0025742958205128203], 0, 0.9823927879333496, 1535874777.1529186], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002453553219512195], 0, 1.027040719985962, 1535875017.0616808], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10667, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002327847418604651], 0, 1.030226230621338, 1535875360.582989], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2091, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002206292543478261], 0, 1.042597770690918, 1535875593.9350574], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0020883826666666665], 0, 1.010864496231079, 1535876024.9933944], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13483, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019737022352941174], 0, 1.0111725330352783, 1535876337.3643184], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0018484203818181818], 0, 1.0398824214935303, 1535876658.2531588], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5419, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001731515724137931], 0, 1.0214545726776123, 1535876970.7406116], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5163, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016142949365079366], 0, 1.0311267375946045, 1535877266.747691], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2603, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0014997511791044775], 0, 1.0348572731018066, 1535877519.6159651], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14251, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013826725616438355], 0, 1.0200920104980469, 1535877922.459238], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0012700887974683545], 0, 1.0383615493774414, 1535878253.9122772], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5931, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011557881149425286], 0, 1.003911018371582, 1535878546.82445], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14251, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010447444166666667], 0, 1.028672218322754, 1535878911.2021189], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0008500219915254237], 0, 2.6243090629577637, 1535879182.8139532], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009340794537037037], 0, 1.006883144378662, 1535879456.4370754], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7239, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.015536609571428573], 0, 1.8323256969451904, 1535879766.6652172], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30299, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007610769357142858], 0, 1.0533397197723389, 1535880060.7316282], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6435, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0070827269333333335], 0, 1.0155055522918701, 1535880375.1276994], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12483, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0065695223125], 0, 1.0247323513031006, 1535880678.8808696], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11619, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006066912529411765], 0, 1.0296063423156738, 1535881041.9360034], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6683, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005583380166666667], 0, 1.0516958236694336, 1535881412.5760958], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13635, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0050572407], 0, 1.0060200691223145, 1535881714.361793], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6435, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004539602347826087], 0, 1.0365831851959229, 1535881994.584638], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12483, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00407884976], 0, 1.0319972038269043, 1535882263.9357245], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4707, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0035435204827586205], 0, 1.0374653339385986, 1535882558.479322], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0030062086176470587], 0, 1.0370540618896484, 1535882885.7239344], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6435, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002500447875], 0, 1.0481789112091064, 1535883289.2406862], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 696, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [8, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.005871429500000001], 0, 2.5013861656188965, 1535883482.7106888], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00200213674], 0, 1.0432822704315186, 1535883818.9882414], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 43651, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.017118677333333335], 0, 1.0458931922912598, 1535884085.650732], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22659, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015015905999999999], 0, 1.036301851272583, 1535884516.738676], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 41603, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.012850740125], 0, 1.0719225406646729, 1535884899.023663], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 42115, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0108603461], 0, 1.0569005012512207, 1535885176.6808953], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006940615], 0, 0.9846491813659668, 1535885703.3665733], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.014634489285714286], 0, 2.629513740539551, 1535885908.6369312], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10603, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003549577275862069], 0, 1.0621600151062012, 1535886377.7250006], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9972, "c": null, "e": [["tile_co", "sp", [56, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008167388230769231], 0, 1.05180025100708, 1535886641.8560708], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6021, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.007011787199999999], 0, 1.0558886528015137, 1535887102.1073425], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5044, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005864639277777778], 0, 1.041900396347046, 1535887521.1474288], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19563, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002194059760869565], 0, 1.0408320426940918, 1535887751.7398622], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2537, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [48, 8]], ["tile_c", "sp", [448, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0051816823], 0, 2.523982524871826, 1535888021.9531574], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18218, "c": null, "e": [["tile_co", "sp", [56, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00506597495], 0, 1.0183184146881104, 1535888354.0414436], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 15763, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0040226566000000005], 0, 1.191983938217163, 1535888821.035866], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 11651, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0042227665], 0, 1.0330696105957031, 1535889030.8109107], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27733, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004347858347826087], 0, 1.0283198356628418, 1535889391.0382214], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24938, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003630558107142857], 0, 0.9896316528320312, 1535889616.0205817], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24713, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003549182551724138], 0, 1.1433401107788086, 1535889969.8400445], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24264, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0058762527222222215], 0, 1.1726973056793213, 1535890225.7783935], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 2639, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.017287808166666665], 0, 1.0220444202423096, 1535890492.34707], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 4543, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.017289727], 0, 1.0607144832611084, 1535890696.2221918], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 2527, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.014392054142857142], 0, 1.0516717433929443, 1535890965.2784097], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 4999, "c": null, "e": [["tile_co", "sp", [8, 20]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012123700333333333], 0, 1.0816552639007568, 1535891206.7271261], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 2527, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.014408261714285714], 0, 1.0397610664367676, 1535891409.7025518], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3847, "c": null, "e": [["tile_co", "sp", [8, 20]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.012110901666666667], 0, 1.0709097385406494, 1535891745.8117652], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2023, "c": null, "e": [["tile_co", "sp", [8, 20]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008708796], 0, 0.9992084503173828, 1535891964.8166866], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 4599, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.011538894888888889], 0, 1.0478708744049072, 1535892182.4845812], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 741, "c": null, "e": [["tile_co", "sp", [4, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.007701358615384615], 0, 1.0212745666503906, 1535892391.559579], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 2415, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011528142555555556], 0, 1.0503599643707275, 1535892593.255601], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3461, "c": null, "e": [["tile_co", "sp", [4, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00769657076923077], 0, 1.0464003086090088, 1535892839.9363909], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1349, "c": null, "e": [["tile_co", "sp", [4, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0069652952], 0, 0.9851541519165039, 1535893047.414964], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 903, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.010649769], 0, 0.9399757385253906, 1535893270.0392303], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2166, "c": null, "e": [["tile_co", "sp", [8, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006042783823529412], 0, 1.0501677989959717, 1535893469.899569], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7113, "c": null, "e": [["tile_co", "sp", [12, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0615500205], 0, 1.47709321975708, 1535893795.3775012], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3083, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00405375552], 0, 0.9996118545532227, 1535894057.1592999], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4874, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004671418136363636], 0, 1.0421631336212158, 1535894366.955129], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2763, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003585902071428571], 0, 1.0156898498535156, 1535894627.9069421], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9690, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0041580177200000006], 0, 1.0501728057861328, 1535894891.9881763], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10706, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001764466192982456], 0, 1.0366666316986084, 1535895250.6422956], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2920, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [48, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0100516841], 0, 2.6344704627990723, 1535895405.7927814], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1840, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.007711641384615385], 0, 2.6450724601745605, 1535895746.3718283], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 9774, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.013986426], 0, 1.2428455352783203, 1535896038.9707313], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6903, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [5, 7]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025128983], 0, 1.034337043762207, 1535896288.076893], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12378, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003096685242424242], 0, 1.0122265815734863, 1535896565.5381186], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4767, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [71, 1]], ["tile_ow", "sp", [71, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.16616388325], 0, 2.945256233215332, 1535896782.4115124], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3487, "c": null, "e": [["tile_co", "sp", [4, 20]], ["tile_oh", "sp", [73, 1]], ["tile_ow", "sp", [73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007888369384615384], 0, 0.9553723335266113, 1535897234.9925163], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 21002, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [21, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.069287843], 0, 1.673440933227539, 1535897615.5549154], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3530, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [21, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0376812965], 0, 2.1774871349334717, 1535897978.9854457], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2019, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [149, 1]], ["tile_ow", "sp", [149, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007448719785714286], 0, 1.049992322921753, 1535898247.2696643], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 342, "c": null, "e": [["tile_co", "sp", [50, 20]], ["tile_oh", "sp", [1, 1]], ["tile_ow", "sp", [1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00032918867763157895], 0, 0.9231925010681152, 1535898429.6051092], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7526, "c": null, "e": [["tile_co", "sp", [320, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003331550806451613], 0, 1.004594087600708, 1535898671.3261254], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2270, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0024626470975609755], 0, 1.016362190246582, 1535898954.1860723], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5114, "c": null, "e": [["tile_co", "sp", [40, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012502429875], 0, 0.9685764312744141, 1535899241.719852], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1513, "c": null, "e": [["tile_c", "sp", [480, 2]], ["tile_h", "sp", [7, 1]], ["tile_w", "sp", [1, 7]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00022876097972972974], 0, 1.0879342555999756, 1535899492.7930477], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9355, "c": null, "e": [["tile_co", "sp", [240, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0012520333375], 0, 1.0149590969085693, 1535899730.144996], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 794, "c": null, "e": [["tile_co", "sp", [40, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0007608128863636364], 0, 0.9922950267791748, 1535899996.440168], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 864, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00022294626948775055], 0, 1.0736589431762695, 1535900210.7079601], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20609, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015832748125], 0, 1.032682180404663, 1535900435.968286], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 696, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0006806117315436242], 0, 1.1230158805847168, 1535900674.445971], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27998, "c": null, "e": [["tile_co", "sp", [72, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001633821612903226], 0, 1.0263967514038086, 1535900935.2902656], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16769, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0010496381875], 0, 1.0584471225738525, 1535901279.1037664], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12358, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007143568714285714], 0, 1.025890588760376, 1535901678.683801], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1299, "c": null, "e": [["tile_c", "sp", [96, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00037633076779026216], 0, 1.0672800540924072, 1535901898.3627403], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11605, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007452176888888889], 0, 1.0416312217712402, 1535902196.3357546], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2278, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0003696600701107011], 0, 1.039217233657837, 1535902533.799068], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 479, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0002685563946666667], 0, 1.081014633178711, 1535902792.8906808], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9363, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008268132459016394], 0, 1.0285654067993164, 1535903220.2973185], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6665, "c": null, "e": [["tile_c", "sp", [96, 2]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010983254130434784], 0, 1.0487258434295654, 1535903441.3031201], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 40927, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0008924621592920354], 0, 1.0460529327392578, 1535903858.48559], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4611, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0006109505182926829], 0, 1.013946294784546, 1535904118.483186], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1098, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [28, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0010525962708333332], 0, 1.0185754299163818, 1535904348.1885536], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 65100, "c": null, "e": [["tile_co", "sp", [4, 6]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0029780487647058825], 0, 1.046910285949707, 1535904632.6540232], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1938, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0039843320769230774], 0, 1.0484378337860107, 1535904867.0949383], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21483, "c": null, "e": [["tile_co", "sp", [36, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0023646463023255814], 0, 1.0329222679138184, 1535905277.0900674], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 79890, "c": null, "e": [["tile_co", "sp", [8, 3]], ["tile_oh", "sp", [14, 4]], ["tile_ow", "sp", [56, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0025665208974358974], 0, 0.9776053428649902, 1535905556.2492056], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1551, "c": null, "e": [["tile_c", "sp", [24, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0036118554642857142], 0, 1.0654077529907227, 1535905799.034823], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 126245, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0068189286666666665], 0, 1.0377330780029297, 1535905956.599957], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 40652, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [16, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0025650832564102564], 0, 1.035742998123169, 1535906283.2929611], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 51123, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0036906727142857142], 0, 1.0533051490783691, 1535906614.9176865], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1960, "c": null, "e": [["tile_co", "sp", [256, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008341790333333333], 0, 1.014880895614624, 1535906931.9357674], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 595, "c": null, "e": [["tile_c", "sp", [512, 2]], ["tile_h", "sp", [7, 1]], ["tile_w", "sp", [1, 7]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002594050465116279], 0, 1.0943217277526855, 1535907208.5442753], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4776, "c": null, "e": [["tile_co", "sp", [256, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004218789333333333], 0, 1.0163333415985107, 1535907597.0871272], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 532, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00019728865680473373], 0, 1.0869140625, 1535907809.6090517], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16693, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007228319285714285], 0, 1.0497119426727295, 1535908108.9082053], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 332, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0005741038114285714], 0, 1.1270930767059326, 1535908383.1919243], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6773, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0036699419285714284], 0, 1.0549159049987793, 1535908701.4121802], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1460, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00036298671582733814], 0, 1.1341958045959473, 1535908956.8284526], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 35751, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007914603846153846], 0, 1.0619268417358398, 1535909176.9554489], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4285, "c": null, "e": [["tile_c", "sp", [128, 2]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016527744918032787], 0, 1.0701408386230469, 1535909421.7176766], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14043, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003930022692307692], 0, 0.9923868179321289, 1535909734.4979897], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3803, "c": null, "e": [["tile_c", "sp", [16, 8]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008821893771929825], 0, 1.1276097297668457, 1535909991.8039658], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 44675, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008854019999999999], 0, 1.110100507736206, 1535910359.75753], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1034, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0035849489285714286], 0, 1.0848300457000732, 1535910618.8165505], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004863413857142857], 0, 1.0441834926605225, 1535910898.740208], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2249, "c": null, "e": [["tile_c", "sp", [16, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002456499731707317], 0, 1.0797441005706787, 1535911146.5250747], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12043, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00649938825], 0, 1.0292408466339111, 1535911470.8550332], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3744, "c": null, "e": [["tile_c", "sp", [32, 1]], ["tile_h", "sp", [14, 8]], ["tile_w", "sp", [28, 4]], ["ann", "an", ["none", "vec", "none"]]], "t": "direct"}], "r": [[0.003867426769230769], 0, 1.8320848941802979, 1535911851.908701], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 52322, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003233817612903226], 0, 1.096975326538086, 1535912133.38638], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3112, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010764590215053764], 0, 0.9866111278533936, 1535912406.3411784], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4366, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000903172117117117], 0, 0.9732120037078857, 1535912606.0711665], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1037, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0029729099117647055], 0, 0.9995296001434326, 1535912835.8323414], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2415, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00041974330962343093], 0, 0.9404454231262207, 1535913123.5399625], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3341, "c": null, "e": [["tile_co", "sp", [24, 2]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005789384971098266], 0, 0.9929215908050537, 1535913346.1333234], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10395, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006026739156626506], 0, 0.9896507263183594, 1535913556.1993494], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3302, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009629180673076923], 0, 1.0051872730255127, 1535913849.924076], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 73082, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [14, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005848908111111111], 0, 1.091404676437378, 1535914210.3322077], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1414, "c": null, "e": [["tile_co", "sp", [50, 20]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0214223508], 0, 1.0292282104492188, 1535914474.3039367], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.002739591162162162], 0, 2.599026918411255, 1535914659.5182357], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1589, "c": null, "e": [["tile_co", "sp", [8, 32]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007175246357142857], 0, 0.9670054912567139, 1535914998.4716396], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 536, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001429850342857143], 0, 0.9885058403015137, 1535915245.6802108], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 768, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.012977006625], 0, 2.427839994430542, 1535915452.2533872], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15015, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022771903636363638], 0, 1.0337705612182617, 1535915682.8874214], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11791, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035621779655172414], 0, 1.025313377380371, 1535915930.6258304], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17806, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002812973833333333], 0, 1.0066864490509033, 1535916141.866153], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4315, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0110032041], 0, 1.0640997886657715, 1535916368.8402455], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10365, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001319706210526316], 0, 1.018779993057251, 1535916636.7979658], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13486, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001811639392857143], 0, 1.039339542388916, 1535917027.6959298], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 13475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0042561011666666666], 0, 1.8484501838684082, 1535917277.8445766], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006180451543209876], 0, 1.006117820739746, 1535917570.1862874], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7995, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0012497871481481483], 0, 1.0166749954223633, 1535917785.5692449], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11042, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.017384289833333334], 0, 1.0812103748321533, 1535918103.3693402], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13474, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028468038055555556], 0, 1.0348665714263916, 1535918474.8294525], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10202, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003321741], 0, 1.0127544403076172, 1535918718.5729113], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1542, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002041306326530612], 0, 0.9959666728973389, 1535918924.246683], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 9438, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004529989043478261], 0, 1.0545332431793213, 1535919284.8428514], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8990, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0008592897435897435], 0, 1.0198874473571777, 1535919677.1728883], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1302, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0015375277727272727], 0, 0.9964268207550049, 1535919924.7992334], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 21843, "c": null, "e": [["tile_co", "sp", [24, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.029487022], 0, 1.1587800979614258, 1535920469.8444839], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [512, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0258434165], 0, 3.5272419452667236, 1535920786.219744], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 768, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [256, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.127067439], 0, 3.7743887901306152, 1535921084.356418], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 668, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [128, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.06694648125], 0, 2.9744529724121094, 1535921454.6264045], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.104547297], 0, 3.675703525543213, 1535921913.4576726], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.06245824925], 0, 3.106839418411255, 1535922393.601476], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2648, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [196, 4]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.1452575885], 0, 6.913934707641602, 1535922825.1677697], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2648, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [196, 4]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0981105615], 0, 3.594774007797241, 1535923390.5251439], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2473, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [784, 4]], ["tile_k", "sp", [8, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.24803750975], 0, 5.54766321182251, 1535923835.5238993], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 19322, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [224, 1]], ["tile_ow", "sp", [56, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.018322846333333333], 0, 1.0951249599456787, 1535924568.7708895], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4302, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011830588588235293], 0, 1.044874906539917, 1535924852.2144072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1612, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0103697488], 0, 1.0422253608703613, 1535925297.5888944], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6384, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010551518315789472], 0, 1.0551557540893555, 1535925673.6977892], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15887, "c": null, "e": [["tile_co", "sp", [64, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0100384315], 0, 1.1225578784942627, 1535926055.9480326], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24867, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001176763058139535], 0, 1.0307064056396484, 1535926372.8655858], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24538, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0105462744], 0, 1.0737485885620117, 1535926705.1027775], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1902, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008390369], 0, 1.0141997337341309, 1535926975.5441978], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 842, "c": null, "e": [["tile_co", "sp", [512, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0084405375], 0, 0.9960105419158936, 1535927239.11576], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1946, "c": null, "e": [["tile_co", "sp", [512, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.018193530833333332], 0, 1.0996158123016357, 1535927501.5333319], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1136, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [512, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.007285116714285714], 0, 2.4856388568878174, 1535927726.1663], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3382, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004621641636363636], 0, 1.0410125255584717, 1535927983.1956265], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15024, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007462784642857143], 0, 1.0591535568237305, 1535928330.9349072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19066, "c": null, "e": [["tile_co", "sp", [128, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0073287949285714285], 0, 1.0026061534881592, 1535928673.0336595], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14480, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.017191845], 0, 1.9116895198822021, 1535928941.95602], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.007402969571428572], 0, 2.6678688526153564, 1535929255.1175828], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2920, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004410804956521739], 0, 1.0100977420806885, 1535929572.0225327], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11907, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008138508692307693], 0, 1.048060655593872, 1535930027.6689358], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30003, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007756009461538462], 0, 1.0650906562805176, 1535930374.4901955], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 37563, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.017396149333333333], 0, 1.0720045566558838, 1535930624.4455588], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 678, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [8, 16]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.012345374222222222], 0, 2.430203914642334, 1535930848.9246736], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6723, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0049223226190476185], 0, 1.9980731010437012, 1535931207.4185019], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8179, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00940216409090909], 0, 1.0163772106170654, 1535931665.7364244], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25491, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009410501181818182], 0, 1.0630242824554443, 1535932123.8683615], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [8, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.01329342475], 0, 2.635051727294922, 1535932379.1936517], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=bcm2837 -target=armv7l-linux-gnueabihf -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 47603, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025904653076923075], 0, 1.0392539501190186, 1535932934.8960934], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 60979, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0108865713], 0, 5.046119213104248, 1535998915.3983145], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2643, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000643373544871795], 0, 7.373241186141968, 1535867769.112659], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1299, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006232570124223602], 0, 4.519797086715698, 1535868141.637669], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006032632409638555], 0, 1.523380994796753, 1535868577.9077752], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2739, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005836808837209303], 0, 5.016199827194214, 1535868915.1114855], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 723, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0005626136179775281], 0, 2.3225531578063965, 1535869557.2066498], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005432204972972973], 0, 1.5021629333496094, 1535869984.736592], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005239464502617802], 0, 3.8839216232299805, 1535870335.8850505], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2739, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005027702663316583], 0, 2.341417074203491, 1535870727.543773], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3507, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00048329306763285026], 0, 3.0169031620025635, 1535871274.136661], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000460935465437788], 0, 1.564884901046753, 1535871797.4949446], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00044189919823788546], 0, 2.2022483348846436, 1535872250.3087072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1427, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004221192362869198], 0, 5.4362242221832275, 1535872606.220856], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004020528514056225], 0, 3.5564537048339844, 1535873052.4636877], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2643, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003818329503816794], 0, 3.863401174545288, 1535873456.5221703], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003615718916967509], 0, 1.5705974102020264, 1535873953.219711], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 412, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [8, 4]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.00015226633333333334], 0, 2.357700824737549, 1535874309.2256474], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 659, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0003398390711864407], 0, 0.856743335723877, 1535874744.9150379], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2804, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.009497779545454546], 0, 3.3152828216552734, 1535875197.8323314], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002411916666666667], 0, 4.946890830993652, 1535875796.5777], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6084, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002284724], 0, 3.7815210819244385, 1535876306.025284], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5219, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021940116304347826], 0, 5.469178915023804, 1535877073.5668974], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13923, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002093388833333333], 0, 4.994829416275024, 1535877848.8405266], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11066, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00201188712], 0, 4.722552061080933, 1535878499.0032315], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019432569807692308], 0, 4.482206344604492, 1535879042.3146672], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14058, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001858024574074074], 0, 5.384204149246216, 1535879880.0850754], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13882, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017594783859649122], 0, 0.8571586608886719, 1535881038.8413334], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10554, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001724831], 0, 3.464539051055908, 1535881652.617043], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10426, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0016409399193548387], 0, 5.915098190307617, 1535882526.7586374], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001569226046875], 0, 5.878464698791504, 1535883166.2701552], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001494628455882353], 0, 0.890733003616333, 1535883838.656456], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001427921281690141], 0, 0.872572660446167, 1535884500.3921692], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6115, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013538966216216216], 0, 0.858823299407959, 1535885046.0352826], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14186, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012763736962025315], 0, 2.7870748043060303, 1535885853.038236], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10554, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011821050465116278], 0, 2.4074759483337402, 1535886573.2000315], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5987, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011328990337078652], 0, 4.174314498901367, 1535887325.4185164], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5731, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010601982421052632], 0, 0.8256945610046387, 1535887838.7875333], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000989827931372549], 0, 3.99890398979187, 1535888387.7525854], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13882, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009269328055555555], 0, 3.5627260208129883, 1535888910.6716933], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14186, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008406298319327731], 0, 2.376335859298706, 1535889310.2788723], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007723444615384615], 0, 0.8594024181365967, 1535889867.1626806], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006945043680555555], 0, 1.6334896087646484, 1535890372.3664432], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 688, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [8, 4]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.00045975260454545456], 0, 2.5292694568634033, 1535890765.7371151], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13674, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006164924329268293], 0, 0.8362984657287598, 1535891429.6714761], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34456, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009818211363636363], 0, 0.8427114486694336, 1535891903.887033], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31587, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0047900266666666665], 0, 3.763829469680786, 1535892723.5280082], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31828, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004445607434782608], 0, 4.628793716430664, 1535893302.3123899], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30147, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004188478166666666], 0, 2.318864583969116, 1535893811.1717522], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5332, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003845980222222222], 0, 0.8565471172332764, 1535894208.3171473], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12291, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035866158214285715], 0, 3.9173243045806885, 1535894707.1015103], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30100, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0032345457741935484], 0, 0.8719861507415771, 1535895272.2647822], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12003, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0029437744285714285], 0, 0.9651789665222168, 1535895860.8605187], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23764, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002601547564102564], 0, 0.9306168556213379, 1535896468.0587988], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022974321333333334], 0, 0.9792115688323975, 1535897198.4402585], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12867, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019340523773584907], 0, 2.3084769248962402, 1535897929.8870792], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30442, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001566614390625], 0, 0.9274990558624268, 1535898548.6593454], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 797, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [8, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.002275420888888889], 0, 3.4180822372436523, 1535898893.1411545], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30690, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 4]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012414517530864198], 0, 3.330500364303589, 1535899823.8126338], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22724, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0105196242], 0, 0.945650577545166, 1535900969.133171], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22851, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009416030727272727], 0, 3.681612491607666, 1535901247.4323485], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 43204, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008122868923076922], 0, 1.628535509109497, 1535901748.5948753], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55499, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0068992716], 0, 1.6304936408996582, 1535902586.87599], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54595, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004381857826086957], 0, 0.8520708084106445, 1535903352.69153], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1297, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [2, 16]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.008924321333333334], 0, 7.0948522090911865, 1535903807.6040685], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18245, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [1, 8]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002318596977272727], 0, 2.9796509742736816, 1535904593.8101113], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24334, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005389464052631579], 0, 0.8880598545074463, 1535905074.1790903], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21971, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0045559385909090915], 0, 3.446683883666992, 1535906098.2110915], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19630, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0038159604074074075], 0, 0.8763372898101807, 1535906346.4639935], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18627, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [8, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0014601200579710144], 0, 4.467316627502441, 1535907038.8601227], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1988, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [32, 12]], ["tile_c", "sp", [32, 14]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.00262698], 0, 2.861510992050171, 1535907562.0645094], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19448, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0033385871666666666], 0, 1.575890064239502, 1535908252.4618568], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 27701, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [1, 8]], ["tile_ow", "sp", [8, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002229480111111111], 0, 2.2805023193359375, 1535909048.7053776], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 11973, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022322292444444444], 0, 4.471539258956909, 1535909697.9482665], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20595, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [1, 8]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0028671103999999997], 0, 6.215193510055542, 1535910254.8434348], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23928, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0024034559523809524], 0, 0.8770339488983154, 1535910828.5631251], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24759, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016375382741935485], 0, 1.0871405601501465, 1535911564.3831959], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24758, "c": null, "e": [["tile_co", "sp", [20, 16]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027002461842105265], 0, 0.9368150234222412, 1535912310.7056985], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 6225, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008414321], 0, 1.5206542015075684, 1535912784.2860506], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 4825, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008365221333333334], 0, 4.472844362258911, 1535913159.860301], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 5945, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0070752508], 0, 4.0602805614471436, 1535913850.0383084], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 4134, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.006018351117647059], 0, 4.156597375869751, 1535914425.7783875], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 2473, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007068676666666667], 0, 1.5000953674316406, 1535915101.6790938], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 2022, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006007800411764706], 0, 2.885835886001587, 1535915516.9348955], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5142, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005489702368421053], 0, 5.466745615005493, 1535915996.0147579], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 1017, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005811262], 0, 0.7953202724456787, 1535916474.2688396], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 1446, "c": null, "e": [["tile_co", "sp", [2, 64]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003521843896551724], 0, 0.8691070079803467, 1535916868.0933282], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 1129, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005793940611111111], 0, 0.878190279006958, 1535917257.9390473], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3398, "c": null, "e": [["tile_co", "sp", [2, 64]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035198022413793105], 0, 4.688335180282593, 1535917781.7513525], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3494, "c": null, "e": [["tile_co", "sp", [2, 64]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0032823500000000003], 0, 4.607094764709473, 1535918255.5288293], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6169, "c": null, "e": [["tile_co", "sp", [6, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005960715647058823], 0, 4.8614442348480225, 1535918765.805774], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5143, "c": null, "e": [["tile_co", "sp", [6, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002999334411764706], 0, 6.295312166213989, 1535919186.8725255], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2955, "c": null, "e": [["tile_co", "sp", [6, 64]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.028731773], 0, 3.482081413269043, 1535919885.1640832], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13685, "c": null, "e": [["tile_co", "sp", [6, 8]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0020132228], 0, 0.908611536026001, 1535920474.0948074], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2048, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00235106088372093], 0, 3.004063606262207, 1535921160.5698774], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3445, "c": null, "e": [["tile_co", "sp", [6, 8]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001796266107142857], 0, 5.962117433547974, 1535922213.3759124], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2160, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0020313012], 0, 2.077986717224121, 1535922418.9550884], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4252, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008341047272727272], 0, 0.8710401058197021, 1535922775.3250818], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3040, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [8, 12]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.005317129736842105], 0, 2.6970064640045166, 1535923635.4515216], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1912, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [4, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0040032706800000005], 0, 2.7741732597351074, 1535924173.7832673], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 9776, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007503583], 0, 1.1826331615447998, 1535924792.5988057], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3567, "c": null, "e": [["tile_co", "sp", [3, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0016502962786885246], 0, 2.9310302734375, 1535925400.1060352], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4848, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016602016557377048], 0, 6.369439363479614, 1535925858.2798464], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2639, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [71, 1]], ["tile_ow", "sp", [71, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0813388135], 0, 11.77689528465271, 1535926661.5109696], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4166, "c": null, "e": [["tile_co", "sp", [5, 16]], ["tile_oh", "sp", [73, 1]], ["tile_ow", "sp", [73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005762948555555555], 0, 0.7283191680908203, 1535927516.1000335], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 28018, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.034560065], 0, 5.980690956115723, 1535928121.3766432], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4796, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [49, 3]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.018146859166666664], 0, 1.0699045658111572, 1535928625.697254], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 556, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [149, 1]], ["tile_ow", "sp", [149, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0030117573823529413], 0, 1.516263723373413, 1535929050.8396485], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1316, "c": null, "e": [["tile_co", "sp", [125, 8]], ["tile_oh", "sp", [1, 1]], ["tile_ow", "sp", [1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0001911963778625954], 0, 1.576277494430542, 1535929421.767032], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7600, "c": null, "e": [["tile_co", "sp", [160, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0020518567346938777], 0, 0.8154218196868896, 1535929753.9009876], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2272, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015157028636363637], 0, 0.778888463973999, 1535930383.1056135], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5308, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007719584153846154], 0, 1.5802838802337646, 1535930898.12776], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1485, "c": null, "e": [["tile_c", "sp", [480, 2]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00014667120321637428], 0, 0.899038553237915, 1535931360.572654], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2187, "c": null, "e": [["tile_co", "sp", [240, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0008593699159663865], 0, 0.8113443851470947, 1535932126.9752927], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5164, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00047235131132075474], 0, 0.8251545429229736, 1535932815.6865196], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 864, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00014296112711864408], 0, 0.8250195980072021, 1535933234.904522], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16407, "c": null, "e": [["tile_co", "sp", [24, 4]], ["tile_oh", "sp", [1, 14]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0010265503571428572], 0, 0.8855535984039307, 1535934024.1596637], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3510, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003143925327102804], 0, 1.114424228668213, 1535934590.1464806], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 36818, "c": null, "e": [["tile_co", "sp", [72, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011010446483516484], 0, 3.1134889125823975, 1535935438.2064934], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8213, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006945154791666667], 0, 0.8350672721862793, 1535936651.8682947], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5239, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004747314502369668], 0, 2.3365349769592285, 1535936943.9951122], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3443, "c": null, "e": [["tile_c", "sp", [96, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002109249460580913], 0, 1.7872483730316162, 1535937341.461135], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 28115, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000503653765], 0, 1.5329773426055908, 1535938202.4450536], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2188, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0002478011831683168], 0, 0.9888298511505127, 1535938702.694558], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2943, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00017513380244755245], 0, 1.6473076343536377, 1535939070.4297774], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9435, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005305880687830688], 0, 3.6460165977478027, 1535939978.0798883], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 5155, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00044363534513274335], 0, 0.9191522598266602, 1535940405.5527337], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 40603, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [7, 4]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005740788114285714], 0, 1.627058982849121, 1535940916.4930634], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22832, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003942261850393701], 0, 0.9206781387329102, 1535941587.36765], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 7233, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [7, 4]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000504555055276382], 0, 0.966200590133667, 1535941948.7192726], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21315, "c": null, "e": [["tile_co", "sp", [6, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017570933508771931], 0, 5.064387321472168, 1535942764.149184], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 9753, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002005842431372549], 0, 0.9285116195678711, 1535943199.869146], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 106203, "c": null, "e": [["tile_co", "sp", [36, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015103684925373134], 0, 0.8551974296569824, 1535943887.3201168], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 56643, "c": null, "e": [["tile_co", "sp", [6, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011664084302325582], 0, 2.3068089485168457, 1535944572.6458676], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 10203, "c": null, "e": [["tile_c", "sp", [24, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [14, 4]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018732999814814815], 0, 5.422122001647949, 1535944983.8811202], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24125, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003803151444444444], 0, 0.7922976016998291, 1535946173.3377676], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20753, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [8, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011816688117647059], 0, 5.2489306926727295, 1535946855.7831771], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 62703, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [8, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021165278333333333], 0, 3.457056760787964, 1535947539.8797631], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4689, "c": null, "e": [["tile_co", "sp", [128, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00508359705], 0, 1.4686310291290283, 1535947977.7092593], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 584, "c": null, "e": [["tile_c", "sp", [512, 2]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0001574408901098901], 0, 6.548422574996948, 1535948318.1734817], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3721, "c": null, "e": [["tile_co", "sp", [128, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0025811819743589742], 0, 3.850900173187256, 1535948813.7793984], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 412, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00012798284760705289], 0, 0.8876736164093018, 1535949190.7315125], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17412, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0045201612608695655], 0, 2.480592966079712, 1535949895.1965342], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1672, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00027638438016528924], 0, 1.2047803401947021, 1535950294.632389], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17412, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002352006418604651], 0, 6.451756477355957, 1535951027.205684], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1460, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00023311199310344828], 0, 0.9189062118530273, 1535951380.6503155], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13450, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0051228038], 0, 0.8912930488586426, 1535952298.922883], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3332, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [2, 14]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006418003949044586], 0, 4.710473299026489, 1535952675.4560266], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 28046, "c": null, "e": [["tile_co", "sp", [64, 4]], ["tile_oh", "sp", [7, 4]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0025147350999999997], 0, 0.9441139698028564, 1535953414.1382484], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3858, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [7, 4]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004539623053097345], 0, 0.959007740020752, 1535953827.4504845], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 56643, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005621323500000001], 0, 4.6010284423828125, 1535954246.9429193], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 5202, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0017888885263157893], 0, 1.0021793842315674, 1535954595.3741252], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54595, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00312765975], 0, 1.5780363082885742, 1535955237.468042], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 4552, "c": null, "e": [["tile_c", "sp", [16, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0010845631612903227], 0, 2.8521361351013184, 1535955596.6039379], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14283, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [14, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003956065307692307], 0, 2.2524709701538086, 1535956138.6537874], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6074, "c": null, "e": [["tile_c", "sp", [8, 4]], ["tile_h", "sp", [28, 4]], ["tile_w", "sp", [56, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0019402786346153847], 0, 2.3598577976226807, 1535956683.909289], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 27722, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001771450052631579], 0, 4.844622850418091, 1535957273.310632], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1193, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004164394398340249], 0, 1.6432487964630127, 1535957796.5350623], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1745, "c": null, "e": [["tile_co", "sp", [6, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000319620089456869], 0, 2.2563931941986084, 1535958639.675572], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4903, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011515568735632183], 0, 4.139292001724243, 1535959278.483072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1097, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00018543192592592592], 0, 3.4323630332946777, 1535959804.9759617], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3383, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0002305343064516129], 0, 0.8248159885406494, 1535960344.050412], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10424, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00032870451307189544], 0, 1.0060405731201172, 1535961053.0445204], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1803, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0004333651379310345], 0, 2.29067063331604, 1535961568.4802325], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 58942, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00322597903125], 0, 3.0223960876464844, 1535962159.820107], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2914, "c": null, "e": [["tile_co", "sp", [250, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008228543230769231], 0, 3.0065977573394775, 1535962700.1219175], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 563, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [64, 4]], ["tile_c", "sp", [16, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.0013582435810810811], 0, 3.894016742706299, 1535963177.9348078], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1461, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003040722090909091], 0, 2.9221959114074707, 1535963921.8925858], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2285, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000548459131147541], 0, 6.378520488739014, 1535964557.3361697], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 795, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [64, 4]], ["tile_c", "sp", [4, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.005401568904761905], 0, 3.269641637802124, 1535965073.8680248], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15635, "c": null, "e": [["tile_co", "sp", [64, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012531336375], 0, 1.714676856994629, 1535966126.2579775], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9665, "c": null, "e": [["tile_co", "sp", [2, 32]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001847373690909091], 0, 1.5412015914916992, 1535966488.1977332], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12885, "c": null, "e": [["tile_co", "sp", [6, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0015237055757575757], 0, 0.8347368240356445, 1535967089.8821263], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 19381, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004879469666666666], 0, 1.9168751239776611, 1535967539.4468913], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24769, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007386800935251798], 0, 0.933753490447998, 1535968392.4546504], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3445, "c": null, "e": [["tile_co", "sp", [6, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0010121947878787877], 0, 0.8142867088317871, 1535969054.4772363], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 13732, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021917415217391307], 0, 0.898996114730835, 1535969746.9326835], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11082, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003494669024390244], 0, 0.9265255928039551, 1535970247.80693], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10204, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006760299121621622], 0, 0.8375835418701172, 1535970588.984347], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2980, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.009070586166666667], 0, 3.9552371501922607, 1535971269.6922448], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2084, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001562722375], 0, 0.8499212265014648, 1535971822.4000003], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1948, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0014310156428571428], 0, 1.5348868370056152, 1535972624.0837798], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1883, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0008660841465517241], 0, 2.962033987045288, 1535973344.156632], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 12352, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025121241999999998], 0, 0.9767675399780273, 1535973933.0644855], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9328, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0004960180049504951], 0, 5.248998165130615, 1535974372.9741683], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6923, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006477972838709677], 0, 2.3997135162353516, 1535974738.8372047], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 104643, "c": null, "e": [["tile_co", "sp", [24, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.01580936614285714], 0, 4.800135612487793, 1535975637.409461], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1347, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [64, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.020777155500000002], 0, 1.9854884147644043, 1533118077.3115017], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 797, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [32, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.07687202383333333], 0, 2.4723472595214844, 1533117116.6460853], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 797, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [16, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0417725765], 0, 2.2609124183654785, 1533116551.92242], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1297, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [128, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.06775810983333333], 0, 2.2776637077331543, 1533115787.1989372], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1297, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.037354383666666664], 0, 2.0939018726348877, 1533115216.5393353], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 902, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 16]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.09080821066666667], 0, 2.445009708404541, 1533114586.8835123], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 902, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 16]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.05564672833333333], 0, 2.553973913192749, 1533113977.947298], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1815, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [784, 4]], ["tile_k", "sp", [4, 16]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.14363715116666667], 0, 2.6377294063568115, 1533113271.215146], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 22430, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [224, 1]], ["tile_ow", "sp", [32, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.010544594333333332], 0, 0.47663116455078125, 1533107299.325099], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4263, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005988825568862275], 0, 2.754915952682495, 1535984632.3167787], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 823, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004571504227272727], 0, 3.3157143592834473, 1535985071.942518], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15807, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000637066949044586], 0, 4.084044933319092, 1535985723.7015972], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 12171, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004710854181818182], 0, 0.9165973663330078, 1535986299.3396235], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12867, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007072563450704225], 0, 4.936697721481323, 1535986885.4430535], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 6483, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00507628185], 0, 1.7728586196899414, 1535987370.647911], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1823, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00506417605], 0, 0.842127799987793, 1535988070.50219], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2187, "c": null, "e": [["tile_co", "sp", [256, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0051378234], 0, 4.041447639465332, 1535988520.9861257], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2139, "c": null, "e": [["tile_co", "sp", [256, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008808258833333334], 0, 0.9389739036560059, 1535988976.246949], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 401, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [256, 2]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.00413005784], 0, 2.33941912651062, 1535989479.813587], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 783, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0022367072666666667], 0, 2.458928346633911, 1535989996.6707149], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6879, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0047689580952380945], 0, 0.8053445816040039, 1535990457.3312116], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19505, "c": null, "e": [["tile_co", "sp", [256, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004741292954545454], 0, 0.8742308616638184, 1535991099.8192878], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8012, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009543757454545455], 0, 0.8688035011291504, 1535991790.2000918], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 643, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.004276464791666667], 0, 4.099792242050171, 1535992255.3163607], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6124, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002429812452380952], 0, 3.9159204959869385, 1535993107.673077], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4708, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00504312015], 0, 0.8100388050079346, 1535993579.885059], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 38772, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00502503705], 0, 2.301513671875, 1535994260.0613396], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 29644, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.009951001], 0, 0.8911936283111572, 1535994868.2353237], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 798, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [8, 16]], ["tile_c", "sp", [8, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.006394444125], 0, 2.431100845336914, 1535995468.0274594], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13348, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026767548947368423], 0, 0.8435056209564209, 1535995980.825545], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 38252, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0055626987777777775], 0, 4.0184502601623535, 1535996374.5503085], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 64299, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006148074823529411], 0, 0.9109959602355957, 1535996888.0533552], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1297, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 16]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.007179314066666666], 0, 2.7369048595428467, 1535997591.3817925], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=rk3399 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 49115, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015982310476190477], 0, 0.9359464645385742, 1535998260.3719687], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 72880, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.106961243], 0, 1.368462085723877, 1536393445.5359874], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2660, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00662297825], 0, 5.104468584060669, 1535985204.9534225], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3428, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0064107423125], 0, 2.5474369525909424, 1535985817.9107852], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 548, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0062003672352941176], 0, 2.553555965423584, 1535986468.886735], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3364, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0059833248823529415], 0, 1.7026195526123047, 1535987138.2243228], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 612, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005771200166666667], 0, 9.08200716972351, 1535987796.1964505], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3332, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005557581], 0, 2.843329668045044, 1535988508.5214128], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 612, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005344445157894737], 0, 7.527868747711182, 1535989151.7860174], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1284, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005132444], 0, 11.459618330001831, 1535989822.737277], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3332, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004918658619047619], 0, 8.568621635437012, 1535990581.9945655], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3364, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004705861363636364], 0, 7.734873294830322, 1535991262.807085], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 612, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0044908224782608695], 0, 5.983682870864868, 1535991932.69856], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2660, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004281380875], 0, 10.717791557312012, 1535992509.8627102], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3428, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00406764124], 0, 6.924135446548462, 1535993181.9482796], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1444, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003852086846153846], 0, 1.9932496547698975, 1535993801.559875], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 580, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0036426504285714287], 0, 2.5371012687683105, 1535994579.6006505], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1078, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [2, 16]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0012311305731707318], 0, 12.708612203598022, 1535995300.8100455], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2596, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0034325926333333336], 0, 9.121357679367065, 1535996228.4724262], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7404, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0884283565], 0, 3.468022346496582, 1535997370.4672306], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5284, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.021894694200000002], 0, 3.583516836166382, 1535999094.502908], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5284, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0211681858], 0, 9.114071607589722, 1535999931.7826157], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13604, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0204467004], 0, 0.8813538551330566, 1536001394.6085665], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5156, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0197442045], 0, 13.256389141082764, 1536002224.532172], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14244, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.019032308666666668], 0, 0.9120655059814453, 1536003284.5927365], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5924, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.018307628666666666], 0, 3.4800479412078857, 1536004523.9970472], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13732, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.017594627166666665], 0, 11.207655668258667, 1536005469.5458837], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13988, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.016873544], 0, 2.5659217834472656, 1536006489.6103947], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13604, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.016171420285714284], 0, 10.037509679794312, 1536007785.2072124], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6052, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015455659], 0, 2.6320958137512207, 1536008733.3458116], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2724, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.014728859], 0, 8.578466653823853, 1536009979.8663187], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13348, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.01401239425], 0, 3.4833552837371826, 1536011159.920498], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10788, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.013315408375], 0, 6.20878267288208, 1536012090.1761513], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5284, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012602411875], 0, 0.8680591583251953, 1536012938.1228185], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13476, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011906025666666667], 0, 2.56451678276062, 1536013752.8513348], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5796, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011184917333333332], 0, 5.074424982070923, 1536015054.0515509], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14244, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0104913839], 0, 5.092036247253418, 1536016173.2625072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5796, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009771616727272727], 0, 1.7279868125915527, 1536017329.2729309], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13739, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0090636815], 0, 5.954016208648682, 1536018365.868858], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6059, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00835277825], 0, 3.672765016555786, 1536019458.7710483], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5419, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007620281285714286], 0, 1.740419626235962, 1536020336.4687092], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13995, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006894636533333333], 0, 11.11193060874939, 1536021311.1605742], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0062132566470588235], 0, 10.250398874282837, 1536022370.9242525], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.004136529], 0, 7.780946731567383, 1536023330.162095], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13611, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005527556736842106], 0, 7.960354804992676, 1536024164.6822608], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13990, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.091285232], 0, 9.398808240890503, 1536025411.6241415], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11572, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.04399574033333333], 0, 2.0391805171966553, 1536026950.0614262], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13588, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.041117087000000004], 0, 5.783163785934448, 1536028263.3573275], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30580, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03826042366666667], 0, 8.968405485153198, 1536029472.1997998], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30580, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03533505833333333], 0, 9.155192852020264, 1536031097.70375], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11572, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03248380925], 0, 6.610016822814941, 1536032935.4708593], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11572, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.02956564925], 0, 5.5461437702178955, 1536034072.0776339], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13300, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.02664786625], 0, 5.316365003585815, 1536034996.7301824], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12148, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0237270104], 0, 10.904694557189941, 1536035888.1208122], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 32020, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0208300818], 0, 9.444387674331665, 1536036964.2430937], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31444, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0178857255], 0, 6.158954858779907, 1536038602.9000514], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12148, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.014954339285714286], 0, 0.886885404586792, 1536039636.2477448], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 658, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [2, 16]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.0224087858], 0, 4.293109655380249, 1536040225.0317628], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13588, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012060928555555555], 0, 1.7368383407592773, 1536041349.7068474], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22084, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.096010694], 0, 7.387532711029053, 1536042506.7926655], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55876, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.08446824], 0, 3.434619665145874, 1536044036.5829544], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54340, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.072683069], 0, 8.759652376174927, 1536045545.3443148], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 53828, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.061100946], 0, 3.8991539478302, 1536047077.8311696], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 53316, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03809013633333334], 0, 7.406723737716675, 1536048232.0081944], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1892, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0634418555], 0, 11.27404522895813, 1536049162.6518457], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3871, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0215669348], 0, 10.555378437042236, 1536050140.859705], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23582, "c": null, "e": [["tile_co", "sp", [28, 16]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.049791205], 0, 7.122688293457031, 1536051669.3044033], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22087, "c": null, "e": [["tile_co", "sp", [24, 16]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.042839361], 0, 5.763508319854736, 1536053034.5223012], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4092, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.042590925], 0, 11.367834568023682, 1536054215.1692817], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18933, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0132371445], 0, 2.744795799255371, 1536055090.4833481], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2542, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [24, 16]], ["tile_c", "sp", [448, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.024951786400000002], 0, 8.358462810516357, 1536056047.5688665], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3700, "c": null, "e": [["tile_co", "sp", [56, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.03079815775], 0, 10.591206550598145, 1536057856.7775118], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 26659, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [8, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0237657292], 0, 3.016745090484619, 1536058831.2717686], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 11139, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0240564018], 0, 6.604193925857544, 1536059644.1951723], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4437, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.02644703225], 0, 1.8660132884979248, 1536060568.1556642], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10196, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0220472276], 0, 13.803077936172485, 1536062077.9383404], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24253, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0217498264], 0, 2.7448790073394775, 1536063502.1823757], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24476, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03612950366666667], 0, 7.907050132751465, 1536064525.2145276], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 1014, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0993892735], 0, 5.139895677566528, 1536065415.6999116], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 4542, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0994397415], 0, 3.6532845497131348, 1536066467.554051], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 2358, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.082898306], 0, 4.4233314990997314, 1536067552.073468], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 2215, "c": null, "e": [["tile_co", "sp", [8, 20]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0684360385], 0, 14.80985975265503, 1536068472.6201303], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 4654, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.082834549], 0, 8.972679615020752, 1536069247.970078], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 967, "c": null, "e": [["tile_co", "sp", [8, 20]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0684171705], 0, 5.030740976333618, 1536070104.589575], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2070, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.038722916], 0, 7.256319284439087, 1536070889.4199767], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 5942, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.066593377], 0, 11.926833152770996, 1536071899.4370933], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 613, "c": null, "e": [["tile_co", "sp", [4, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.04484594133333333], 0, 12.572870254516602, 1536072879.1274586], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 5942, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.066320086], 0, 6.31308913230896, 1536073635.3825917], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 549, "c": null, "e": [["tile_co", "sp", [4, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.045139644], 0, 9.387719631195068, 1536074613.6114504], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3524, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0312199415], 0, 11.02207612991333, 1536075443.112281], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1295, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.04625590433333333], 0, 8.238906621932983, 1536076230.7780614], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4998, "c": null, "e": [["tile_co", "sp", [8, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.033103249], 0, 3.8261072635650635, 1536076978.7727256], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1286, "c": null, "e": [["tile_co", "sp", [32, 12]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.39137453], 0, 34.057485818862915, 1536077789.728279], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13128, "c": null, "e": [["tile_co", "sp", [2, 24]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0233539218], 0, 8.249753952026367, 1536080094.117024], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9412, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.02795104225], 0, 7.835646390914917, 1536080827.5301204], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16808, "c": null, "e": [["tile_co", "sp", [2, 24]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0207374092], 0, 0.8452932834625244, 1536081521.4709458], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4932, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0247770228], 0, 9.753630638122559, 1536082273.9992583], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8068, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0102894957], 0, 0.8591609001159668, 1536083143.8162096], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2707, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [8, 12]], ["tile_c", "sp", [96, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.04492477533333333], 0, 10.002772331237793, 1536083996.0711129], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1747, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [8, 12]], ["tile_c", "sp", [16, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.03405050166666667], 0, 7.309311389923096, 1536084957.1458223], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 2046, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.080201849], 0, 8.470266342163086, 1536086946.2742312], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6728, "c": null, "e": [["tile_co", "sp", [2, 24]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015452175], 0, 4.61653208732605, 1536088243.4868565], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4932, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.018476856333333333], 0, 3.3900060653686523, 1536089007.2164743], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4767, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [71, 1]], ["tile_ow", "sp", [71, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.9601713445], 0, 20.146138906478882, 1536091153.100271], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4407, "c": null, "e": [["tile_co", "sp", [4, 20]], ["tile_oh", "sp", [73, 1]], ["tile_ow", "sp", [73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.034946037666666666], 0, 2.560015916824341, 1536093762.3698134], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 20961, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.364702647], 0, 57.98618674278259, 1536096079.7633862], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4143, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.183219009], 0, 3.758673906326294, 1536099790.693681], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1084, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [149, 1]], ["tile_ow", "sp", [149, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.03139255225], 0, 10.344429731369019, 1536101391.630419], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 262, "c": null, "e": [["tile_co", "sp", [50, 20]], ["tile_oh", "sp", [1, 1]], ["tile_ow", "sp", [1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0014022152222222222], 0, 9.823032855987549, 1536102224.0373216], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3350, "c": null, "e": [["tile_co", "sp", [320, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0210386474], 0, 3.2289891242980957, 1536102974.238015], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2470, "c": null, "e": [["tile_co", "sp", [20, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015844152857142857], 0, 3.472846508026123, 1536103602.5578265], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1014, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.007985261307692308], 0, 0.9079644680023193, 1536104302.7712083], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1487, "c": null, "e": [["tile_c", "sp", [240, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012059493253012048], 0, 1.9175469875335693, 1536105298.1293778], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12043, "c": null, "e": [["tile_co", "sp", [240, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0072342140714285715], 0, 5.957295179367065, 1536106301.5047193], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4182, "c": null, "e": [["tile_co", "sp", [10, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004800831476190476], 0, 5.051286458969116, 1536107117.9638586], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 444, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009205425229357797], 0, 6.735369682312012, 1536107702.959085], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3895, "c": null, "e": [["tile_co", "sp", [6, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.009530027363636364], 0, 6.973094940185547, 1536108520.7487903], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1788, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00324296864516129], 0, 1.0459721088409424, 1536109194.446854], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 36062, "c": null, "e": [["tile_co", "sp", [72, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009632026272727273], 0, 8.70404577255249, 1536109941.4148066], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4097, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0062922288125], 0, 1.7302289009094238, 1536110850.5155654], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5190, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004207820625], 0, 4.241033554077148, 1536111777.5968788], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 547, "c": null, "e": [["tile_c", "sp", [96, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001973005431372549], 0, 4.385333776473999, 1536112618.4858267], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27733, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004275248083333333], 0, 6.122188091278076, 1536113521.4249365], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1942, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0021312031914893616], 0, 2.5548458099365234, 1536114357.4573731], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2943, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011694673720930233], 0, 5.072686672210693, 1536115218.6615684], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23368, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004920701952380952], 0, 5.94102668762207, 1536116076.477695], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 2621, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004541308086956522], 0, 1.008664608001709, 1536116852.2033715], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 40997, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0051385694], 0, 10.43580937385559, 1536117760.6478994], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9112, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00364323625], 0, 0.8713490962982178, 1536118706.6047802], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 7128, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00412026092], 0, 0.9984474182128906, 1536119484.1909046], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10310, "c": null, "e": [["tile_co", "sp", [2, 12]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.013428558125], 0, 0.8911924362182617, 1536120159.0575712], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 12976, "c": null, "e": [["tile_c", "sp", [72, 2]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [7, 8]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015647953428571428], 0, 1.0471694469451904, 1536120936.4925685], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 103925, "c": null, "e": [["tile_co", "sp", [18, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.014572951142857143], 0, 10.828694105148315, 1536122076.1910734], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11334, "c": null, "e": [["tile_co", "sp", [2, 12]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.009141526636363638], 0, 0.869736909866333, 1536123357.2741227], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3951, "c": null, "e": [["tile_c", "sp", [24, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.013296637625], 0, 1.0238304138183594, 1536124227.1008575], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19327, "c": null, "e": [["tile_co", "sp", [6, 16]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.029333936], 0, 8.717342376708984, 1536125497.9321861], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55103, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00944388809090909], 0, 5.16627836227417, 1536126571.208744], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 50464, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.01602791957142857], 0, 10.481632232666016, 1536127548.4142838], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 972, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0538108185], 0, 7.416002511978149, 1536128387.275845], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 101, "c": null, "e": [["tile_c", "sp", [256, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0013316991688311689], 0, 1.928370475769043, 1536129063.0628688], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3568, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0270013085], 0, 3.765385389328003, 1536129974.4256575], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 92, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0008266065327868853], 0, 0.9601027965545654, 1536130640.2921314], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6604, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.04397093633333333], 0, 1.0184991359710693, 1536131844.749581], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 822, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028401455555555556], 0, 7.667769193649292, 1536132726.2101436], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13813, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0221652226], 0, 8.030568838119507, 1536133690.1328871], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1919, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015889717142857141], 0, 5.5886194705963135, 1536134455.519258], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14638, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.046359602], 0, 3.0045220851898193, 1536135391.9573114], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1685, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006141458058823529], 0, 1.0209779739379883, 1536136216.9901872], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15286, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.023870212199999997], 0, 0.9487848281860352, 1536136856.1833947], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3802, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [14, 2]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0036302355], 0, 10.92769718170166, 1536137617.7127235], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 53828, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.049572218333333334], 0, 7.075160026550293, 1536138479.2779188], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6921, "c": null, "e": [["tile_c", "sp", [64, 2]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [7, 8]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.01431693542857143], 0, 2.9511704444885254, 1536139654.7421224], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 56388, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.02657734925], 0, 6.9130542278289795, 1536140416.7236364], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 5889, "c": null, "e": [["tile_c", "sp", [16, 4]], ["tile_h", "sp", [28, 2]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008887952833333332], 0, 1.02897310256958, 1536141440.1043384], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 56074, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.03094035925], 0, 9.081524848937988, 1536142214.3214188], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3732, "c": null, "e": [["tile_c", "sp", [32, 1]], ["tile_h", "sp", [28, 4]], ["tile_w", "sp", [28, 4]], ["ann", "an", ["none", "vec", "none"]]], "t": "direct"}], "r": [[0.014405389000000001], 0, 4.119186162948608, 1536143409.6923523], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 52263, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.01563523157142857], 0, 3.746095657348633, 1536144399.1464067], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 480, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004561452454545454], 0, 5.0171496868133545, 1536145055.6683118], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1886, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003772162962962963], 0, 6.02678918838501, 1536145715.914727], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5999, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0188976855], 0, 3.6448419094085693, 1536146456.2684476], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2583, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017948420357142857], 0, 5.017563104629517, 1536147553.9738386], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 888, "c": null, "e": [["tile_co", "sp", [2, 24]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00252121745], 0, 10.349458932876587, 1536148371.0074573], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8283, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0030418320000000004], 0, 2.5419938564300537, 1536149243.9752967], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3383, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0051230957], 0, 7.264456510543823, 1536150068.9042654], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11973, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.03026434075], 0, 4.665707588195801, 1536151236.7542784], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2886, "c": null, "e": [["tile_co", "sp", [50, 20]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.095531283], 0, 10.996755361557007, 1536152580.7919078], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.01396799875], 0, 9.230791330337524, 1536153945.1142707], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4000, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003110784696969697], 0, 10.347509145736694, 1536154818.2446988], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2244, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0060861833529411765], 0, 7.644370079040527, 1536155431.0104601], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 708, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0545944645], 0, 5.0825982093811035, 1536156101.967289], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6087, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011823206111111111], 0, 3.363551378250122, 1536156797.3972626], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2048, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.018655831666666668], 0, 9.16552734375, 1536157691.3797135], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13646, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.012624696875], 0, 4.719533205032349, 1536159071.1554952], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 18653, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0528823615], 0, 2.0929510593414307, 1536160210.3209066], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23806, "c": null, "e": [["tile_co", "sp", [16, 12]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0064070591875], 0, 1.841479778289795, 1536161637.0200956], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13326, "c": null, "e": [["tile_co", "sp", [4, 12]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008418026083333334], 0, 3.641582489013672, 1536162855.5396953], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2467, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.023726238], 0, 9.779823064804077, 1536164096.9034014], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5667, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0031158892424242423], 0, 1.7324836254119873, 1536165208.2339666], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4060, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006593819625], 0, 8.10648226737976, 1536166380.0293431], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5410, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.1056160045], 0, 10.496391773223877, 1536168424.1510243], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2820, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [55, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.015793224571428573], 0, 8.22214674949646, 1536169574.822832], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10084, "c": null, "e": [["tile_co", "sp", [2, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [55, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.016040679999999998], 0, 4.058653116226196, 1536170713.991018], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1463, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.009992259], 0, 10.457476139068604, 1536171622.5953014], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11902, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.02753597075], 0, 10.296802759170532, 1536172527.8882258], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4596, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [55, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004533768956521739], 0, 3.5457844734191895, 1536173781.0732815], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8823, "c": null, "e": [["tile_co", "sp", [2, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0075965217857142855], 0, 6.614614248275757, 1536174655.4988232], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 97337, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.1589562825], 0, 28.049460649490356, 1536177577.2210584], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1771, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [8, 2]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [512, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.152017326], 0, 37.43001127243042, 1536179514.2314026], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 608, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [512, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.5756763175], 0, 37.720627784729004, 1536181367.2407982], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 758, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.306084444], 0, 14.39254641532898, 1536183539.2286177], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1771, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [98, 2]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.5395060455], 0, 61.31764054298401, 1536186291.0563264], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1896, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [98, 2]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.311042132], 0, 27.3410165309906, 1536278808.4519851], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2648, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [196, 4]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.693819379], 0, 6.901161432266235, 1536282601.0950372], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2479, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [392, 2]], ["tile_k", "sp", [8, 16]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.4342901205], 0, 31.95503568649292, 1536286181.557321], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2648, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [784, 4]], ["tile_k", "sp", [8, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[1.016162306], 0, 34.12473940849304, 1536288930.9953587], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 81735, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [224, 1]], ["tile_ow", "sp", [112, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.109269532], 0, 6.661062479019165, 1536292458.8620472], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 884, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.006781742733333334], 0, 2.5316274166107178, 1536293363.1361797], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1622, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0741942105], 0, 7.260768175125122, 1536294413.6586921], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6520, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0058008065], 0, 5.178553104400635, 1536295348.5083904], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 6231, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0605661385], 0, 7.528777837753296, 1536296230.522926], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13012, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006671732733333334], 0, 9.478656530380249, 1536297077.062968], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 23379, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.06260136], 0, 4.352974891662598, 1536297975.3722384], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1764, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.053987766], 0, 4.471951961517334, 1536298782.3274467], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2116, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0539957875], 0, 11.578891515731812, 1536299586.051567], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2116, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.107573803], 0, 12.74745225906372, 1536374774.4937918], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1063, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [512, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.03672958466666667], 0, 3.272906541824341, 1536375989.6812162], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3244, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.027036744], 0, 3.5768511295318604, 1536376906.10542], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6376, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.044541668666666666], 0, 3.478752613067627, 1536378211.2710476], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15008, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.044386678000000006], 0, 0.9905264377593994, 1536379287.0828285], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19232, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.091334751], 0, 12.037572383880615, 1536380471.1364822], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.042157331], 0, 12.140062093734741, 1536381791.813934], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6808, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0234645674], 0, 7.370471239089966, 1536382988.1830661], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31156, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.046766849], 0, 9.168155908584595, 1536383984.3123465], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 38584, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.047209113999999996], 0, 9.272347211837769, 1536385066.4190762], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 38224, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0964747025], 0, 1.2207269668579102, 1536386388.4979112], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 758, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [8, 16]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0522491935], 0, 6.515461206436157, 1536387469.376385], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13012, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.02551019625], 0, 2.569166421890259, 1536388291.9547625], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 47100, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0504954185], 0, 10.576186180114746, 1536389090.895174], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 61132, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.052133712], 0, 5.421657085418701, 1536390820.5900688], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1767, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [8, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.060729868], 0, 4.567668199539185, 1536391328.8810341], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=pynq -target=armv7a-linux-eabi -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 49340, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.013793001125], 0, 3.6062705516815186, 1536392247.2690873], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 30249, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0067798176875], 0, 3.2185189723968506, 1536121438.3648748], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2579, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006973204], 0, 1.3633522987365723, 1535867857.4162548], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2708, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0007745112153846154], 0, 5.821341276168823, 1535868613.0246267], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 748, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.000693746455782313], 0, 3.441767930984497, 1535869789.1854405], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1292, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006682881933333333], 0, 2.3297817707061768, 1535870726.534538], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1452, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006345391454545454], 0, 5.6044347286224365, 1535871751.6955447], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005793312988505747], 0, 0.7147471904754639, 1535872670.0655181], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3508, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006227516832298136], 0, 4.035978555679321, 1535873488.912006], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2732, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005943591637426901], 0, 0.9178712368011475, 1535874466.0940254], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3572, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004929007586206897], 0, 1.403226613998413, 1535875134.0138905], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2740, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0004830804278846154], 0, 6.199410915374756, 1535875987.7227733], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 756, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0005473559890710382], 0, 1.6840357780456543, 1535876800.9521306], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1524, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004310354077253219], 0, 1.413905143737793, 1535877455.5635722], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 660, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00040739834136546184], 0, 2.1056830883026123, 1535878107.1478257], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3412, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004486360269058296], 0, 4.943973541259766, 1535878766.931437], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1331, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003857062274368231], 0, 1.4150216579437256, 1535879332.3861043], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1030, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [16, 8]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "none"]]], "t": "winograd"}], "r": [[0.00015473625557206537], 0, 2.471970796585083, 1535880173.9594421], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2739, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00037653900375939845], 0, 4.647382020950317, 1535880820.8591723], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12893, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008452473916666666], 0, 0.8132140636444092, 1535881704.7049239], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5963, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002154798319148936], 0, 2.345853090286255, 1535882369.645099], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13771, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0020856445416666666], 0, 1.4828360080718994, 1535883215.4536846], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13508, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022473032222222224], 0, 1.4799690246582031, 1535883947.1987643], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14307, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019453024807692307], 0, 1.4401798248291016, 1535884682.074935], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13643, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018731192037037036], 0, 0.7684931755065918, 1535885595.9936612], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6115, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018014695], 0, 7.233718156814575, 1535886381.2281792], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14051, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001734797051724138], 0, 6.814533948898315, 1535887310.4025786], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10436, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001867467203703704], 0, 3.4698684215545654, 1535888026.1399605], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13795, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015929976825396824], 0, 2.3382205963134766, 1535888824.8254836], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14307, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015188762575757578], 0, 7.851373910903931, 1535889599.335516], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014442262], 0, 3.376384735107422, 1535890306.9581919], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013807505753424658], 0, 0.7294833660125732, 1535891267.0679455], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5219, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001311038948051948], 0, 3.6140637397766113, 1535892429.984995], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5579, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012501822875], 0, 4.060142278671265, 1535893236.5809393], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011812867647058823], 0, 5.279164791107178, 1535894190.4662569], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011198148111111112], 0, 2.784979820251465, 1535895053.2702043], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010510091145833334], 0, 4.410005331039429, 1535896361.6936448], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5835, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009828890980392157], 0, 8.209392786026001, 1535897575.8223896], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14307, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009119697], 0, 5.198476791381836, 1535898737.7374442], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008478813559322034], 0, 1.611907720565796, 1535900001.3891747], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13771, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007724078538461538], 0, 1.5435268878936768, 1535900650.594195], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13923, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007093383169014084], 0, 9.112094163894653, 1535901749.3960054], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006334021257861635], 0, 3.1378026008605957, 1535902683.3292675], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1317, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.00037911537121212123], 0, 2.5609993934631348, 1535903351.2208252], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5322, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006244824037267081], 0, 2.7735486030578613, 1535904277.3955193], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34518, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0090734375], 0, 8.924296617507935, 1535905260.9196935], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30683, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004496422086956522], 0, 0.8161787986755371, 1535906092.935598], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12579, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003940985576923077], 0, 1.7681026458740234, 1535906912.192131], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31835, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0039139423076923075], 0, 0.8010337352752686, 1535907837.1773732], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12579, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003391805533333333], 0, 2.0879979133605957, 1535908735.9604592], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11715, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0030947127575757577], 0, 3.8538825511932373, 1535909702.246306], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30147, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002802748833333333], 0, 2.141653060913086, 1535911021.0215058], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30107, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026796189736842104], 0, 2.8362505435943604, 1535911918.362335], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 32163, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002243530088888889], 0, 2.7420859336853027, 1535912796.8561633], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001957421884615385], 0, 0.7765591144561768, 1535913875.102971], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30723, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016838975833333334], 0, 0.7851724624633789, 1535915283.7894862], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30395, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014903492647058824], 0, 4.427369832992554, 1535916177.9326692], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 535, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [8, 4]], ["tile_c", "sp", [16, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0013692567567567568], 0, 2.8027522563934326, 1535916861.0110612], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30723, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00112436797752809], 0, 7.195612192153931, 1535917812.5680432], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22851, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009785464], 0, 2.941189765930176, 1535918952.5005338], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 57155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008618533], 0, 3.7676258087158203, 1535920143.2995026], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55107, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007471019357142858], 0, 5.736740827560425, 1535921089.8946888], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54475, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0062680013125], 0, 0.9397869110107422, 1535921996.895203], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 44740, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003583203142857143], 0, 8.052263259887695, 1535922775.5954125], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1292, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.006663444], 0, 9.646049976348877, 1535923425.3580048], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19001, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0024084325476190476], 0, 6.948801040649414, 1535924407.356199], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24824, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0055109101052631575], 0, 1.399707317352295, 1535925762.9869232], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26835, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004737949818181818], 0, 0.7538621425628662, 1535927070.7897084], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 19224, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003961718769230769], 0, 1.6735901832580566, 1535928384.6178102], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24377, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015094760597014925], 0, 2.1761128902435303, 1535929283.5266976], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1988, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [32, 12]], ["tile_c", "sp", [32, 14]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0034115625], 0, 2.796121597290039, 1535929898.2997253], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18552, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003448814655172414], 0, 3.5635452270507812, 1535930866.9282033], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 27749, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001940625], 0, 7.755219221115112, 1535931827.0118506], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 27603, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018429640181818181], 0, 0.8039741516113281, 1535932609.0807226], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20691, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0029632589142857143], 0, 2.7548351287841797, 1535933899.0524952], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24600, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0024730818048780486], 0, 0.7370038032531738, 1535934671.719636], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24547, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016274776857142857], 0, 6.345768451690674, 1535935444.3502874], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9804, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002555807275], 0, 0.8150131702423096, 1535936533.6883445], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 8023, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.010131380166666667], 0, 5.966093063354492, 1535937624.5787191], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 6917, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.009679166692307692], 0, 10.309670686721802, 1535938882.5727322], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3108, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "vec", "none"]]], "t": "direct"}], "r": [[0.0063670246875], 0, 0.6734812259674072, 1535939841.8675945], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 7356, "c": null, "e": [["tile_co", "sp", [160, 1]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.005615335611111111], 0, 2.9724414348602295, 1535941134.9305546], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 8694, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.004564630681818182], 0, 3.107525587081909, 1535942028.1641054], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 5400, "c": null, "e": [["tile_co", "sp", [160, 1]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "t": "direct"}], "r": [[0.0037127121851851853], 0, 2.900451183319092, 1535942805.446378], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 988, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0050244792], 0, 0.9805982112884521, 1535943429.6996932], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 7085, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.004225021666666666], 0, 4.7361555099487305, 1535943882.846985], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 5033, "c": null, "e": [["tile_co", "sp", [64, 2]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.003025089285714286], 0, 6.492930889129639, 1535944670.548397], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 7855, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.0043831295217391306], 0, 2.5758743286132812, 1535945356.4632545], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3921, "c": null, "e": [["tile_co", "sp", [64, 2]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.002792028361111111], 0, 0.7271623611450195, 1535945975.6412723], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2612, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003896875], 0, 0.762338399887085, 1535946454.310427], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4851, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.005894393411764706], 0, 3.0600411891937256, 1535946919.2557883], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2213, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022917968636363634], 0, 4.185027122497559, 1535947360.9372318], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5157, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.03768385425], 0, 2.5912251472473145, 1535948053.9145346], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13493, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0019152810566037735], 0, 0.917884349822998, 1535948544.4498672], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4764, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016814323], 0, 1.4706261157989502, 1535949240.9026003], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17523, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016700868166666667], 0, 0.7155423164367676, 1535949709.1002097], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12156, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019220027358490567], 0, 0.9091367721557617, 1535950275.0830157], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10208, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008011333280000001], 0, 5.996466159820557, 1535950708.5686529], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3064, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [6, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0033552604], 0, 2.8562121391296387, 1535951368.0915186], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1280, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [9, 9]], ["tile_k", "sp", [32, 3]], ["tile_c", "sp", [8, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0026423399268292683], 0, 3.523475408554077, 1535951967.595978], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 9425, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [5, 7]], ["tile_ow", "sp", [35, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004961954380952381], 0, 0.8645095825195312, 1535952349.911913], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12853, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011124884222222223], 0, 1.521202564239502, 1535953141.5501883], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2496, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0011914744523809525], 0, 1.3707749843597412, 1535953635.322519], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4571, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [71, 1]], ["tile_ow", "sp", [1, 71]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.093674479], 0, 15.636468887329102, 1535954293.5397668], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1844, "c": null, "e": [["tile_co", "sp", [10, 8]], ["tile_oh", "sp", [73, 1]], ["tile_ow", "sp", [73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004203081583333333], 0, 3.316103458404541, 1535955254.1675594], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 27807, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [21, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0269822915], 0, 3.5311408042907715, 1535956095.5218112], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10195, "c": null, "e": [["tile_co", "sp", [16, 2]], ["tile_oh", "sp", [49, 3]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.016545907714285716], 0, 1.1963660717010498, 1535956738.4667482], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 507, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [149, 1]], ["tile_ow", "sp", [149, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002893482142857143], 0, 4.071887969970703, 1535957352.6064842], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1764, "c": null, "e": [["tile_co", "sp", [125, 8]], ["tile_oh", "sp", [1, 1]], ["tile_ow", "sp", [1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004332273593073593], 0, 1.3035788536071777, 1535957754.5327072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1192, "c": null, "e": [["tile_co", "sp", [160, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0019142000754716981], 0, 1.6121251583099365, 1535958033.5743163], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6250, "c": null, "e": [["tile_co", "sp", [20, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013602638933333333], 0, 2.869635820388794, 1535958487.9459636], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2090, "c": null, "e": [["tile_co", "sp", [40, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007885416614173228], 0, 0.6892850399017334, 1535958916.6421394], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1494, "c": null, "e": [["tile_c", "sp", [60, 16]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[9.648331402714931e-05], 0, 3.0957162380218506, 1535959459.1312845], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4987, "c": null, "e": [["tile_co", "sp", [240, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005570775444444444], 0, 0.886345386505127, 1535960113.486169], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2092, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00039246978244274807], 0, 2.249979019165039, 1535960746.7005074], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 864, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[9.530261669829222e-05], 0, 0.9123079776763916, 1535961209.5734022], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21077, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008493776525423729], 0, 0.9564721584320068, 1535961970.534097], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4434, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000200709376], 0, 0.8978424072265625, 1535962380.0124834], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 37217, "c": null, "e": [["tile_co", "sp", [72, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007940575748031497], 0, 0.7120976448059082, 1535963440.473609], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8021, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005760416666666667], 0, 0.7351577281951904, 1535963998.0267239], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5330, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00040161458], 0, 0.7181363105773926, 1535964650.2331035], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3379, "c": null, "e": [["tile_c", "sp", [96, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0001350284852546917], 0, 1.6671388149261475, 1535965082.9408898], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26711, "c": null, "e": [["tile_co", "sp", [24, 16]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00032661848534201954], 0, 2.8192408084869385, 1535966061.7003074], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4903, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00020283985656565655], 0, 4.000597238540649, 1535966893.9300532], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2945, "c": null, "e": [["tile_c", "sp", [24, 8]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00011407912884834663], 0, 0.9913613796234131, 1535967352.980321], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9651, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004345621385281385], 0, 1.4508092403411865, 1535967911.8562243], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6667, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002679348411458333], 0, 0.8778612613677979, 1535968405.4646528], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10255, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0003498561328671329], 0, 3.230353593826294, 1535969249.1668587], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9435, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00033762100336700337], 0, 4.053060293197632, 1535969793.8551376], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 5448, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [4, 7]], ["tile_w", "sp", [28, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00036073703597122304], 0, 3.5192062854766846, 1535970270.5058587], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54597, "c": null, "e": [["tile_co", "sp", [3, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014741727941176472], 0, 0.809891939163208, 1535971022.8079114], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 9678, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [4, 14]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0010114057171717171], 0, 1.1268048286437988, 1535971539.519611], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 76925, "c": null, "e": [["tile_co", "sp", [18, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011376174509803922], 0, 1.1307761669158936, 1535972141.278059], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55109, "c": null, "e": [["tile_co", "sp", [3, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009441381698113207], 0, 2.1401026248931885, 1535972952.7344315], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 10203, "c": null, "e": [["tile_c", "sp", [24, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [14, 4]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008842288070175439], 0, 3.7907180786132812, 1535973417.0360305], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25335, "c": null, "e": [["tile_co", "sp", [24, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0027220157567567566], 0, 0.8117961883544922, 1535973966.1247025], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22302, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [7, 16]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011130671333333334], 0, 1.5748915672302246, 1535974622.1091774], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 65049, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [14, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019024760983606557], 0, 4.330333232879639, 1535975318.45034], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4822, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004502943826086956], 0, 2.8856000900268555, 1535975757.623309], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 454, "c": null, "e": [["tile_c", "sp", [128, 8]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[9.530560775047258e-05], 0, 3.24080491065979, 1535976221.6874697], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 994, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0022784838], 0, 0.7129216194152832, 1535976718.8039546], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 533, "c": null, "e": [["tile_c", "sp", [64, 8]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[9.03847514893617e-05], 0, 0.8161959648132324, 1535977128.541298], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7453, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00376772762962963], 0, 6.457433223724365, 1535977568.535984], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1632, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00017829211229946523], 0, 0.7968342304229736, 1535977998.6842341], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17083, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019315705], 0, 1.611405611038208, 1535978497.184628], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1893, "c": null, "e": [["tile_c", "sp", [32, 8]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000153387013803681], 0, 2.1249988079071045, 1535978875.9867256], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34842, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004321915038461539], 0, 3.3214404582977295, 1535979432.3609753], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4286, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00036056842086330934], 0, 0.8227465152740479, 1535979870.2205324], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 35859, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019263321346153846], 0, 4.08035683631897, 1535980441.4232445], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3810, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003220073022508039], 0, 1.5286197662353516, 1535980823.2337446], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 42699, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004556060636363636], 0, 1.0215740203857422, 1535981286.1096067], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 5154, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [7, 8]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0010547014464285715], 0, 0.8195693492889404, 1535981740.9459593], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8836, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.002104937083333333], 0, 1.5973365306854248, 1535982414.2158563], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 5847, "c": null, "e": [["tile_c", "sp", [16, 4]], ["tile_h", "sp", [8, 7]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005481176413043478], 0, 1.0518739223480225, 1535982873.7177157], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 77849, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0032088704375], 0, 2.232079267501831, 1535983221.1078334], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6032, "c": null, "e": [["tile_c", "sp", [8, 4]], ["tile_h", "sp", [8, 14]], ["tile_w", "sp", [112, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009347034504504504], 0, 2.727337598800659, 1535983659.6081777], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 25274, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [28, 4]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010112273962264151], 0, 1.1193735599517822, 1535984264.1494517], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2257, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00037541744569288387], 0, 2.8062586784362793, 1535984784.104372], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1623, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00032136895268138803], 0, 1.4531347751617432, 1535985297.6182828], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 6191, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007309648686131386], 0, 1.1906604766845703, 1535985799.5098386], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4793, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0001272067743979721], 0, 2.2798943519592285, 1535986748.0467825], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1863, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00021190854449152543], 0, 2.492400884628296, 1535987419.9554431], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4275, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003266417184466019], 0, 0.7313265800476074, 1535987971.1034648], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3722, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00041355027272727273], 0, 0.8023056983947754, 1535988344.670053], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 30880, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0023402979534883723], 0, 0.8697738647460938, 1535988895.7753098], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1316, "c": null, "e": [["tile_co", "sp", [125, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.006784132], 0, 2.8145053386688232, 1535989669.969148], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 518, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [64, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.0008794407894736842], 0, 2.558967351913452, 1535990265.593547], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 741, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00021783882289416848], 0, 3.5209481716156006, 1535991028.3412025], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 549, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0004910896666666667], 0, 2.2042293548583984, 1535991562.7885213], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 795, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [64, 4]], ["tile_c", "sp", [4, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.003511099137931035], 0, 3.2780706882476807, 1535992075.3326764], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6088, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008145155772357723], 0, 3.941746234893799, 1535992592.3353176], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5212, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012894564615384616], 0, 1.0187790393829346, 1536019303.8000207], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3603, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0011537595747126437], 0, 2.396380662918091, 1536019797.697808], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24545, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00346008064516129], 0, 1.101011037826538, 1536020572.4407218], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24479, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004525640995475113], 0, 1.4700887203216553, 1536021452.7269256], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13843, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009440153301886793], 0, 1.5362191200256348, 1536021925.2563477], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2883, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001569848625], 0, 7.309131145477295, 1536022322.493493], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5924, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002169019892933619], 0, 0.8522911071777344, 1536022973.9088604], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4083, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000645098826923077], 0, 0.8601901531219482, 1536023423.6702647], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 14274, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007118194466666667], 0, 2.094909429550171, 1536023918.8025923], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2596, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0010894927608695651], 0, 0.7477781772613525, 1536024328.3871667], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10131, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010180397777777778], 0, 0.9929845333099365, 1536097563.6325727], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7002, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006909913793103448], 0, 1.3022198677062988, 1536097988.655825], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 9802, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0017505298103448275], 0, 0.8012533187866211, 1536098502.4173906], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4960, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00027001647721179627], 0, 4.2481467723846436, 1536098994.535059], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8522, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006101767696969697], 0, 0.6880800724029541, 1536099477.7670798], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 22085, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [14, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0101779687], 0, 4.515950918197632, 1536100066.6133456], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1343, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [64, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.013762586833333333], 0, 2.0050904750823975, 1533102459.8475103], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 797, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [32, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.04702907983333333], 0, 2.572512149810791, 1533101388.8959608], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 795, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [128, 4]], ["tile_c", "sp", [16, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.023498090166666666], 0, 2.263050079345703, 1533100829.2756784], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1347, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [32, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.040334461833333335], 0, 2.449192762374878, 1533100036.573083], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1322, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.023134722166666667], 0, 2.081481456756592, 1533099469.6317222], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1843, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [196, 4]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.047807465333333334], 0, 1.9650189876556396, 1533098865.244454], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1815, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [196, 4]], ["tile_k", "sp", [8, 16]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0302609375], 0, 1.8535256385803223, 1533098147.8221877], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1885, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [784, 4]], ["tile_k", "sp", [4, 16]], ["tile_c", "sp", [8, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.07199887149999999], 0, 2.7277400493621826, 1533097296.825846], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 42597, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [112, 2]], ["tile_ow", "sp", [32, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009713454833333334], 0, 0.35669827461242676, 1533090293.6936846], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4224, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00044601486721991703], 0, 2.852078676223755, 1536107943.0204308], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 903, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003737635037037037], 0, 5.141981601715088, 1536108532.5755713], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5980, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00041375903719008265], 0, 2.1418228149414062, 1536109168.6666536], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3243, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0032984039032258063], 0, 3.2871994972229004, 1536109859.5890794], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30100, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004674491302325581], 0, 0.7585835456848145, 1536110384.2188175], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24347, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003826059], 0, 5.195935249328613, 1536111188.517972], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3464, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004463224608695652], 0, 3.124676465988159, 1536111620.2835436], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5164, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0045538589090909085], 0, 3.415370464324951, 1536112035.360572], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1036, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.007579389857142857], 0, 2.9171316623687744, 1536112430.6656375], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 809, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [64, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.004700213090909091], 0, 6.611178874969482, 1536112898.7217844], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3264, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0018582272222222222], 0, 2.0837552547454834, 1536113358.7907305], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11604, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0038069444444444444], 0, 5.336071729660034, 1536114266.0101695], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18615, "c": null, "e": [["tile_co", "sp", [128, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003853906269230769], 0, 0.9708969593048096, 1536114827.3898988], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18396, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0063206054375], 0, 5.031559944152832, 1536115279.0848536], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1293, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [128, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.00324741264516129], 0, 2.6970198154449463, 1536115698.3109655], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11596, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001570084640625], 0, 4.061263084411621, 1536116236.1538243], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004001080259259259], 0, 3.1615545749664307, 1536116734.631115], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 37683, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003977155172413793], 0, 6.039116621017456, 1536117219.0879], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31022, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.011737557888888889], 0, 1.7811172008514404, 1536117701.5345201], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 785, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [32, 4]], ["tile_c", "sp", [16, 8]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.00393106973076923], 0, 2.9623961448669434, 1536118125.8196707], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24628, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0021959465652173914], 0, 2.068652629852295, 1536118723.4467762], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 47771, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00485018752], 0, 2.5046238899230957, 1536119190.8156369], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26644, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004184917541666667], 0, 3.6412549018859863, 1536119852.2769718], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1322, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 16]], ["tile_c", "sp", [16, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.004828693181818181], 0, 2.8729515075683594, 1536120307.99372], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=kirin970 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 47156, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011112673666666666], 0, 2.6773269176483154, 1536120918.5177624], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 74280, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006116716294117647], 0, 0.8584136962890625, 1536136638.2027256], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1460, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00046159566820276496], 0, 2.95910906791687, 1535941993.1768527], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3412, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004467060044642857], 0, 2.235837697982788, 1535942408.7816665], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3476, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004326605603448276], 0, 0.7859692573547363, 1535942850.2488618], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3540, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004209106260504202], 0, 3.943957805633545, 1535943183.6446962], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3508, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004031792128514056], 0, 0.79543137550354, 1535943529.2815614], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 596, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00038899551550387597], 0, 0.7725718021392822, 1535944090.1169941], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 596, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.000372919219330855], 0, 1.5162627696990967, 1535944625.438159], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2804, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003659160547445256], 0, 0.7525911331176758, 1535945211.808394], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2740, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003512473135888502], 0, 0.7581408023834229, 1535945744.7994215], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3412, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00033481504347826084], 0, 5.8844146728515625, 1535946162.4519515], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1300, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00031751293015873014], 0, 0.7655775547027588, 1535946649.967913], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3444, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00030708161585365855], 0, 0.7408092021942139, 1535947074.464277], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1492, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002853505539772727], 0, 0.7644622325897217, 1535947532.5670466], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3412, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00028458438873239436], 0, 0.7970550060272217, 1535948018.1016922], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3572, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00025816669230769234], 0, 0.7271606922149658, 1535948575.0510943], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1195, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 2]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [8, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.00010214286034658512], 0, 2.4247498512268066, 1535949045.7754717], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2644, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0002292443986332574], 0, 0.7795882225036621, 1535949353.6020837], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17213, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005905738941176471], 0, 0.8417942523956299, 1535949765.8698359], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5451, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015041504925373136], 0, 3.5004403591156006, 1535950330.1632025], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13667, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014682587536231884], 0, 0.7581527233123779, 1535950898.9669318], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5987, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014156916805555557], 0, 1.736725091934204, 1535951492.6076767], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5963, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013576704594594595], 0, 5.34473180770874, 1535951950.0357766], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013055885974025975], 0, 0.7965493202209473, 1535952665.2261236], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13387, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012602240875], 0, 0.7512173652648926, 1535953298.6050177], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6091, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012149762650602409], 0, 0.8382539749145508, 1535953794.190265], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011681681162790697], 0, 0.7328383922576904, 1535954135.7971358], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13923, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011121060222222222], 0, 1.7831757068634033, 1535954830.7671313], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001061046705263158], 0, 2.5931198596954346, 1535955480.0854414], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13515, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010108897272727273], 0, 1.540914535522461, 1535955939.1311157], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009658679807692307], 0, 3.4772584438323975, 1535956446.672383], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10315, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009186564403669725], 0, 4.74798583984375, 1535956897.021324], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5219, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008720512695652174], 0, 2.4436705112457275, 1535957613.145204], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008314683140495867], 0, 3.915058135986328, 1535958207.8747637], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13411, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007814938125], 0, 4.3618268966674805, 1535958800.3345053], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13923, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007332041459854015], 0, 3.497318983078003, 1535959189.192695], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14179, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006860520821917809], 0, 2.2991044521331787, 1535959777.7926369], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5963, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006380530764331211], 0, 5.401664972305298, 1535960165.6129115], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14155, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005960429702380952], 0, 0.7711784839630127, 1535960651.797142], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005431081621621622], 0, 3.1935198307037354, 1535961264.0639458], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13539, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004965620346534654], 0, 4.099796295166016, 1535961891.6806257], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6091, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004484145267857143], 0, 0.812124490737915, 1535962420.3919532], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1292, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [4, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.00025928030569948187], 0, 2.262220621109009, 1535962838.8832164], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5859, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003971391746031746], 0, 7.697132349014282, 1535963452.9493062], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14151, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006106042235294117], 0, 3.5176310539245605, 1535964560.6628072], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30147, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002976758882352941], 0, 0.8086378574371338, 1535964988.2985084], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31835, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0029329735142857145], 0, 0.8191285133361816, 1535965597.717552], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30435, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026061367435897434], 0, 1.5136566162109375, 1535966101.0201373], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11620, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022368555333333335], 0, 2.2662553787231445, 1535966675.2630477], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31299, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022873868], 0, 0.7781345844268799, 1535967162.4006915], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13443, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001994288392156863], 0, 2.5566022396087646, 1535967728.3448937], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12867, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017639758771929824], 0, 0.7674558162689209, 1535968246.0435636], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 32163, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015642718125], 0, 0.7397537231445312, 1535969048.7722178], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13731, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013989794583333334], 0, 0.839015007019043, 1535969844.125493], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12291, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001175374395348837], 0, 0.7773828506469727, 1535970394.6976082], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31299, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009907440784313725], 0, 0.7486202716827393, 1535970847.9451635], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 797, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [8, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0009663617788461538], 0, 3.552513837814331, 1535971315.763483], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 30147, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007712635954198473], 0, 3.1421215534210205, 1535972363.6402466], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22339, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0066678166], 0, 1.5053436756134033, 1535972844.4946856], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 44235, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00596304594117647], 0, 1.6008901596069336, 1535973814.1225545], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8836, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004498021304347826], 0, 0.7791235446929932, 1535974217.553452], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 43211, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00430811025], 0, 0.7888033390045166, 1535974853.8475828], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54083, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025212072750000003], 0, 0.7554500102996826, 1535975355.3463664], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 642, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 8]], ["tile_c", "sp", [128, 1]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["none", "vec"]]], "t": "winograd"}], "r": [[0.004188739541666667], 0, 3.201399326324463, 1535975877.2563932], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25049, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016959888644067797], 0, 1.5403258800506592, 1535976640.6686215], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18244, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [1, 8]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003957033653846154], 0, 1.5361175537109375, 1535977213.0488367], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4293, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0034360263666666667], 0, 4.060506820678711, 1535978029.85694], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23704, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028549380833333333], 0, 4.766369819641113, 1535978878.776696], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24377, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010665976947368421], 0, 5.131404876708984, 1535979710.8643892], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1988, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [32, 12]], ["tile_c", "sp", [32, 14]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.002397347714285714], 0, 5.791667222976685, 1535980428.2624853], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18958, "c": null, "e": [["tile_co", "sp", [112, 4]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00250306665], 0, 0.8071002960205078, 1535981256.144897], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 27763, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [1, 8]], ["tile_ow", "sp", [4, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013564043289473685], 0, 9.21124815940857, 1535982519.389711], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 22213, "c": null, "e": [["tile_co", "sp", [48, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001288959794871795], 0, 10.213921308517456, 1535983435.9151478], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26835, "c": null, "e": [["tile_co", "sp", [96, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002124398612244898], 0, 1.9007179737091064, 1535984360.0390837], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24110, "c": null, "e": [["tile_co", "sp", [80, 4]], ["tile_oh", "sp", [2, 4]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017948513749999999], 0, 0.7613134384155273, 1535985469.1000557], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24769, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001159165], 0, 0.939647912979126, 1535986292.9722545], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24140, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001731463896551724], 0, 0.8111846446990967, 1535987385.798208], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 7910, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.0062763059375], 0, 10.834023475646973, 1535988419.320336], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 2828, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "vec", "none"]]], "t": "direct"}], "r": [[0.005759601277777778], 0, 5.599284887313843, 1535989656.7510784], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3556, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.004752486272727272], 0, 2.1798794269561768, 1535990993.3723483], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 6396, "c": null, "e": [["tile_co", "sp", [160, 1]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "none", "none"]]], "t": "direct"}], "r": [[0.003950773615384615], 0, 0.6831107139587402, 1535992313.8172636], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 8022, "c": null, "e": [["tile_co", "sp", [192, 1]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.004773958809523809], 0, 2.7595438957214355, 1535993145.9615872], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 5688, "c": null, "e": [["tile_co", "sp", [160, 1]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "t": "direct"}], "r": [[0.003786443296296296], 0, 0.7469737529754639, 1535993986.654902], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4060, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004338464], 0, 0.7445175647735596, 1535994916.3777652], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 7141, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0034488545], 0, 0.7720198631286621, 1535996179.6438222], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 5001, "c": null, "e": [["tile_co", "sp", [64, 2]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0024576920731707317], 0, 0.7179040908813477, 1535997264.1533923], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 8751, "c": null, "e": [["tile_co", "sp", [96, 2]], ["tile_oh", "sp", [1, 17]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.003503288793103448], 0, 2.6482205390930176, 1535998196.78333], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 3953, "c": null, "e": [["tile_co", "sp", [64, 2]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.002310253869565217], 0, 2.9624388217926025, 1535999597.0328796], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2804, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00331382846875], 0, 1.5189454555511475, 1536000446.0136447], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1155, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [1, 17]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.004685320090909091], 0, 46.51941108703613, 1536001319.828308], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5093, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001938968185185185], 0, 1.504150390625, 1536002081.8102407], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1481, "c": null, "e": [["tile_co", "sp", [12, 32]], ["tile_oh", "sp", [17, 1]], ["tile_ow", "sp", [17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.03355650075], 0, 5.259204864501953, 1536003322.6832743], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13333, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0013345612533333334], 0, 0.9124705791473389, 1536004568.2724583], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9804, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001694917186440678], 0, 2.9813454151153564, 1536005426.2075477], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17363, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012870453974358975], 0, 1.4026844501495361, 1536006391.7072964], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1964, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0015260552272727272], 0, 2.2616052627563477, 1536007383.934484], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10304, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006232327826086957], 0, 0.926898717880249, 1536008493.0196972], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3016, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [12, 8]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0028510896944444445], 0, 2.379835605621338, 1536009395.5150921], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1888, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [27, 3]], ["tile_k", "sp", [12, 8]], ["tile_c", "sp", [8, 8]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.00206633458], 0, 2.381316661834717, 1536010099.009254], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 11931, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004905352666666666], 0, 0.7998201847076416, 1536010933.7697515], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13683, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [35, 1]], ["tile_ow", "sp", [5, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009398793738317757], 0, 0.7745063304901123, 1536011673.5613468], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9669, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [7, 5]], ["tile_ow", "sp", [7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011259902666666667], 0, 0.9310727119445801, 1536012430.978725], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1195, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [1, 71]], ["tile_ow", "sp", [71, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.08028360175], 0, 29.017651081085205, 1536014404.113639], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4324, "c": null, "e": [["tile_co", "sp", [10, 8]], ["tile_oh", "sp", [73, 1]], ["tile_ow", "sp", [73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0032777725161290323], 0, 4.3437371253967285, 1536015732.9121451], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11644, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [49, 3]], ["tile_ow", "sp", [49, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0218599188], 0, 9.608185529708862, 1536016567.452563], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5043, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [147, 1]], ["tile_ow", "sp", [21, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.011463942111111111], 0, 4.409928560256958, 1536017613.2931557], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1995, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [149, 1]], ["tile_ow", "sp", [149, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0023340378863636363], 0, 0.715843915939331, 1536018349.7271018], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 660, "c": null, "e": [["tile_co", "sp", [125, 8]], ["tile_oh", "sp", [1, 1]], ["tile_ow", "sp", [1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003419902218430034], 0, 7.9540696144104, 1536019451.7778199], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1392, "c": null, "e": [["tile_co", "sp", [80, 16]], ["tile_oh", "sp", [1, 7]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001382412383561644], 0, 0.7335708141326904, 1536019849.6141517], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1096, "c": null, "e": [["tile_co", "sp", [40, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0011404255617977529], 0, 0.714956521987915, 1536020656.7628424], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 938, "c": null, "e": [["tile_co", "sp", [40, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.000644472923566879], 0, 6.011553049087524, 1536021494.4979725], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1494, "c": null, "e": [["tile_c", "sp", [60, 16]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[7.947598967434472e-05], 0, 0.8504102230072021, 1536022065.737844], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9355, "c": null, "e": [["tile_co", "sp", [240, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00046429763181818183], 0, 1.4487085342407227, 1536023055.4891448], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1948, "c": null, "e": [["tile_co", "sp", [20, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00034875311846689893], 0, 0.701000452041626, 1536024161.773847], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 866, "c": null, "e": [["tile_c", "sp", [72, 8]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[8.410173017170892e-05], 0, 0.8886661529541016, 1536024852.1844578], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8021, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006683209], 0, 10.230619192123413, 1536026092.4883912], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4434, "c": null, "e": [["tile_c", "sp", [144, 4]], ["tile_h", "sp", [1, 14]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0001803295963963964], 0, 0.8243341445922852, 1536027098.249329], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15041, "c": null, "e": [["tile_co", "sp", [72, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006251896335403727], 0, 0.7476546764373779, 1536028156.0684233], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8309, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005497778723404256], 0, 0.667487382888794, 1536028853.4242327], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9474, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003175000316455696], 0, 2.5827298164367676, 1536029834.974365], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3365, "c": null, "e": [["tile_c", "sp", [48, 8]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00011543759608745685], 0, 0.8591289520263672, 1536030780.8958592], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27735, "c": null, "e": [["tile_co", "sp", [24, 16]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00026018414545454544], 0, 1.4775424003601074, 1536031947.9872396], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5127, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [1, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00016132788870967742], 0, 0.7692306041717529, 1536032790.1603823], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2257, "c": null, "e": [["tile_c", "sp", [48, 4]], ["tile_h", "sp", [7, 2]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0001032701909650924], 0, 9.397353649139404, 1536033568.4576824], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23475, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003465056], 0, 7.670989513397217, 1536034921.8032045], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 6613, "c": null, "e": [["tile_c", "sp", [24, 8]], ["tile_h", "sp", [2, 14]], ["tile_w", "sp", [28, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00022719872505543238], 0, 0.8938412666320801, 1536035759.7288659], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21847, "c": null, "e": [["tile_co", "sp", [12, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00027976529085872574], 0, 0.7315948009490967, 1536037102.8543599], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23475, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00026518273015873017], 0, 0.7950952053070068, 1536038242.0583954], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 7158, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [4, 7]], ["tile_w", "sp", [14, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00030752623312883435], 0, 0.7969005107879639, 1536039060.4195528], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 55107, "c": null, "e": [["tile_co", "sp", [6, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011874045294117646], 0, 2.205906629562378, 1536039976.9889479], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 12633, "c": null, "e": [["tile_c", "sp", [36, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007591801590909091], 0, 0.7970850467681885, 1536040501.8343825], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 102365, "c": null, "e": [["tile_co", "sp", [18, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008861643628318585], 0, 8.273169755935669, 1536042257.2534866], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54083, "c": null, "e": [["tile_co", "sp", [6, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [4, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007908973858267716], 0, 1.4867935180664062, 1536042776.9733043], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 10047, "c": null, "e": [["tile_c", "sp", [24, 4]], ["tile_h", "sp", [4, 14]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007126467887323943], 0, 8.059999465942383, 1536043678.891334], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 52935, "c": null, "e": [["tile_co", "sp", [24, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001961125607843137], 0, 8.941946506500244, 1536044506.7819371], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 52302, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [7, 16]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007323559854014598], 0, 3.670271873474121, 1536045333.9466982], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25562, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [7, 16]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001290033088607595], 0, 1.4837305545806885, 1536046845.941301], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4866, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003595197535714286], 0, 7.988141059875488, 1536047717.4362721], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 455, "c": null, "e": [["tile_c", "sp", [64, 16]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[8.416339243697479e-05], 0, 0.837181568145752, 1536048525.7668655], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4822, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001790993482142857], 0, 4.5119569301605225, 1536049440.6071227], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 532, "c": null, "e": [["tile_c", "sp", [128, 4]], ["tile_h", "sp", [1, 7]], ["tile_w", "sp", [7, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[7.268952906976744e-05], 0, 1.4634675979614258, 1536050072.0866299], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6813, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003054424212121212], 0, 0.7896997928619385, 1536050978.7675412], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1624, "c": null, "e": [["tile_c", "sp", [32, 16]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [14, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00016730340166666666], 0, 0.9480326175689697, 1536051693.3728569], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13663, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [2, 7]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0016121931969696971], 0, 1.7503433227539062, 1536052865.078714], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1496, "c": null, "e": [["tile_c", "sp", [64, 4]], ["tile_h", "sp", [2, 7]], ["tile_w", "sp", [7, 2]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0001378155013661202], 0, 0.8744897842407227, 1536053571.8750489], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15447, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0033389691], 0, 1.4921560287475586, 1536054475.6807044], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3261, "c": null, "e": [["tile_c", "sp", [32, 8]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [28, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0003167326265822785], 0, 1.4690158367156982, 1536055117.3643768], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 35535, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016242830153846155], 0, 6.273205041885376, 1536056066.076082], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3762, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [7, 4]], ["tile_w", "sp", [28, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002757388715846994], 0, 2.9622366428375244, 1536056697.2431805], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10500, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [7, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0033952017333333333], 0, 0.7387659549713135, 1536057248.632653], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 5154, "c": null, "e": [["tile_c", "sp", [32, 4]], ["tile_h", "sp", [7, 8]], ["tile_w", "sp", [56, 1]], ["ann", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006799153466666666], 0, 3.4481821060180664, 1536058048.4736161], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21636, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016133150161290323], 0, 0.72365403175354, 1536058950.2996428], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 5896, "c": null, "e": [["tile_c", "sp", [16, 4]], ["tile_h", "sp", [14, 4]], ["tile_w", "sp", [28, 2]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00046885758796296297], 0, 0.8069319725036621, 1536059633.62642], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15544, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [112, 1]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0021159040833333333], 0, 0.7083666324615479, 1536060448.1761312], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 7837, "c": null, "e": [["tile_c", "sp", [16, 2]], ["tile_h", "sp", [7, 16]], ["tile_w", "sp", [112, 1]], ["ann", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000738090347826087], 0, 0.8959550857543945, 1536061198.4866538], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 13928, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0008192095121951219], 0, 0.9119460582733154, 1536062349.3075526], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2341, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00029580171386430675], 0, 0.7234361171722412, 1536063113.8110507], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 823, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0002619901020942408], 0, 0.7018110752105713, 1536064146.437332], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4903, "c": null, "e": [["tile_co", "sp", [48, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006805546352201257], 0, 15.530133247375488, 1536066006.759754], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1041, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0001030567248459959], 0, 7.316335916519165, 1536067128.7199745], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4423, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00017060078777589134], 0, 4.35484766960144, 1536068170.3420224], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10323, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0002564866717948718], 0, 0.7209694385528564, 1536069324.1666481], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3642, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003258745928338762], 0, 0.8143470287322998, 1536070072.5710378], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 14080, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 2]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.001591630203125], 0, 0.824331521987915, 1536070848.1415775], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5346, "c": null, "e": [["tile_co", "sp", [250, 4]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.006333343176470588], 0, 2.593400239944458, 1536071585.850969], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1293, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.000706720986111111], 0, 3.076237440109253, 1536072758.8508039], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3153, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00017383130729166666], 0, 1.4083540439605713, 1536073630.053251], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1193, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [13, 1]], ["tile_ow", "sp", [1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003926916235294118], 0, 0.7123100757598877, 1536074407.5020707], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 798, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 1]], ["tile_k", "sp", [16, 16]], ["tile_c", "sp", [4, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.0029866654857142855], 0, 7.6771934032440186, 1536075353.6672528], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12280, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000649594896103896], 0, 2.0672879219055176, 1536076363.577217], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12155, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00141263438028169], 0, 7.622973442077637, 1536077855.1415608], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17373, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [9, 3]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011555885977011494], 0, 13.869017601013184, 1536078547.5583801], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 19381, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002675782846153846], 0, 0.8574180603027344, 1536079966.455821], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4317, "c": null, "e": [["tile_co", "sp", [24, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0004206574605809128], 0, 2.891443967819214, 1536081166.8564], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6483, "c": null, "e": [["tile_co", "sp", [12, 4]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007450000740740741], 0, 8.193702220916748, 1536082017.2788978], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5955, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011495224772727272], 0, 0.9198944568634033, 1536083212.3625515], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10916, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00017084748122866893], 0, 3.5794410705566406, 1536084563.5972736], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10419, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [27, 1]], ["tile_ow", "sp", [3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005322740793650794], 0, 0.7264156341552734, 1536085145.7893326], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 14274, "c": null, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005642460055555556], 0, 0.9713246822357178, 1536085944.0637481], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2468, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0007759262713178294], 0, 0.7431638240814209, 1536086847.995349], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3987, "c": null, "e": [["tile_co", "sp", [4, 8]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00100951886], 0, 1.498164176940918, 1536097676.4999], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3242, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006660534569536424], 0, 0.7463428974151611, 1536098375.0879269], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 12378, "c": null, "e": [["tile_co", "sp", [16, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [5, 11]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014504417246376813], 0, 6.887804269790649, 1536099631.9051716], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1824, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.00021897342764578832], 0, 0.7701969146728516, 1536100591.7201855], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3222, "c": null, "e": [["tile_co", "sp", [4, 4]], ["tile_oh", "sp", [55, 1]], ["tile_ow", "sp", [11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005445233804347826], 0, 0.7250428199768066, 1536101330.403957], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 50549, "c": null, "e": [["tile_co", "sp", [12, 8]], ["tile_oh", "sp", [28, 4]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008715647583333333], 0, 3.316070079803467, 1536102375.0753405], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1318, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [64, 8]], ["tile_c", "sp", [128, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.010575870800000001], 0, 11.251509428024292, 1536103135.0365064], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 795, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [128, 4]], ["tile_c", "sp", [32, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.04815382], 0, 3.5421621799468994, 1536104048.8652966], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 795, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [128, 4]], ["tile_c", "sp", [16, 16]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.020693179], 0, 16.602781772613525, 1536105412.3545794], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1262, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [64, 4]], ["tile_c", "sp", [256, 1]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.04217726975], 0, 15.700594425201416, 1536106393.0490034], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1292, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.021821543800000003], 0, 4.234783887863159, 1536107547.1928937], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1810, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [98, 8]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [64, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.04427553825], 0, 5.108673572540283, 1536108599.9776504], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1810, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [98, 8]], ["tile_k", "sp", [16, 8]], ["tile_c", "sp", [32, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0246804192], 0, 2.633969783782959, 1536109586.8460913], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1843, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [784, 4]], ["tile_k", "sp", [8, 8]], ["tile_c", "sp", [16, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.06368745425], 0, 8.785919904708862, 1536110688.884073], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 43599, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [224, 1]], ["tile_ow", "sp", [32, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005805616888888889], 0, 0.8277938365936279, 1536112314.9338953], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4184, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000358496875], 0, 0.8179452419281006, 1536113018.2201066], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 863, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003176745125], 0, 5.019597053527832, 1536114385.0759907], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5836, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003304881862745098], 0, 0.7812585830688477, 1536115286.897814], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 12180, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0028490410833333334], 0, 0.9238848686218262, 1536115960.9596996], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13396, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003732395037313433], 0, 2.1984589099884033, 1536116943.5517302], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24011, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003117249363636364], 0, 0.8778603076934814, 1536117930.977674], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1824, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035912057142857144], 0, 8.9507896900177, 1536119050.9417675], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 844, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.003615182642857143], 0, 2.324625015258789, 1536119908.7735555], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 796, "c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.006235870705882353], 0, 7.047893524169922, 1536120631.9302435], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 794, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [1, 4]], ["tile_k", "sp", [32, 16]], ["tile_c", "sp", [128, 4]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0038201894444444442], 0, 3.5701382160186768, 1536121430.8213255], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1864, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001554832676923077], 0, 3.2758054733276367, 1536122249.4893742], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6852, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0029987334411764703], 0, 6.5263512134552, 1536123533.6109035], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18934, "c": null, "e": [["tile_co", "sp", [128, 8]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0030692616666666665], 0, 1.4980154037475586, 1536124380.903211], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7627, "c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005740417222222222], 0, 11.873262643814087, 1536125633.8073392], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1293, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [2, 8]], ["tile_k", "sp", [32, 8]], ["tile_c", "sp", [128, 2]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0027190230263157895], 0, 2.55869197845459, 1536126541.0445106], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12172, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0013052278205128206], 0, 0.7870054244995117, 1536127468.6283941], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12003, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00336386428125], 0, 0.7970542907714844, 1536128258.7289355], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17113, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00317720734375], 0, 6.9983744621276855, 1536129401.310512], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16803, "c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [2, 14]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0064564818125], 0, 4.580580234527588, 1536130443.8960125], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 775, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [7, 7]], ["tile_k", "sp", [32, 4]], ["tile_c", "sp", [32, 4]], ["ann_reduce", "an", ["unroll"]], ["ann_spatial", "an", ["vec", "unroll"]]], "t": "winograd"}], "r": [[0.003081264515151515], 0, 2.786720037460327, 1536131175.868099], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12244, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018306837678571427], 0, 1.5416374206542969, 1536131981.988531], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 49962, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [7, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003746728777777778], 0, 7.905523061752319, 1536133079.3056982], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26644, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00314575065625], 0, 0.7575502395629883, 1536134028.532529], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1347, "c": null, "e": [["tile_size", "ot", 4], ["tile_p", "sp", [49, 4]], ["tile_k", "sp", [4, 16]], ["tile_c", "sp", [8, 8]], ["ann_reduce", "an", ["none"]], ["ann_spatial", "an", ["unroll", "vec"]]], "t": "winograd"}], "r": [[0.0034228753333333333], 0, 3.2874855995178223, 1536134679.4625976], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=snapdragon835 -target=arm64-linux-android -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 49347, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008544545338983051], 0, 0.7857861518859863, 1536135556.627556], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "int8"], ["TENSOR", [64, 3, 7, 7], "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 3, 224, 224, "int8"], [64, 3, 7, 7, "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {"i": 72880, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [56, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.105920721], 0, 0.729074239730835, 1533135908.2729535], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"i": 9467, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.1141719645], 0, 0.752866268157959, 1533136336.6250343], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"i": 17980, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0135751615], 0, 0.43703341484069824, 1533136851.4134429], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [128, 64, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [128, 64, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"i": 24531, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.062420649], 0, 0.6215760707855225, 1533137098.0733955], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [128, 64, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [128, 64, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"i": 12436, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0065981785], 0, 0.4342308044433594, 1533137426.0706506], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [128, 128, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [128, 128, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"i": 4947, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.1117400415], 0, 0.778343677520752, 1533137683.4026408], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [256, 128, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [256, 128, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"i": 6519, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0603407095], 0, 0.6446278095245361, 1533138053.628069], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [256, 128, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [256, 128, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"i": 15448, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0058095355], 0, 0.4593832492828369, 1533138393.4752011], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [256, 256, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [256, 256, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"i": 2487, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.11036658], 0, 0.7984218597412109, 1533138719.775083], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [512, 256, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [512, 256, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"i": 3222, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.073809475], 0, 0.6738097667694092, 1533139197.6142242], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [512, 256, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [512, 256, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"i": 3444, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.006798731], 0, 0.47246456146240234, 1533139549.1411946], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "int8"], ["TENSOR", [512, 512, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 7, 7, "int8"], [512, 512, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"i": 1622, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.107875749], 0, 0.7853262424468994, 1533139813.8459315], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 72880, "c": null, "e": [["tile_co", "sp", [8, 4]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [56, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.105920721], 0, 0.729074239730835, 1533135908.2729535], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 9467, "c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.1141719645], 0, 0.752866268157959, 1533136336.6250343], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 17980, "c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0135751615], 0, 0.43703341484069824, 1533136851.4134429], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24531, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.062420649], 0, 0.6215760707855225, 1533137098.0733955], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12436, "c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0065981785], 0, 0.4342308044433594, 1533137426.0706506], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4947, "c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.1117400415], 0, 0.778343677520752, 1533137683.4026408], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 6519, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0603407095], 0, 0.6446278095245361, 1533138053.628069], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15448, "c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0058095355], 0, 0.4593832492828369, 1533138393.4752011], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2487, "c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.11036658], 0, 0.7984218597412109, 1533138719.775083], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3222, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.073809475], 0, 0.6738097667694092, 1533139197.6142242], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3444, "c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.006798731], 0, 0.47246456146240234, 1533139549.1411946], "v": 0.1}
+{"i": ["llvm -device=arm_cpu -model=vtacpu -target=armv7a-linux-eabi", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1622, "c": null, "e": [["tile_co", "sp", [128, 4]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.107875749], 0, 0.7853262424468994, 1533139813.8459315], "v": 0.1}
+# int8/fp32 for dcgan and int8 for resnet on ultra96
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 128, 32, 32], "int8"], ["TENSOR", [128, 3, 4, 4], "int8"], [2, 2], [1, 1], "int32"], {}, ["conv2d_transpose_nchw", [1, 128, 32, 32, "int8"], [128, 3, 4, 4, "int8"], [2, 2], [1, 1], "int32"], {"t": "direct", "i": 11596, "e": [["tile_co", "sp", [3, 1]], ["tile_oh", "sp", [16, 4]], ["tile_ow", "sp", [16, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "c": null}], "r": [[0.007695752153846154], 0, 18.223355054855347, 1542405515.4346437]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 128, 32, 32], "float32"], ["TENSOR", [128, 3, 4, 4], "float32"], [2, 2], [1, 1], "float32"], {}, ["conv2d_transpose_nchw", [1, 128, 32, 32, "float32"], [128, 3, 4, 4, "float32"], [2, 2], [1, 1], "float32"], {"t": "direct", "i": 11498, "e": [["tile_co", "sp", [3, 1]], ["tile_oh", "sp", [16, 4]], ["tile_ow", "sp", [16, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "c": null}], "r": [[0.009704522727272728], 0, 44.45757484436035, 1542406268.4606514]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 256, 16, 16], "int8"], ["TENSOR", [256, 128, 4, 4], "int8"], [2, 2], [1, 1], "int32"], {}, ["conv2d_transpose_nchw", [1, 256, 16, 16, "int8"], [256, 128, 4, 4, "int8"], [2, 2], [1, 1], "int32"], {"t": "direct", "i": 25163, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [16, 2]], ["tile_ow", "sp", [8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "c": null}], "r": [[0.04691446666666666], 0, 36.08643412590027, 1542406774.5246825]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 256, 16, 16], "float32"], ["TENSOR", [256, 128, 4, 4], "float32"], [2, 2], [1, 1], "float32"], {}, ["conv2d_transpose_nchw", [1, 256, 16, 16, "float32"], [256, 128, 4, 4, "float32"], [2, 2], [1, 1], "float32"], {"t": "direct", "i": 30970, "e": [["tile_co", "sp", [32, 4]], ["tile_oh", "sp", [16, 2]], ["tile_ow", "sp", [4, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "c": null}], "r": [[0.07177340633333333], 0, 36.92069888114929, 1542407439.0180466]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 512, 8, 8], "int8"], ["TENSOR", [512, 256, 4, 4], "int8"], [2, 2], [1, 1], "int32"], {}, ["conv2d_transpose_nchw", [1, 512, 8, 8, "int8"], [512, 256, 4, 4, "int8"], [2, 2], [1, 1], "int32"], {"t": "direct", "i": 5277, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [8, 2]], ["tile_ow", "sp", [4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "c": null}], "r": [[0.04968457666666667], 0, 51.032613039016724, 1542407767.037206]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 512, 8, 8], "float32"], ["TENSOR", [512, 256, 4, 4], "float32"], [2, 2], [1, 1], "float32"], {}, ["conv2d_transpose_nchw", [1, 512, 8, 8, "float32"], [512, 256, 4, 4, "float32"], [2, 2], [1, 1], "float32"], {"t": "direct", "i": 24257, "e": [["tile_co", "sp", [64, 4]], ["tile_oh", "sp", [16, 1]], ["tile_ow", "sp", [1, 16]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "c": null}], "r": [[0.090142901], 0, 43.57327890396118, 1542408683.9351382]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 1024, 4, 4], "int8"], ["TENSOR", [1024, 512, 4, 4], "int8"], [2, 2], [1, 1], "int32"], {}, ["conv2d_transpose_nchw", [1, 1024, 4, 4, "int8"], [1024, 512, 4, 4, "int8"], [2, 2], [1, 1], "int32"], {"t": "direct", "i": 3803, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [8, 1]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "c": null}], "r": [[0.06295115433333333], 0, 50.96684241294861, 1542408917.2494385]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d_transpose_nchw", [["TENSOR", [1, 1024, 4, 4], "float32"], ["TENSOR", [1024, 512, 4, 4], "float32"], [2, 2], [1, 1], "float32"], {}, ["conv2d_transpose_nchw", [1, 1024, 4, 4, "float32"], [1024, 512, 4, 4, "float32"], [2, 2], [1, 1], "float32"], {"t": "direct", "i": 13414, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [4, 2]], ["tile_ow", "sp", [1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "c": null}], "r": [[0.158133852], 0, 48.270063638687134, 1542409845.450727]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "int8"], ["TENSOR", [512, 512, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 7, 7, "int8"], [512, 512, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 3463, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "c": null}], "r": [[0.0101057844], 0, 33.18108820915222, 1542405669.9648728]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [512, 256, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [512, 256, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 3304, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "c": null}], "r": [[0.0007376986544117648], 0, 59.29027462005615, 1542406703.149588]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [512, 256, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [512, 256, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 1863, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "c": null}], "r": [[0.005961500647058824], 0, 31.356482982635498, 1542407534.460094]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [256, 256, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [256, 256, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 3243, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "c": null}], "r": [[0.009931650545454545], 0, 40.08094239234924, 1542408549.1960964]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [256, 128, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [256, 128, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 5836, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "c": null}], "r": [[0.00077054506870229], 0, 42.55009865760803, 1542409564.7757497]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [256, 128, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [256, 128, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 6817, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [7, 2]], ["tile_ow", "sp", [7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "c": null}], "r": [[0.005663803388888889], 0, 36.390716314315796, 1542410510.5673108]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [128, 128, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [128, 128, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 32075, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "c": null}], "r": [[0.009820255363636364], 0, 21.711212158203125, 1542411124.5747042]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [128, 64, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [128, 64, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 30964, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "c": null}], "r": [[0.0009000952767857143], 0, 5.5258800983428955, 1542411696.0540035]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [128, 64, 3, 3], "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [128, 64, 3, 3, "int8"], [2, 2], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 13308, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "c": null}], "r": [[0.0061031527647058825], 0, 32.75572371482849, 1542412244.7890527]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 49452, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "c": null}], "r": [[0.0019729396078431373], 0, 33.34551644325256, 1542412866.1647928]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 9978, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "c": null}], "r": [[0.010652710999999999], 0, 17.48495626449585, 1542413395.9526582]}
+{"v": 0.1, "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "int8"], ["TENSOR", [64, 3, 7, 7], "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 3, 224, 224, "int8"], [64, 3, 7, 7, "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {"t": "direct", "i": 29487, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [28, 4]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "c": null}], "r": [[0.0130459995], 0, 12.096744775772095, 1542414453.2667437]}
+{"v": 0.1, "r": [[0.005599333333333334], 0, 20.886932611465454, 1542405538.643406], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "int8"], ["TENSOR", [512, 2048, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 2048, 7, 7, "int8"], [512, 2048, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 1704}]}
+{"v": 0.1, "r": [[0.005618416444444444], 0, 51.378814458847046, 1542406556.8303523], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "int8"], ["TENSOR", [2048, 512, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 7, 7, "int8"], [2048, 512, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 2188}]}
+{"v": 0.1, "r": [[0.011231821888888889], 0, 33.91471076011658, 1542407455.8026795], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "int8"], ["TENSOR", [2048, 1024, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 1024, 14, 14, "int8"], [2048, 1024, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [128, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 1132}]}
+{"v": 0.1, "r": [[0.0100646098], 0, 30.24740719795227, 1542408586.7219546], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "int8"], ["TENSOR", [512, 512, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 7, 7, "int8"], [512, 512, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [64, 8]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 1863}]}
+{"v": 0.1, "r": [[0.0028580094], 0, 15.562670707702637, 1542410185.1138933], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "int8"], ["TENSOR", [512, 1024, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 1024, 14, 14, "int8"], [512, 1024, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [7, 1]], ["tile_ow", "sp", [1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct", "i": 3344}]}
+{"v": 0.1, "r": [[0.005779640277777778], 0, 39.99040937423706, 1542411087.6936805], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "int8"], ["TENSOR", [256, 1024, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 1024, 14, 14, "int8"], [256, 1024, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 3100}]}
+{"v": 0.1, "r": [[0.0057653328333333335], 0, 16.175684213638306, 1542411596.7556822], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [1024, 256, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [1024, 256, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 8012}]}
+{"v": 0.1, "r": [[0.01180704], 0, 1.7591967582702637, 1542412138.3311837], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "int8"], ["TENSOR", [1024, 512, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 28, 28, "int8"], [1024, 512, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [64, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 7484}]}
+{"v": 0.1, "r": [[0.009933850727272727], 0, 2.825345039367676, 1542412603.555186], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "int8"], ["TENSOR", [256, 256, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 14, 14, "int8"], [256, 256, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [32, 8]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct", "i": 12603}]}
+{"v": 0.1, "r": [[0.002977942735294118], 0, 8.401044607162476, 1542413151.3596654], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "int8"], ["TENSOR", [256, 512, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 28, 28, "int8"], [256, 512, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [14, 1]], ["tile_ow", "sp", [2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct", "i": 12172}]}
+{"v": 0.1, "r": [[0.006230185647058823], 0, 20.572247743606567, 1542413704.2630906], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "int8"], ["TENSOR", [128, 512, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 512, 28, 28, "int8"], [128, 512, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct", "i": 24628}]}
+{"v": 0.1, "r": [[0.0064848833125], 0, 18.38726830482483, 1542414173.1151903], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [512, 128, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [512, 128, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 15664}]}
+{"v": 0.1, "r": [[0.013075736625], 0, 16.801048755645752, 1542414616.7927444], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "int8"], ["TENSOR", [512, 256, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 56, 56, "int8"], [512, 256, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [32, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 8464}]}
+{"v": 0.1, "r": [[0.009789951363636364], 0, 0.9385528564453125, 1542415000.349857], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "int8"], ["TENSOR", [128, 128, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 128, 28, 28, "int8"], [128, 128, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [16, 8]], ["tile_oh", "sp", [14, 2]], ["tile_ow", "sp", [7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct", "i": 32075}]}
+{"v": 0.1, "r": [[0.0038914148461538464], 0, 9.693146228790283, 1542415351.4980526], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "int8"], ["TENSOR", [128, 256, 1, 1], "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 56, 56, "int8"], [128, 256, 1, 1, "int8"], [2, 2], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [8, 16]], ["tile_oh", "sp", [28, 1]], ["tile_ow", "sp", [4, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 4756}]}
+{"v": 0.1, "r": [[0.007500206285714286], 0, 11.497221231460571, 1542415662.669843], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "int8"], ["TENSOR", [64, 256, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 256, 56, 56, "int8"], [64, 256, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct", "i": 49452}]}
+{"v": 0.1, "r": [[0.007966708076923077], 0, 10.025586128234863, 1542415988.2491486], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [256, 64, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [256, 64, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [16, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct", "i": 26140}]}
+{"v": 0.1, "r": [[0.0106324581], 0, 5.028797149658203, 1542416306.2014444], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 3, 3], "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 3, 3, "int8"], [1, 1], [1, 1], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [28, 2]], ["tile_ow", "sp", [14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 9978}]}
+{"v": 0.1, "r": [[0.0019684391568627452], 0, 0.8258552551269531, 1542416660.841285], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "int8"], ["TENSOR", [64, 64, 1, 1], "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 64, 56, 56, "int8"], [64, 64, 1, 1, "int8"], [1, 1], [0, 0], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [4, 16]], ["tile_oh", "sp", [56, 1]], ["tile_ow", "sp", [8, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct", "i": 10028}]}
+{"v": 0.1, "r": [[0.01306751625], 0, 4.587785959243774, 1542417231.9395165], "i": ["llvm -device=arm_cpu -model=ultra96 -target=aarch64-linux-gnu -mattr=+neon", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "int8"], ["TENSOR", [64, 3, 7, 7], "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {}, ["conv2d", [1, 3, 224, 224, "int8"], [64, 3, 7, 7, "int8"], [2, 2], [3, 3], [1, 1], "NCHW", "int32"], {"c": null, "e": [["tile_co", "sp", [8, 8]], ["tile_oh", "sp", [28, 4]], ["tile_ow", "sp", [56, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct", "i": 57487}]}
diff --git a/tophub/cuda_v0.05.log b/tophub/cuda_v0.05.log
new file mode 100644
index 0000000..108aed3
--- /dev/null
+++ b/tophub/cuda_v0.05.log
@@ -0,0 +1,485 @@
+# This is the pre-tuned parameters for cuda backend
+# TVM downloaded this during compilation
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8797896, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [64, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.69732536806342e-05], 0, 14.116215467453003, 1535423916.7184713], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 734204, "c": null, "e": [["tile_f", "sp", [28, 2, 8, 1]], ["tile_y", "sp", [2, 2, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [64, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.882545781556573e-05], 0, 20.854647159576416, 1535425623.6315196], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1240970, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 2]], ["tile_y", "sp", [2, 2, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.448735867117118e-05], 0, 14.069890975952148, 1535426980.678347], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5709779, "c": null, "e": [["tile_f", "sp", [10, 1, 16, 2]], ["tile_y", "sp", [2, 1, 2, 2]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.139676807712908e-05], 0, 3.966081380844116, 1535427983.421136], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13233096, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [40, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.6700240116420085e-05], 0, 3.608333110809326, 1535429000.509115], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 232980, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [6, 1, 32, 2]], ["tile_x", "sp", [1, 1, 4, 1]], ["tile_rc", "sp", [28, 16]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00011755589655172413], 0, 5.878081798553467, 1535431279.4730268], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3731075, "c": null, "e": [["tile_f", "sp", [14, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.2090259632072194e-05], 0, 1.3185603618621826, 1535433032.3829937], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 35416378, "c": null, "e": [["tile_f", "sp", [24, 2, 8, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [16, 24]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.5191243902439025e-05], 0, 1.303452730178833, 1535434456.1185555], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 13970090, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [16, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.3322369696969696e-05], 0, 1.4410252571105957, 1535436450.3439708], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5330090, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.552532625833839e-05], 0, 1.3521144390106201, 1535437568.5237384], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3690668, "c": null, "e": [["tile_f", "sp", [20, 2, 8, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.795118104265403e-05], 0, 1.2909648418426514, 1535438405.1098573], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10006142, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 1]], ["tile_y", "sp", [8, 1, 1, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.094121392190153e-05], 0, 1.2878923416137695, 1535439480.1925032], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 28729475, "c": null, "e": [["tile_f", "sp", [10, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [48, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.494733763996555e-05], 0, 1.3043124675750732, 1535440944.6158926], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 776877, "c": null, "e": [["tile_f", "sp", [8, 4, 6, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00016294033007600435], 0, 2.581367254257202, 1535441832.1955879], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 325452, "c": null, "e": [["tile_f", "sp", [8, 1, 6, 4]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00016267638502673795], 0, 1.3517541885375977, 1535442556.6070833], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 88756, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00014167337005649717], 0, 2.514183282852173, 1535443548.5128298], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 449924, "c": null, "e": [["tile_f", "sp", [4, 1, 10, 4]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011452480775075988], 0, 4.953140497207642, 1535444579.1761086], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 669438, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00014245359474671668], 0, 3.571627140045166, 1535445303.0586505], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 446246, "c": null, "e": [["tile_f", "sp", [5, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011218677481481482], 0, 1.3602855205535889, 1535446031.3722363], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 220455, "c": null, "e": [["tile_f", "sp", [4, 5, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.427650495049505e-05], 0, 1.2197649478912354, 1535446918.1282773], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 61876, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00011426011119573495], 0, 18.739489793777466, 1535447697.4223692], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 129655, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.613525476429288e-05], 0, 8.062903881072998, 1535448397.9567559], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 448948, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011292374210526315], 0, 2.1787657737731934, 1535449108.6007366], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 129623, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.355220323212537e-05], 0, 3.0357697010040283, 1535449805.546825], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 120055, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [32, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.593256546572935e-05], 0, 10.24105978012085, 1535450762.5778947], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 228674, "c": null, "e": [["tile_f", "sp", [8, 2, 3, 4]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [64, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.298381360619469e-05], 0, 1.2305030822753906, 1535452260.8384964], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 227677, "c": null, "e": [["tile_f", "sp", [4, 2, 6, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [24, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.262685946622186e-05], 0, 2.3495943546295166, 1535453153.0011835], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 426376, "c": null, "e": [["tile_f", "sp", [16, 4, 3, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [144, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00046931365937500003], 0, 5.672942876815796, 1535454505.8647752], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2956827, "c": null, "e": [["tile_f", "sp", [2, 6, 4, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [12, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.304847466960352e-05], 0, 9.784206628799438, 1535455827.5935478], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 225833, "c": null, "e": [["tile_f", "sp", [2, 4, 4, 2]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [12, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.915220448617632e-05], 0, 1.400207281112671, 1535456732.6478782], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1129045, "c": null, "e": [["tile_f", "sp", [2, 2, 4, 3]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.0527908595641644e-05], 0, 2.0959362983703613, 1535457623.5808153], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 462356, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.608307039884671e-05], 0, 7.058563470840454, 1535458449.1994894], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 322573, "c": null, "e": [["tile_f", "sp", [2, 4, 4, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [8, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.7956799473873014e-05], 0, 1.3385558128356934, 1535459233.2346888], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1352202, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [6, 2, 2, 4]], ["tile_x", "sp", [1, 3, 108, 1]], ["tile_rc", "sp", [12, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[5.601382798507463e-05], 0, 11.555405378341675, 1535461246.940517], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1901044, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [6, 1, 2, 8]], ["tile_x", "sp", [1, 3, 108, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[4.3254536418166237e-05], 0, 1.3233392238616943, 1535463738.0164495], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 3275017, "c": null, "e": [["tile_f", "sp", [2, 1, 32, 1]], ["tile_y", "sp", [5, 1, 1, 7]], ["tile_x", "sp", [7, 1, 5, 1]], ["tile_rc", "sp", [16, 3]], ["tile_ry", "sp", [1, 5]], ["tile_rx", "sp", [1, 5]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.947820815677967e-05], 0, 1.3620657920837402, 1535466035.195244], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1774165, "c": null, "e": [["tile_f", "sp", [2, 2, 4, 3]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [12, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.3459930018703244e-05], 0, 4.270485162734985, 1535466912.1237135], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1085972, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [8, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.796061236268851e-05], 0, 1.4543447494506836, 1535467856.116219], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 922093, "c": null, "e": [["tile_f", "sp", [4, 8, 3, 2]], ["tile_y", "sp", [71, 1, 1, 1]], ["tile_x", "sp", [1, 1, 71, 1]], ["tile_rc", "sp", [40, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0006553213608695652], 0, 1.5329456329345703, 1535470367.360386], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 70574, "c": null, "e": [["tile_f", "sp", [5, 8, 2, 1]], ["tile_y", "sp", [73, 1, 1, 1]], ["tile_x", "sp", [1, 1, 73, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.883320232858991e-05], 0, 1.5573954582214355, 1535471271.1996536], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 8943001, "c": null, "e": [["tile_f", "sp", [2, 1, 32, 1]], ["tile_y", "sp", [21, 1, 1, 7]], ["tile_x", "sp", [7, 1, 7, 3]], ["tile_rc", "sp", [32, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00025052284166666663], 0, 7.156659126281738, 1535472591.5201485], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6050372, "c": null, "e": [["tile_f", "sp", [1, 1, 32, 1]], ["tile_y", "sp", [21, 7, 1, 1]], ["tile_x", "sp", [7, 1, 7, 3]], ["tile_rc", "sp", [16, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00012198324857839155], 0, 3.08567214012146, 1535473653.6231658], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23791, "c": null, "e": [["tile_f", "sp", [2, 2, 1, 8]], ["tile_y", "sp", [149, 1, 1, 1]], ["tile_x", "sp", [1, 1, 149, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.216306946826758e-05], 0, 5.255294322967529, 1535474702.044478], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 117516, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 4, 16, 1]], ["tile_x", "sp", [1, 7, 7, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00018763357], 0, 1.6003353595733643, 1535405501.9874778], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1172502, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 2, 8, 4]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0005314307929824561], 0, 15.780593156814575, 1535406995.4651527], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1062454, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [16, 4, 4, 2]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0002953229862475442], 0, 1.9394707679748535, 1535408976.5928106], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 796885, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [4, 2, 8, 4]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00037394815158924206], 0, 3.6207523345947266, 1535410469.9387088], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 714385, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [4, 2, 8, 4]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00023493554432348364], 0, 4.635826826095581, 1535411899.2398431], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1813376, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 8, 2]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00042691407670454546], 0, 19.30599021911621, 1535414354.3659015], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1603376, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 8, 2]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0002890588754789272], 0, 3.4835450649261475, 1535415894.133812], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2195318, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 8, 4]], ["tile_x", "sp", [49, 4, 16, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0006257575269709544], 0, 5.687942981719971, 1535418141.8687968], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 59919263, "c": null, "e": [["tile_f", "sp", [1, 2, 16, 2]], ["tile_y", "sp", [56, 1, 1, 4]], ["tile_x", "sp", [7, 2, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.754253240324032e-05], 0, 1.5378711223602295, 1535422580.7674332], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 58555, "c": null, "e": [["tile_f", "sp", [16, 2, 16, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00015904646927966104], 0, 1.3764936923980713, 1535340979.3568566], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 143095, "c": null, "e": [["tile_f", "sp", [64, 2, 16, 1]], ["tile_y", "sp", [1, 7, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.65306638061585e-05], 0, 2.2808291912078857, 1535342441.0873826], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 335004, "c": null, "e": [["tile_f", "sp", [16, 2, 32, 2]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [256, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00028060819402985073], 0, 1.6567697525024414, 1535343548.7764354], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 132040, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.38442889305816e-05], 0, 1.886913776397705, 1535346388.4123077], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2521885, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.780090051813472e-05], 0, 2.096433401107788, 1535347370.988212], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2965852, "c": null, "e": [["tile_f", "sp", [32, 4, 8, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.840749328644501e-05], 0, 34.52979779243469, 1535349067.3907256], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1723190, "c": null, "e": [["tile_f", "sp", [16, 4, 16, 1]], ["tile_y", "sp", [7, 1, 1, 2]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00011870797707509881], 0, 6.502109050750732, 1535351117.3144069], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1795299, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 4]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [128, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.030473995983936e-05], 0, 14.208871603012085, 1535355244.8065631], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6652823, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.824087160589604e-05], 0, 8.829635620117188, 1535357359.3047445], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4022069, "c": null, "e": [["tile_f", "sp", [16, 4, 8, 1]], ["tile_y", "sp", [7, 4, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.836637116564418e-05], 0, 30.88849425315857, 1535359778.4165032], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13520886, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.964673237753884e-05], 0, 21.609076023101807, 1535362579.700074], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7564855, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.5890965484408476e-05], 0, 7.409880638122559, 1535366580.0144885], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6625940, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [56, 1, 1, 1]], ["tile_x", "sp", [1, 4, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.789036865624003e-05], 0, 1.2796452045440674, 1535367873.0130634], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3542617, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 2]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.3173960469779435e-05], 0, 19.429253578186035, 1535370537.716797], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 222539, "c": null, "e": [["tile_f", "sp", [64, 2, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00012062812861736334], 0, 1.9216992855072021, 1535376563.6045246], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 25761, "c": null, "e": [["tile_f", "sp", [256, 1, 4, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.3759017449004675e-06], 0, 4.439143419265747, 1535377557.3893585], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 108194, "c": null, "e": [["tile_f", "sp", [64, 2, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.706907663125948e-05], 0, 1.3412449359893799, 1535379150.708214], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 19810, "c": null, "e": [["tile_f", "sp", [256, 1, 2, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.498452183395762e-06], 0, 1.8959953784942627, 1535380034.9811802], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 771840, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.098708732737612e-05], 0, 2.6358609199523926, 1535381939.087032], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 310860, "c": null, "e": [["tile_f", "sp", [512, 1, 1, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.404539118416505e-06], 0, 1.145503282546997, 1535383090.2789602], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 715461, "c": null, "e": [["tile_f", "sp", [32, 4, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.3200288139329807e-05], 0, 5.702603101730347, 1535384171.1888306], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 191400, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [1, 1, 14, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.514348936170213e-06], 0, 1.3499267101287842, 1535385287.8302827], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3339832, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 2]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.51544672033393e-05], 0, 1.2489631175994873, 1535386621.2483017], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1172160, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [2, 1, 7, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.0338754319654428e-05], 0, 5.082122564315796, 1535387899.0255547], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5190596, "c": null, "e": [["tile_f", "sp", [8, 4, 8, 1]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.6632782280701753e-05], 0, 3.803677797317505, 1535390486.2065094], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 894840, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [1, 1, 28, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.2479111885844008e-05], 0, 8.535689115524292, 1535391579.1142876], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4126080, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [2, 1, 4, 7]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.559869505094614e-05], 0, 3.6279289722442627, 1535395987.3606555], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7948975, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.3504540112817302e-05], 0, 1.3222031593322754, 1535397725.9387324], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2886240, "c": null, "e": [["tile_f", "sp", [64, 1, 1, 1]], ["tile_y", "sp", [7, 1, 4, 2]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.393993827751196e-05], 0, 1.3483922481536865, 1535398924.3910995], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 23485665, "c": null, "e": [["tile_f", "sp", [1, 8, 8, 1]], ["tile_y", "sp", [28, 2, 2, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [8, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.6321584531743247e-05], 0, 1.2850267887115479, 1535400525.906705], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4740848, "c": null, "e": [["tile_f", "sp", [32, 1, 1, 1]], ["tile_y", "sp", [8, 1, 2, 7]], ["tile_x", "sp", [1, 1, 112, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5317467108618054e-05], 0, 1.272824764251709, 1535401884.0289147], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 41999093, "c": null, "e": [["tile_f", "sp", [1, 1, 8, 4]], ["tile_y", "sp", [16, 7, 1, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.3977960968793665e-05], 0, 1.3667681217193604, 1535403596.0402923], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 344206, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 1, 16, 4]], ["tile_x", "sp", [1, 1, 8, 2]], ["tile_rc", "sp", [32, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[8.610198970251716e-05], 0, 1.4480016231536865, 1535511176.105762], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 110915, "c": null, "e": [["tile_f", "sp", [16, 2, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.729347361717613e-05], 0, 1.236013412475586, 1535512971.3886647], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 736155, "c": null, "e": [["tile_f", "sp", [16, 2, 16, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [7, 1, 1, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00016977524406779662], 0, 2.0983633995056152, 1535514397.926689], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 66732, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 2, 16, 2]], ["tile_x", "sp", [1, 7, 7, 1]], ["tile_rc", "sp", [16, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[7.205142459173871e-05], 0, 1.4618885517120361, 1535515568.3814023], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 528026, "c": null, "e": [["tile_f", "sp", [8, 4, 8, 1]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [8, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.621085209760274e-05], 0, 10.320522785186768, 1535516689.8571618], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5183051, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.29051653250774e-05], 0, 1.2669603824615479, 1535518191.0489676], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 520156, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 2, 4, 4]], ["tile_x", "sp", [1, 2, 98, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[6.1538798195242e-05], 0, 2.2867848873138428, 1535519772.1041968], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1996855, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.2646983313669309e-05], 0, 1.2524137496948242, 1535521080.8447835], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 31178187, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [2, 1, 14, 1]], ["tile_rc", "sp", [32, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.736485554060117e-05], 0, 1.3498356342315674, 1535522336.8640018], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13628176, "c": null, "e": [["tile_f", "sp", [2, 8, 4, 1]], ["tile_y", "sp", [56, 1, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5144373877056628e-05], 0, 4.291016101837158, 1535524033.4773188], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 263327, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 1, 4, 8]], ["tile_x", "sp", [1, 2, 98, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[6.067211160533765e-05], 0, 2.9739716053009033, 1535525609.763939], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 100826, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [62, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.591533125897558e-05], 0, 1.1918675899505615, 1535541649.1359475], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 73946, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [60, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.460836556169429e-05], 0, 18.39060950279236, 1535542453.8870785], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 77786, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.402473731477324e-05], 0, 1.248830795288086, 1535543187.1269765], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 102746, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 14]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.9833012105576504e-05], 0, 1.2429931163787842, 1535543904.137366], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 200666, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [54, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.1719856631578944e-05], 0, 10.93128776550293, 1535544717.37846], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10586, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.109051445683602e-05], 0, 1.3161144256591797, 1535545449.2476206], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16346, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 25]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.168217405126168e-05], 0, 2.2161407470703125, 1535546141.5243394], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14426, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.887415226179018e-05], 0, 1.545259714126587, 1535546886.594689], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10581, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 23]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.998375162548765e-05], 0, 3.117043972015381, 1535547588.8117917], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 64581, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [44, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.866922851002865e-05], 0, 4.251639366149902, 1535548336.165861], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 123861, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [12, 56]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.3642525614914618e-05], 0, 14.377593994140625, 1535549085.618711], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12506, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [40, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.506711887295765e-05], 0, 2.169248342514038, 1535549800.977766], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54981, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [38, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.5828701796599296e-05], 0, 1.7529754638671875, 1535550758.379618], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 135386, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [48, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.3415936738157693e-05], 0, 3.331063747406006, 1535551509.8037014], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54981, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [34, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.3128844173042964e-05], 0, 13.201400756835938, 1535552201.2710478], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 77269, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 1, 8, 4]], ["tile_x", "sp", [1, 1, 16, 1]], ["tile_rc", "sp", [1, 128]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[1.2395253346554289e-05], 0, 1.358180284500122, 1535553731.9634063], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 85466, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.104165770137252e-05], 0, 1.2656316757202148, 1535554370.9764986], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2686720, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011719296046511628], 0, 6.276942491531372, 1535555350.7657874], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 507382, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [62, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.409438910926794e-05], 0, 12.190387964248657, 1535556229.9904573], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4783222, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [24, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.688655548743102e-05], 0, 12.86388874053955, 1535557808.5567095], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 507382, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.016689842966923e-05], 0, 1.346564769744873, 1535558795.3216088], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 728214, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.165852137701277e-05], 0, 1.437685489654541, 1535560456.7344701], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1060336, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [36, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.687965669339914e-05], 0, 8.346641778945923, 1535561443.6733437], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 169462, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.555442124430956e-05], 0, 5.441713571548462, 1535562357.9449513], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3093622, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [20, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.046558225458468e-05], 0, 15.71262526512146, 1535563464.8701277], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 230902, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.234744090267983e-05], 0, 6.256928205490112, 1535564394.6937435], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 138742, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [46, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.1410453554889675e-05], 0, 6.182973146438599, 1535565178.7289805], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 538129, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [88, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.870800489312387e-05], 0, 2.2132797241210938, 1535565942.2794137], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1864854, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [24, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.229561833046472e-05], 0, 1.3550889492034912, 1535567724.982913], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1219702, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.477221223354958e-05], 0, 1.387446641921997, 1535568750.6347365], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 476656, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [76, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.4095696955929124e-05], 0, 4.452491521835327, 1535569599.26808], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1521136, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [48, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.237845636867933e-05], 0, 1.8098704814910889, 1535570428.9661424], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1213969, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [68, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.1499605432303545e-05], 0, 2.1479368209838867, 1535571183.5289185], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 108049, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.9696199288256228e-05], 0, 1.2669003009796143, 1535571947.9111545], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4168854, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [12, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.56696117003079e-05], 0, 10.758132457733154, 1535573553.205647], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1096854, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [16, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.320895284327323e-05], 0, 1.196434497833252, 1535575073.1209912], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1213936, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [52, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.5349264418212478e-05], 0, 12.102940082550049, 1535575957.618665], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1644016, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [48, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.3481232681782646e-05], 0, 4.663531303405762, 1535576847.956911], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 108016, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [44, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.2001159360797536e-05], 0, 2.2442550659179688, 1535577776.5404787], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1951216, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [16, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.9427484680025856e-05], 0, 2.6154847145080566, 1535578615.708037], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 169489, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [36, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.9033601698137118e-05], 0, 2.3652172088623047, 1535579507.9600186], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 20361, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 8, 2]], ["tile_x", "sp", [7, 1, 7, 1]], ["tile_rc", "sp", [8, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[2.5000261515844497e-05], 0, 12.293764114379883, 1535580383.487767], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 937489, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.758315576652602e-05], 0, 8.173214673995972, 1535581094.4785075], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3606637, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 2]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.120121898597627e-05], 0, 1.2265729904174805, 1535582613.3266957], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6024857, "c": null, "e": [["tile_f", "sp", [8, 4, 4, 1]], ["tile_y", "sp", [7, 2, 2, 1]], ["tile_x", "sp", [2, 1, 14, 1]], ["tile_rc", "sp", [48, 10]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.305960671378092e-05], 0, 1.3379249572753906, 1535583757.1784189], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9148823, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [32, 14]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.7931751037344395e-05], 0, 1.3435583114624023, 1535585470.9696934], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5692823, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [26, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.826996172402702e-05], 0, 1.3278939723968506, 1535587266.8153222], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7368199, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [2, 1, 14, 1]], ["tile_rc", "sp", [32, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.423708785377358e-05], 0, 12.33829402923584, 1535590002.951454], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5692823, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 2]], ["tile_rc", "sp", [22, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.094261456628478e-05], 0, 2.164489984512329, 1535590750.0318356], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9491839, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.790092597271349e-05], 0, 6.321907997131348, 1535591875.0671048], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15249775, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [24, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.548012210675484e-05], 0, 11.097039699554443, 1535592960.3593595], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2385743, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.167456558414171e-05], 0, 3.707690954208374, 1535594113.3787842], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12561814, "c": null, "e": [["tile_f", "sp", [4, 1, 4, 8]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 14]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.746549643705463e-05], 0, 6.304758310317993, 1535595088.5418801], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3912199, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [2, 1, 14, 1]], ["tile_rc", "sp", [16, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.471388193643998e-05], 0, 1.2050926685333252, 1535597295.1675448], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8339839, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [8, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.1152598169014084e-05], 0, 1.3271257877349854, 1535598536.5009234], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 209989, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 1, 8, 4]], ["tile_x", "sp", [7, 1, 14, 2]], ["tile_rc", "sp", [4, 32]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[2.9420414022718368e-05], 0, 1.34669828414917, 1535600086.3033242], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2380823, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [8, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.8549898443291326e-05], 0, 1.366112232208252, 1535601471.6817746], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9488423, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.121302610346464e-05], 0, 2.1284446716308594, 1535604383.0814273], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21776455, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [28, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.357535901778154e-05], 0, 2.2265069484710693, 1535605660.2191901], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 46344983, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [14, 1, 1, 4]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.734667162471396e-05], 0, 11.561939001083374, 1535606906.8334565], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13273108, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [7, 2, 1, 4]], ["tile_x", "sp", [7, 1, 8, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.074225447909754e-05], 0, 5.721418619155884, 1535608547.5546439], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2525783, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.997187543252595e-05], 0, 1.9399163722991943, 1535610416.8082263], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21716480, "c": null, "e": [["tile_f", "sp", [2, 2, 8, 4]], ["tile_y", "sp", [7, 4, 2, 1]], ["tile_x", "sp", [7, 1, 8, 1]], ["tile_rc", "sp", [16, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.383850146429376e-05], 0, 6.234032154083252, 1535611391.4232936], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 204395, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 1, 4, 8]], ["tile_x", "sp", [7, 1, 14, 2]], ["tile_rc", "sp", [8, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[7.82145242313705e-05], 0, 2.6251773834228516, 1535612238.94909], "v": 0.1}
+{"i": ["cuda -model=titanx", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 3739701, "c": null, "e": [["tile_f", "sp", [1, 8, 8, 1]], ["tile_y", "sp", [112, 1, 1, 1]], ["tile_x", "sp", [1, 7, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.801264570803718e-05], 0, 1.4477019309997559, 1535614680.9461718], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 184840, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.870914716312056e-05], 0, 1.8239610195159912, 1535340502.6141953], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 28435, "c": null, "e": [["tile_f", "sp", [64, 2, 16, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 1, 7]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.277302426160337e-05], 0, 1.2559425830841064, 1535341522.856728], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 276692, "c": null, "e": [["tile_f", "sp", [16, 4, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00014070911287313433], 0, 1.2780311107635498, 1535342217.8151824], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 267206, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 1, 16, 4]], ["tile_x", "sp", [1, 1, 8, 2]], ["tile_rc", "sp", [32, 16]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[5.8324577519379845e-05], 0, 1.5616421699523926, 1535343022.7710319], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54600, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.612890325552826e-05], 0, 1.328430414199829, 1535343755.3037415], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1120051, "c": null, "e": [["tile_f", "sp", [8, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.8724735294117643e-05], 0, 1.302473783493042, 1535344763.7252576], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1589056, "c": null, "e": [["tile_f", "sp", [16, 4, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.555625907000511e-05], 0, 1.39487886428833, 1535345590.4587004], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3917436, "c": null, "e": [["tile_f", "sp", [32, 2, 8, 2]], ["tile_y", "sp", [7, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.519168720173537e-05], 0, 5.207885265350342, 1535347591.9824076], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 35585, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [16, 2, 1, 8]], ["tile_x", "sp", [1, 1, 49, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[4.633917572575664e-05], 0, 2.31782603263855, 1535348964.346931], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 147911, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 2]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.6686891858624934e-05], 0, 1.2715818881988525, 1535349821.0645304], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2579787, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.9005322125603865e-05], 0, 4.739739656448364, 1535351179.3912873], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9659582, "c": null, "e": [["tile_f", "sp", [4, 4, 32, 1]], ["tile_y", "sp", [7, 1, 1, 4]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.9628529266067004e-05], 0, 1.2400987148284912, 1535352477.7351887], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1152842, "c": null, "e": [["tile_f", "sp", [4, 4, 32, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.6994346804511274e-05], 0, 1.7129621505737305, 1535353304.6144009], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 435771, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 4, 4, 2]], ["tile_x", "sp", [2, 2, 49, 1]], ["tile_rc", "sp", [8, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[3.5854004442366145e-05], 0, 1.4333581924438477, 1535354268.01083], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7756834, "c": null, "e": [["tile_f", "sp", [1, 2, 64, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.339707406829877e-05], 0, 1.3113553524017334, 1535355686.5242553], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16853780, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [56, 1, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.6182050362193858e-05], 0, 1.6612598896026611, 1535356896.8242538], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26764352, "c": null, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [14, 1, 1, 4]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.3482599499765514e-05], 0, 5.5789735317230225, 1535358026.6404517], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 43253325, "c": null, "e": [["tile_f", "sp", [1, 4, 8, 2]], ["tile_y", "sp", [112, 1, 1, 1]], ["tile_x", "sp", [1, 7, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.1339853208482515e-05], 0, 1.3942315578460693, 1535361338.131421], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 121867, "c": null, "e": [["tile_f", "sp", [64, 2, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.052611001410438e-05], 0, 1.422020435333252, 1535362447.7432306], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 25761, "c": null, "e": [["tile_f", "sp", [256, 1, 4, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.5548221728452815e-06], 0, 1.1385302543640137, 1535363136.5671299], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 153954, "c": null, "e": [["tile_f", "sp", [64, 2, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.496040041879944e-05], 0, 3.2046194076538086, 1535363578.4696045], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 19800, "c": null, "e": [["tile_f", "sp", [512, 1, 1, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.5496212216845505e-06], 0, 2.240983724594116, 1535364179.7086353], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1324913, "c": null, "e": [["tile_f", "sp", [32, 2, 4, 2]], ["tile_y", "sp", [2, 1, 7, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.495791946107784e-05], 0, 8.172152519226074, 1535364741.963095], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 254540, "c": null, "e": [["tile_f", "sp", [512, 1, 1, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.878607927128222e-06], 0, 1.8255438804626465, 1535365493.1087449], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 198461, "c": null, "e": [["tile_f", "sp", [32, 4, 4, 1]], ["tile_y", "sp", [2, 1, 7, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.3922045382165605e-05], 0, 3.5875909328460693, 1535366144.9359715], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 191400, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [1, 1, 14, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.2485037789792506e-06], 0, 1.1708734035491943, 1535366785.9483972], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1168397, "c": null, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.2439865390367933e-05], 0, 4.479329586029053, 1535368220.7407424], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1172325, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [1, 2, 7, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.0971720167125235e-06], 0, 1.5838329792022705, 1535368964.1501336], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7504397, "c": null, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [8, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.2662810974597377e-05], 0, 1.6510009765625, 1535370215.2665138], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 852720, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [1, 1, 14, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.131742867638056e-06], 0, 3.9804959297180176, 1535370929.9425554], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21004976, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.1988892341356673e-05], 0, 3.9282212257385254, 1535371747.987445], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3356280, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [7, 2, 1, 4]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.514915720717863e-06], 0, 1.3253803253173828, 1535372645.1413662], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3340948, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.1697967719298246e-05], 0, 4.008930683135986, 1535373250.6288095], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2346036, "c": null, "e": [["tile_f", "sp", [64, 1, 1, 1]], ["tile_y", "sp", [14, 2, 2, 1]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5159026368760064e-05], 0, 2.2290358543395996, 1535373978.812705], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7059804, "c": null, "e": [["tile_f", "sp", [1, 4, 16, 1]], ["tile_y", "sp", [56, 1, 1, 2]], ["tile_x", "sp", [2, 2, 28, 1]], ["tile_rc", "sp", [2, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.749397590782123e-05], 0, 5.936538457870483, 1535375090.9615788], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4741632, "c": null, "e": [["tile_f", "sp", [32, 1, 1, 1]], ["tile_y", "sp", [7, 1, 2, 8]], ["tile_x", "sp", [1, 1, 112, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.589733033477568e-06], 0, 1.2825348377227783, 1535375960.1960888], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 33218209, "c": null, "e": [["tile_f", "sp", [1, 4, 8, 1]], ["tile_y", "sp", [28, 4, 1, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.010032462686567e-06], 0, 1.3537733554840088, 1535377010.26167], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 75716, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 4, 16, 1]], ["tile_x", "sp", [1, 7, 7, 1]], ["tile_rc", "sp", [32, 16]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00011470169923664122], 0, 1.6527025699615479, 1535378530.7402716], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1169651, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 2, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0002516921139028475], 0, 2.42842698097229, 1535379736.2749858], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 861650, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0001474504529411765], 0, 1.4133880138397217, 1535380740.653147], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 332746, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [64, 4]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[0.00022330364635958397], 0, 8.144659042358398, 1535381829.4200964], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 448246, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0001520982654508612], 0, 5.780686140060425, 1535382802.5636008], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1813376, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 8, 2]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00027669260477941176], 0, 6.847986459732056, 1535385301.3655443], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1309376, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 8, 2]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00020200159623149397], 0, 4.39305853843689, 1535386443.8930204], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2124768, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 4, 8]], ["tile_x", "sp", [49, 4, 16, 1]], ["tile_rc", "sp", [16, 4]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00048387853225806453], 0, 7.0284013748168945, 1535388162.2723079], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 128325251, "c": null, "e": [["tile_f", "sp", [1, 2, 16, 2]], ["tile_y", "sp", [112, 2, 1, 1]], ["tile_x", "sp", [7, 1, 16, 2]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.3269261095100864e-05], 0, 1.526334524154663, 1535389977.55576], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 110920, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.3203476454537452e-05], 0, 1.3433687686920166, 1535350110.543228], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 737474, "c": null, "e": [["tile_f", "sp", [32, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.82280392670157e-05], 0, 2.3709828853607178, 1535350877.6429849], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1161671, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 2]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.388735292477555e-06], 0, 1.473917007446289, 1535353812.2598486], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5183051, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.732052392947103e-05], 0, 1.3970696926116943, 1535355518.4190004], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4876831, "c": null, "e": [["tile_f", "sp", [2, 2, 32, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.025047433035715e-06], 0, 1.389564037322998, 1535358581.229104], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 25804827, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [32, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.629003140096618e-05], 0, 1.433337926864624, 1535360709.7012227], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2338581, "c": null, "e": [["tile_f", "sp", [1, 8, 8, 1]], ["tile_y", "sp", [56, 1, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.529723254546355e-06], 0, 2.2757744789123535, 1535362192.8535852], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 28125, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 2, 2, 8]], ["tile_x", "sp", [1, 2, 98, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[3.7446556775642626e-05], 0, 3.851289749145508, 1535364277.7394183], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8797896, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [64, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.1822753712346206e-05], 0, 1.297123670578003, 1535391282.1496253], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 868691, "c": null, "e": [["tile_f", "sp", [14, 1, 16, 2]], ["tile_y", "sp", [2, 2, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.868886900129702e-05], 0, 6.360664129257202, 1535392201.6822503], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8156810, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 2]], ["tile_y", "sp", [2, 1, 2, 2]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.840593352694418e-05], 0, 1.2624528408050537, 1535392929.3049552], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5573259, "c": null, "e": [["tile_f", "sp", [40, 2, 4, 1]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [64, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.779546543485186e-05], 0, 1.4392306804656982, 1535394058.0221245], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8394696, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [40, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.04307966765739e-05], 0, 5.328145265579224, 1535394950.332171], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 368726, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [12, 1, 32, 1]], ["tile_x", "sp", [1, 1, 2, 2]], ["tile_rc", "sp", [28, 16]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[8.091512297734628e-05], 0, 3.577399730682373, 1535396588.173184], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13407875, "c": null, "e": [["tile_f", "sp", [14, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.643936534740545e-05], 0, 1.5609734058380127, 1535398040.0147324], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 10648394, "c": null, "e": [["tile_f", "sp", [12, 2, 16, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [24, 16]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.7830928399183824e-05], 0, 1.3832390308380127, 1535398954.3528938], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 26392394, "c": null, "e": [["tile_f", "sp", [12, 2, 16, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [12, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.5432101456639564e-05], 0, 1.7413804531097412, 1535399941.1696503], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1683420, "c": null, "e": [["tile_f", "sp", [12, 4, 8, 1]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [40, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.873922139303483e-05], 0, 1.2541792392730713, 1535400723.5157862], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3731075, "c": null, "e": [["tile_f", "sp", [10, 1, 16, 2]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.614155295545155e-05], 0, 1.513995885848999, 1535402527.1831696], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2614152, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [8, 1, 1, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.91711075310559e-05], 0, 1.7116973400115967, 1535404126.9687028], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6364244, "c": null, "e": [["tile_f", "sp", [10, 1, 32, 1]], ["tile_y", "sp", [4, 2, 1, 1]], ["tile_x", "sp", [2, 1, 4, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.243342838654012e-05], 0, 1.4600000381469727, 1535405934.3806696], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 626349, "c": null, "e": [["tile_f", "sp", [8, 4, 6, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.678651497695853e-05], 0, 3.809021234512329, 1535407164.0710874], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 776877, "c": null, "e": [["tile_f", "sp", [8, 4, 6, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.552685176738883e-05], 0, 4.9299421310424805, 1535408070.3316078], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 540340, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.962713978985839e-05], 0, 2.6122801303863525, 1535408801.494736], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 234784, "c": null, "e": [["tile_f", "sp", [8, 4, 5, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.338450232361639e-05], 0, 6.433482646942139, 1535409694.7188215], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 733989, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 3]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.02349209171736e-05], 0, 1.5989787578582764, 1535410517.0261497], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 19838, "c": null, "e": [["tile_f", "sp", [8, 1, 5, 4]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.227214694894146e-05], 0, 6.451417684555054, 1535411577.5658758], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 220455, "c": null, "e": [["tile_f", "sp", [4, 5, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.308726506364922e-05], 0, 5.126100778579712, 1535412695.7309365], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 147966, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.653925856228829e-05], 0, 1.454458475112915, 1535413420.592295], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 22102, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.2646732013385384e-05], 0, 1.247044324874878, 1535414154.7640848], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 233982, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.647298758465011e-05], 0, 2.7552649974823, 1535415102.1237185], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 175734, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.234621726941408e-05], 0, 4.006468772888184, 1535415876.2920923], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 45174, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [96, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.568510146989094e-05], 0, 2.430988311767578, 1535416680.345598], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 35137, "c": null, "e": [["tile_f", "sp", [16, 1, 3, 4]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [64, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.250999441808541e-05], 0, 1.5181522369384766, 1535418207.809458], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 664874, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [12, 8]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.5635400142483974e-05], 0, 2.2501699924468994, 1535419527.4759367], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 979233, "c": null, "e": [["tile_f", "sp", [16, 8, 3, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [144, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002292356246200608], 0, 4.310108661651611, 1535420473.161267], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2741780, "c": null, "e": [["tile_f", "sp", [4, 4, 3, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.7069100853548967e-05], 0, 1.666454553604126, 1535421234.2659154], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 223628, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [5, 1, 7, 1]], ["tile_x", "sp", [5, 1, 7, 1]], ["tile_rc", "sp", [12, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.0876184275521405e-05], 0, 1.3121867179870605, 1535422386.6796708], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 161306, "c": null, "e": [["tile_f", "sp", [3, 4, 4, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.6682681009561928e-05], 0, 1.4955830574035645, 1535423237.755808], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 75304, "c": null, "e": [["tile_f", "sp", [4, 2, 4, 2]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.9178238174379953e-05], 0, 2.4978082180023193, 1535423921.9796646], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 120301, "c": null, "e": [["tile_f", "sp", [2, 4, 4, 1]], ["tile_y", "sp", [7, 1, 5, 1]], ["tile_x", "sp", [5, 1, 7, 1]], ["tile_rc", "sp", [8, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.0797558232075604e-05], 0, 1.6644856929779053, 1535425238.4115841], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4409403, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [3, 1, 4, 8]], ["tile_x", "sp", [1, 4, 81, 1]], ["tile_rc", "sp", [4, 24]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[3.319664817679558e-05], 0, 2.4141483306884766, 1535427557.7824433], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2449305, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [3, 4, 4, 2]], ["tile_x", "sp", [1, 4, 81, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[2.5309661492238357e-05], 0, 1.4485929012298584, 1535429272.3281975], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 4140319, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 2]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [16, 3]], ["tile_ry", "sp", [1, 5]], ["tile_rx", "sp", [1, 5]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.883521648815319e-05], 0, 9.28942060470581, 1535430556.1709695], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 663060, "c": null, "e": [["tile_f", "sp", [4, 4, 3, 1]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.1977059175372836e-05], 0, 2.836064100265503, 1535431600.069139], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1083788, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 1]], ["tile_y", "sp", [5, 1, 7, 1]], ["tile_x", "sp", [5, 1, 7, 1]], ["tile_rc", "sp", [8, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.4761395499656088e-05], 0, 1.3095550537109375, 1535432754.9066033], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 922099, "c": null, "e": [["tile_f", "sp", [6, 4, 4, 2]], ["tile_y", "sp", [71, 1, 1, 1]], ["tile_x", "sp", [1, 1, 71, 1]], ["tile_rc", "sp", [40, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00037999983291139237], 0, 1.333446979522705, 1535433665.455902], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 39282, "c": null, "e": [["tile_f", "sp", [5, 2, 2, 4]], ["tile_y", "sp", [73, 1, 1, 1]], ["tile_x", "sp", [1, 1, 73, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.2667290093627303e-05], 0, 1.5402162075042725, 1535434456.8025699], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 18733533, "c": null, "e": [["tile_f", "sp", [1, 8, 8, 1]], ["tile_y", "sp", [49, 1, 1, 3]], ["tile_x", "sp", [3, 1, 49, 1]], ["tile_rc", "sp", [16, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00013859728162511541], 0, 4.481893539428711, 1535435277.8664289], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6050372, "c": null, "e": [["tile_f", "sp", [1, 1, 32, 1]], ["tile_y", "sp", [21, 7, 1, 1]], ["tile_x", "sp", [7, 1, 7, 3]], ["tile_rc", "sp", [16, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.415943226079594e-05], 0, 1.405092716217041, 1535437071.6393325], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27342, "c": null, "e": [["tile_f", "sp", [1, 8, 4, 1]], ["tile_y", "sp", [149, 1, 1, 1]], ["tile_x", "sp", [1, 1, 149, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.0394228291887484e-05], 0, 2.09767746925354, 1535437921.065245], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 108510, "c": null, "e": [["tile_f", "sp", [4, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [8, 124]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.010334864080232e-05], 0, 1.5134379863739014, 1535510578.9815462], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 237150, "c": null, "e": [["tile_f", "sp", [4, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [48, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.905045659898477e-05], 0, 1.168386697769165, 1535511125.4239872], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 100826, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.9401293562564634e-05], 0, 1.3234310150146484, 1535511541.6431155], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 168015, "c": null, "e": [["tile_f", "sp", [32, 1, 4, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.4511918993879114e-05], 0, 2.1898694038391113, 1535512300.6017644], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 212186, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [18, 48]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.7799368040748638e-05], 0, 1.2502412796020508, 1535512769.6244454], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 118106, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.7712727648030198e-05], 0, 1.373544454574585, 1535513219.419708], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 152670, "c": null, "e": [["tile_f", "sp", [4, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [40, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.6151012154458428e-05], 0, 1.2929022312164307, 1535513641.1468432], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 152666, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.6367799781778505e-05], 0, 1.3143672943115234, 1535514062.8160617], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8666, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [46, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.5746391098981842e-05], 0, 1.349114179611206, 1535514467.9158273], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 118106, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [44, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5157039765845782e-05], 0, 1.1946027278900146, 1535514920.2627847], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 116175, "c": null, "e": [["tile_f", "sp", [32, 1, 4, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [24, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.1306232711966325e-05], 0, 1.6486470699310303, 1535515516.133535], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 137310, "c": null, "e": [["tile_f", "sp", [4, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.3417131932021466e-05], 0, 1.2014439105987549, 1535515972.7182794], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31706, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [38, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.3372493136031379e-05], 0, 1.2704458236694336, 1535516428.3896213], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 187226, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [12, 48]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.2496573377156944e-05], 0, 1.2879750728607178, 1535517005.3094254], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 112346, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [2, 272]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.2027159786420146e-05], 0, 1.5824146270751953, 1535517491.8704495], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 75730, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 16, 1]], ["tile_x", "sp", [2, 1, 4, 2]], ["tile_rc", "sp", [2, 64]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[8.30705703527327e-06], 0, 6.790492057800293, 1535518080.136758], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8666, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.1499615007281368e-05], 0, 1.2344355583190918, 1535518522.5797985], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 254349, "c": null, "e": [["tile_f", "sp", [16, 4, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.065668067906225e-05], 0, 1.1561696529388428, 1535519470.7963555], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 507382, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [62, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.5209433299731184e-05], 0, 1.4269888401031494, 1535520041.0994248], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2903547, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [60, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.521613414019163e-05], 0, 1.337665319442749, 1535520693.654924], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1244662, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.378593229744728e-05], 0, 1.4942715167999268, 1535521353.5999422], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1701616, "c": null, "e": [["tile_f", "sp", [16, 2, 4, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [2, 1, 7, 1]], ["tile_rc", "sp", [32, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.22431098265896e-05], 0, 5.287805080413818, 1535522295.2589383], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2473462, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [54, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.2237321375630003e-05], 0, 2.585005283355713, 1535522937.8135786], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 169462, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.1596151432272925e-05], 0, 4.9706690311431885, 1535523507.5953581], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1859062, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [50, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.053800984548065e-05], 0, 1.2881567478179932, 1535524279.4223301], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 230907, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.9923293531677513e-05], 0, 3.857074737548828, 1535524878.4241247], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 876022, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [46, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.9062312920848685e-05], 0, 7.322245359420776, 1535525558.424067], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1459707, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [44, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.8408509690873404e-05], 0, 3.769604444503784, 1535526161.9636657], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2504187, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [42, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.768176361065284e-05], 0, 3.958688497543335, 1535526768.0907528], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 200187, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [40, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.7005916666666666e-05], 0, 1.2610886096954346, 1535527376.862039], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1244662, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [38, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.613369627876963e-05], 0, 1.232668161392212, 1535528073.8818192], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2196982, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [36, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5461951984371785e-05], 0, 6.229677200317383, 1535528790.1099694], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 138747, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [34, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.4684608750978857e-05], 0, 4.6408140659332275, 1535529488.5018454], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1060342, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.3865336010338778e-05], 0, 1.2443079948425293, 1535530103.911455], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 323062, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [30, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.3085105267748126e-05], 0, 2.159207344055176, 1535530789.2544587], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2442769, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [8, 56]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.1815016365066875e-05], 0, 2.2030670642852783, 1535531790.218784], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1275382, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [26, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.1618833707517225e-05], 0, 1.2819890975952148, 1535532505.0172954], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 230902, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [24, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.0968731876644256e-05], 0, 8.432388544082642, 1535533185.603903], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 169467, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [22, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.0255205892405495e-05], 0, 2.2535476684570312, 1535533836.1736326], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1490427, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [20, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.516839233599796e-06], 0, 5.255681037902832, 1535534486.7834814], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1920507, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [18, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.70503696181966e-06], 0, 1.3429267406463623, 1535535109.3423762], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 26631, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 2, 4, 2]], ["tile_x", "sp", [1, 1, 49, 1]], ["tile_rc", "sp", [1, 128]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[1.2433500124264767e-05], 0, 2.377492904663086, 1535535687.8450365], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 415222, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.865961681411298e-06], 0, 6.301952362060547, 1535536188.88553], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9030482, "c": null, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [2, 7, 2, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.274914188034188e-05], 0, 1.4122755527496338, 1535537441.3448718], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2193775, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [24, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.385060962482191e-05], 0, 1.294996738433838, 1535538590.7639327], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1425743, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.4060694725028058e-05], 0, 1.290968894958496, 1535539867.135623], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7953748, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [26, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.562269837745517e-05], 0, 4.757439851760864, 1535540693.427137], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 11217748, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [8, 48]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.2392697674418605e-05], 0, 1.334132194519043, 1535541874.2043667], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3345628, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [22, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.204317429577465e-05], 0, 4.975332975387573, 1535543293.7426958], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9491839, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.6622234280030944e-05], 0, 2.4568378925323486, 1535544351.0567925], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2387839, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [8, 36]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.6870072167266188e-05], 0, 1.5000736713409424, 1535545510.5365977], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5841778, "c": null, "e": [["tile_f", "sp", [4, 1, 16, 2]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5711118778753657e-05], 0, 1.3064296245574951, 1535546967.341969], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3345623, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 14]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.3947885016136469e-05], 0, 3.1074414253234863, 1535548002.22679], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2001748, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [4, 48]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.2573500542978866e-05], 0, 1.5108652114868164, 1535549122.14382], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8337775, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [8, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.355404564444722e-06], 0, 1.492844581604004, 1535550192.1845133], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 209989, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 1, 8, 4]], ["tile_x", "sp", [7, 1, 14, 2]], ["tile_rc", "sp", [4, 32]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[1.5450798312493645e-05], 0, 1.48533034324646, 1535551019.5225036], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5267818, "c": null, "e": [["tile_f", "sp", [4, 1, 16, 2]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.9109391835766e-06], 0, 3.3182780742645264, 1535552192.8403792], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2566943, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [14, 2, 1, 2]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.759123719542516e-05], 0, 3.0082433223724365, 1535553514.2241092], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21004996, "c": null, "e": [["tile_f", "sp", [4, 2, 4, 4]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [32, 7]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.3200856352368304e-05], 0, 1.5159938335418701, 1535554700.846309], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 37133000, "c": null, "e": [["tile_f", "sp", [2, 2, 8, 4]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [16, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.982201131176821e-05], 0, 1.5415749549865723, 1535555668.8435583], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5644948, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [8, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.5671416794937575e-05], 0, 1.374856948852539, 1535558237.050249], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 32528480, "c": null, "e": [["tile_f", "sp", [2, 2, 8, 4]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [8, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.5659986445626108e-05], 0, 4.918555021286011, 1535560222.7970054], "v": 0.1}
+{"i": ["cuda -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 203213, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 1, 8, 4]], ["tile_x", "sp", [7, 1, 28, 1]], ["tile_rc", "sp", [8, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[4.457051544860368e-05], 0, 2.7779624462127686, 1535560904.415111], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 58336, "e": [["tile_f", "sp", [8, 4, 16, 1]], ["tile_y", "sp", [1, 7, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "c": null}], "r": [[0.0005035017], 0, 12.29834508895874, 1554955488.1772192], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 254471, "e": [["tile_f", "sp", [64, 4, 8, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0005412969999999999], 0, 1.7040576934814453, 1554977839.9065943], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 142144, "e": [["tile_f", "sp", [16, 1, 32, 4]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 7, 1, 1]], ["tile_rc", "sp", [256, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "c": null}], "r": [[0.0013291223], 0, 3.519747257232666, 1554970439.7118416], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 9586, "e": [["tile_f", "sp", [8, 1, 16, 4]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [256, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "c": null}], "r": [[0.00040170630000000004], 0, 10.144535064697266, 1554962211.3523424], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 2488262, "e": [["tile_f", "sp", [8, 4, 4, 2]], ["tile_y", "sp", [1, 1, 14, 1]], ["tile_x", "sp", [1, 1, 2, 7]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.00034071700000000004], 0, 9.662159442901611, 1554948918.5204923], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 3552430, "e": [["tile_f", "sp", [32, 8, 4, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0003303043], 0, 9.292404174804688, 1554959345.3456247], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 1647687, "e": [["tile_f", "sp", [8, 8, 16, 1]], ["tile_y", "sp", [7, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [128, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "c": null}], "r": [[0.0013268225999999998], 0, 5.178571462631226, 1554965249.3803976], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 2168137, "e": [["tile_f", "sp", [2, 4, 32, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 7, 2, 1]], ["tile_rc", "sp", [256, 2]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.000460863], 0, 7.6866724491119385, 1554966621.4672701], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 8146507, "e": [["tile_f", "sp", [4, 8, 1, 4]], ["tile_y", "sp", [7, 2, 2, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [128, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0004175422], 0, 1.8580596446990967, 1554945646.8539002], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 14933965, "e": [["tile_f", "sp", [8, 4, 2, 8]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.000431368], 0, 6.44800329208374, 1554980848.7567935], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 6837970, "e": [["tile_f", "sp", [8, 2, 4, 8]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [128, 2]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "c": null}], "r": [[0.0011816415], 0, 9.641122579574585, 1554953178.4609714], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "i": 519574, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 1, 4, 8]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0004963299], 0, 12.552602767944336, 1554943327.7279453], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 9077604, "e": [["tile_f", "sp", [1, 1, 32, 4]], ["tile_y", "sp", [14, 1, 2, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.000269107], 0, 8.641182899475098, 1554969244.0934107], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 25451286, "e": [["tile_f", "sp", [1, 8, 4, 2]], ["tile_y", "sp", [14, 2, 1, 2]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.00044651580000000003], 0, 8.235926151275635, 1554954272.9459844], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 32067807, "e": [["tile_f", "sp", [4, 16, 2, 2]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 1, 56, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.00044522379999999996], 0, 8.113616228103638, 1554957196.4803026], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 20617002, "e": [["tile_f", "sp", [1, 8, 4, 2]], ["tile_y", "sp", [56, 1, 1, 1]], ["tile_x", "sp", [1, 2, 28, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0001289671], 0, 4.256742000579834, 1554979811.828667], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 62969413, "e": [["tile_f", "sp", [1, 1, 8, 8]], ["tile_y", "sp", [28, 1, 1, 4]], ["tile_x", "sp", [4, 2, 14, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0008444969], 0, 8.917776823043823, 1554974316.7219121], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 71826, "c": null, "e": [["tile_f", "sp", [16, 4, 16, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0005240697591623037], 0, 1.9879345893859863, 1538684133.1681948], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 21258, "c": null, "e": [["tile_f", "sp", [32, 2, 8, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.178158100381194e-05], 0, 1.0447304248809814, 1538685070.5457134], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 113016, "c": null, "e": [["tile_f", "sp", [16, 2, 32, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0003075978134556575], 0, 1.165560245513916, 1538686093.9553363], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 16299, "c": null, "e": [["tile_f", "sp", [128, 1, 4, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.744394313505425e-05], 0, 1.0555353164672852, 1538686880.8105361], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 761442, "c": null, "e": [["tile_f", "sp", [16, 8, 4, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002860706657142857], 0, 1.1914925575256348, 1538688643.803028], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 312850, "c": null, "e": [["tile_f", "sp", [256, 1, 2, 1]], ["tile_y", "sp", [1, 1, 2, 7]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.69354909512761e-05], 0, 1.1330385208129883, 1538690229.7487793], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 705122, "c": null, "e": [["tile_f", "sp", [16, 8, 4, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00015319955436447167], 0, 1.1951820850372314, 1538691299.4314046], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 234144, "c": null, "e": [["tile_f", "sp", [128, 1, 2, 1]], ["tile_y", "sp", [1, 1, 7, 2]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.102415431348724e-05], 0, 1.0164296627044678, 1538692213.639821], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3280722, "c": null, "e": [["tile_f", "sp", [4, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.000316804914556962], 0, 1.1764893531799316, 1538692968.977005], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 1173975, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [1, 1, 4, 7]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.466998284862043e-05], 0, 4.699853897094727, 1538693939.3165033], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5128722, "c": null, "e": [["tile_f", "sp", [4, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0001820941309090909], 0, 4.953519344329834, 1538694726.737019], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 1043160, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [2, 2, 7, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010517174790794979], 0, 1.7660229206085205, 1538695922.5100384], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 3358080, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [2, 1, 4, 7]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001251102944099379], 0, 1.777782917022705, 1538698543.7329175], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2886240, "c": null, "e": [["tile_f", "sp", [64, 1, 1, 1]], ["tile_y", "sp", [7, 1, 4, 2]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001965908774319066], 0, 1.0119473934173584, 1538701164.2750542], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 56412935, "c": null, "e": [["tile_f", "sp", [1, 2, 16, 2]], ["tile_y", "sp", [28, 4, 1, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [2, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00018805118421052631], 0, 2.126875638961792, 1538702679.443884], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 4725392, "c": null, "e": [["tile_f", "sp", [32, 1, 1, 1]], ["tile_y", "sp", [4, 1, 4, 7]], ["tile_x", "sp", [2, 1, 56, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00012950566279069768], 0, 1.7964670658111572, 1538703580.0487945], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 16795901, "c": null, "e": [["tile_f", "sp", [1, 1, 8, 4]], ["tile_y", "sp", [28, 2, 1, 2]], ["tile_x", "sp", [2, 1, 56, 1]], ["tile_rc", "sp", [1, 3]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00015486618730650156], 0, 19.915329933166504, 1538704740.3396301], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 78023, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [62, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00012308115431348724], 0, 1.0024254322052002, 1538788111.7026997], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22343, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [48, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.885210177865612e-05], 0, 1.075761079788208, 1538789049.4131649], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8903, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00011706999299065421], 0, 5.105271100997925, 1538789964.8743596], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12743, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [56, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00011415141807909604], 0, 5.740638494491577, 1538790799.898342], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20455, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [36, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.283645821854913e-05], 0, 1.0991449356079102, 1538792263.1661618], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 91463, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010605849119170984], 0, 1.1531589031219482, 1538793276.2868276], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14719, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [40, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.286428052805281e-05], 0, 10.84407353401184, 1538794516.506616], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 16610, "c": null, "e": [["tile_f", "sp", [8, 2, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 24]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.069063076923077e-05], 0, 1.1301770210266113, 1538796067.3274465], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 124107, "c": null, "e": [["tile_f", "sp", [8, 2, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [46, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.65327084942085e-05], 0, 3.662139415740967, 1538796965.1628647], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 122339, "c": null, "e": [["tile_f", "sp", [2, 2, 16, 2]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [22, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.34424737718568e-05], 0, 1.3078110218048096, 1538798665.5558236], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 24290, "c": null, "e": [["tile_f", "sp", [8, 2, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [24, 28]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.355548569332356e-05], 0, 15.131038427352905, 1538800102.8032422], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 168258, "c": null, "e": [["tile_f", "sp", [4, 8, 4, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.260219464544139e-05], 0, 11.109286546707153, 1538801733.528239], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 31827, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [1, 7, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [38, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.92368214532872e-05], 0, 5.3235862255096436, 1538802662.3141403], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 57028, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [36, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.943618764845606e-05], 0, 2.8046820163726807, 1538803707.6444323], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 32068, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [34, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.5462947339246124e-05], 0, 3.488588809967041, 1538804940.014018], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 85025, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 4, 8, 1]], ["tile_x", "sp", [1, 2, 8, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[3.387354208754209e-05], 0, 2.096045732498169, 1538806335.7966588], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 47428, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [1, 1, 1, 7]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.335046297282898e-05], 0, 7.114248037338257, 1538807356.784124], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 2666034, "e": [["tile_f", "sp", [16, 4, 4, 2]], ["tile_y", "sp", [2, 1, 7, 1]], ["tile_x", "sp", [1, 7, 2, 1]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0006444008999999999], 0, 4.5529913902282715, 1554950358.2710702], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1613479, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [62, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00018629811131725419], 0, 1.412217378616333, 1538811130.1064463], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4624039, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [60, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00018082491335740073], 0, 1.389707326889038, 1538812582.4735014], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1613479, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [58, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00017594096335078532], 0, 32.868462324142456, 1538814037.147122], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2657959, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [56, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00016909465551839464], 0, 2.948920965194702, 1538815326.2985299], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1643783, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [108, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00017283390705679863], 0, 1.053091049194336, 1538816925.362273], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 599308, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [52, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.000165996498349835], 0, 1.1660890579223633, 1538818077.9467108], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2965159, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [50, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00015221640881458965], 0, 1.501213550567627, 1538820032.9254172], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 783628, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00015289620182094084], 0, 1.3272161483764648, 1538821218.725461], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1982119, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [46, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001415780463483146], 0, 1.4632749557495117, 1538822696.5618985], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1889959, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [44, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001352428239247312], 0, 1.414778232574463, 1538824048.3295507], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3241639, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [42, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00012970178120978121], 0, 1.4736008644104004, 1538825535.8499055], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2657959, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [40, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001242407141089109], 0, 3.6424829959869385, 1538827139.7220483], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 846263, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [2, 1, 1, 7]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [76, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.000124436], 0, 10.96009349822998, 1538828279.2371376], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3487399, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [36, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011318031354983202], 0, 13.00906491279602, 1538829970.9019978], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1613479, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [34, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010710978525641025], 0, 2.2806811332702637, 1538832060.504336], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1367695, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010063262914572865], 0, 10.810802698135376, 1538833448.3748267], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1060108, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [30, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00010090847177419355], 0, 1.173581838607788, 1538834655.558986], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2350759, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [28, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.046342356115108e-05], 0, 1.443748950958252, 1538836551.1019008], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2012839, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [26, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.418914297729183e-05], 0, 1.418027400970459, 1538838687.6621108], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2197159, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [24, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.74511379044684e-05], 0, 1.4423730373382568, 1538840059.6083984], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 537868, "c": null, "e": [["tile_f", "sp", [2, 4, 16, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [22, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.273062309368192e-05], 0, 14.969782829284668, 1538841445.5116246], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1859183, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [40, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[6.237981982543641e-05], 0, 16.40942907333374, 1538842541.3345072], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2934411, "c": null, "e": [["tile_f", "sp", [4, 4, 4, 2]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [36, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.029475389251632e-05], 0, 1.5659475326538086, 1538843875.1409004], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 21032, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 4, 4, 2]], ["tile_x", "sp", [1, 1, 49, 1]], ["tile_rc", "sp", [4, 32]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00010605685370950888], 0, 2.8770177364349365, 1538845070.5571578], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 661098, "c": null, "e": [["tile_f", "sp", [4, 8, 4, 1]], ["tile_y", "sp", [1, 7, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.281321889696452e-05], 0, 1.3150086402893066, 1538846421.29054], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 11142848, "e": [["tile_f", "sp", [4, 4, 8, 2]], ["tile_y", "sp", [4, 1, 7, 1]], ["tile_x", "sp", [2, 7, 2, 1]], ["tile_rc", "sp", [128, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0009459629], 0, 10.448927640914917, 1554972391.3655684], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10422097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [60, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002770921961325967], 0, 1.2362148761749268, 1538850912.3550613], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6198097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [56, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00025754312082262214], 0, 1.3426780700683594, 1538852293.8814673], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2934097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [52, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002442188317073171], 0, 2.747875928878784, 1538854025.399341], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4086097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [48, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00022825630227272728], 0, 6.214647531509399, 1538855270.4081788], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5238097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [44, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00021150272151898732], 0, 2.7867863178253174, 1538856985.0964792], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6225984, "c": null, "e": [["tile_f", "sp", [2, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [40, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00019690313976377952], 0, 1.1801960468292236, 1538858663.3631353], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7926097, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 8]], ["tile_y", "sp", [7, 1, 4, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [36, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00018197597272727272], 0, 1.2835686206817627, 1538860130.2034986], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4113984, "c": null, "e": [["tile_f", "sp", [2, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0001616460323101777], 0, 1.2255666255950928, 1538861855.9401083], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3157223, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [4, 1, 1, 7]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [28, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00014706854852941176], 0, 1.8666982650756836, 1538863752.199898], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6421223, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [4, 1, 1, 7]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.000128685567394095], 0, 22.308583974838257, 1538865795.462426], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3153984, "c": null, "e": [["tile_f", "sp", [2, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00011284384026996625], 0, 17.547949075698853, 1538867716.5715816], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 152856, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [1, 4, 4, 2]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0001753623992994746], 0, 2.021149158477783, 1538869263.9188836], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3729984, "c": null, "e": [["tile_f", "sp", [2, 8, 8, 1]], ["tile_y", "sp", [4, 7, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.381299249530957e-05], 0, 1.9887633323669434, 1538871065.2399995], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 36378013, "e": [["tile_f", "sp", [2, 32, 2, 1]], ["tile_y", "sp", [28, 1, 1, 2]], ["tile_x", "sp", [1, 1, 56, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0008156606], 0, 6.791729211807251, 1554976592.3804111], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12509816, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [28, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0004905505902439024], 0, 1.1805377006530762, 1538874648.0994287], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25565816, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00042907040170940173], 0, 4.26771354675293, 1538876153.6036224], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21725816, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0003679489267399267], 0, 1.2252421379089355, 1538878120.6407452], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8669816, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00030473472948328267], 0, 3.462869644165039, 1538879663.278558], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22493816, "c": null, "e": [["tile_f", "sp", [2, 4, 8, 2]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [12, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002451529803921569], 0, 4.241109609603882, 1538881371.0287638], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 108068, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 4, 4]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[0.0004426895309734513], 0, 2.5171635150909424, 1538882574.138559], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 18653840, "c": null, "e": [["tile_f", "sp", [2, 2, 8, 4]], ["tile_y", "sp", [28, 1, 2, 1]], ["tile_x", "sp", [1, 7, 8, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001896496231060606], 0, 2.462634563446045, 1538884874.1319482], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "i": 410647, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 2, 16, 4]], ["tile_x", "sp", [1, 4, 4, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0006187037], 0, 8.767261505126953, 1554945239.17839], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 107671, "c": null, "e": [["tile_f", "sp", [4, 2, 32, 2]], ["tile_y", "sp", [1, 7, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]]}], "r": [[0.00011170939999999998], 0, 2.5888760089874268, 1555022071.900302], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 608610, "c": null, "e": [["tile_f", "sp", [8, 1, 32, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 1, 7]], ["tile_rc", "sp", [128, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]]}], "r": [[0.0005576546], 0, 2.253612995147705, 1555024034.1775575], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "i": 79987, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 16, 8]], ["tile_x", "sp", [1, 7, 7, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0004759948], 0, 8.792563199996948, 1554940745.0696976], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 1457647, "c": null, "e": [["tile_f", "sp", [2, 4, 32, 1]], ["tile_y", "sp", [2, 7, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]]}], "r": [[9.957760000000001e-05], 0, 1.0042572021484375, 1555021138.3377924], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 7778627, "c": null, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [1, 2, 7, 1]], ["tile_x", "sp", [2, 7, 1, 1]], ["tile_rc", "sp", [128, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]]}], "r": [[0.0004268699], 0, 8.907147645950317, 1555019702.5898302], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 6028824, "c": null, "e": [["tile_f", "sp", [2, 8, 8, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [8, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]]}], "r": [[0.0001096032], 0, 11.373619556427002, 1555025987.4431183], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 30965602, "c": null, "e": [["tile_f", "sp", [2, 1, 16, 4]], ["tile_y", "sp", [14, 1, 2, 1]], ["tile_x", "sp", [1, 7, 4, 1]], ["tile_rc", "sp", [64, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]]}], "r": [[0.0004450338], 0, 4.735571622848511, 1555023493.8178582], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "i": 321677, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 4, 2]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0005076439], 0, 14.493378162384033, 1554942130.618302], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 93434, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [8, 1, 8, 8]], ["tile_x", "sp", [1, 7, 7, 1]], ["tile_rc", "sp", [128, 4]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0016002932380952382], 0, 1.2927846908569336, 1538706207.6221724], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1171807, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 4, 16, 2]], ["tile_x", "sp", [2, 7, 14, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.005323976947368421], 0, 3.491469383239746, 1538707562.3229945], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1062454, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [16, 4, 4, 2]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.002699300078947368], 0, 1.712977647781372, 1538710204.4787867], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 796847, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [8, 4, 4, 2]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.003582328821428571], 0, 30.718233823776245, 1538711290.5169294], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 301901, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [8, 2, 2, 8]], ["tile_x", "sp", [1, 7, 28, 1]], ["tile_rc", "sp", [32, 4]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "winograd"}], "r": [[0.002160468680851064], 0, 2.2771291732788086, 1538713514.750547], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1813400, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 2, 8, 4]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [16, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0040044463200000005], 0, 12.513367891311646, 1538715498.4120855], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1309376, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 4, 8, 2]], ["tile_x", "sp", [7, 7, 16, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.002707434108108108], 0, 13.540417194366455, 1538717126.7233148], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2195328, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 4, 8]], ["tile_x", "sp", [49, 4, 16, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.005546605210526316], 0, 2.467291831970215, 1538719045.851881], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 161059560, "c": null, "e": [["tile_f", "sp", [1, 4, 4, 4]], ["tile_y", "sp", [28, 1, 1, 8]], ["tile_x", "sp", [14, 1, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0005955082426035502], 0, 1.6560072898864746, 1538721566.8455422], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3813136, "c": null, "e": [["tile_f", "sp", [6, 1, 8, 4]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [2, 1, 4, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00019577969667318982], 0, 16.34182333946228, 1538724300.7337055], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7045458, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 7]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 2, 4, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00032905157565789474], 0, 1.4203288555145264, 1538725181.1495411], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7828109, "c": null, "e": [["tile_f", "sp", [12, 1, 8, 4]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0003022847069486405], 0, 18.06054711341858, 1538726233.047221], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8705133, "c": null, "e": [["tile_f", "sp", [10, 4, 8, 1]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00026277220997375327], 0, 1.9644010066986084, 1538727120.9285002], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8301678, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [64, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0001213984012121212], 0, 1.8440706729888916, 1538729278.8008165], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 215626, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 192, 1]], ["tile_x", "sp", [1, 1, 1, 4]], ["tile_rc", "sp", [112, 4]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0006950035655172414], 0, 31.39168119430542, 1538730783.9491055], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5710194, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 7]], ["tile_y", "sp", [1, 2, 2, 2]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [80, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00021037706289308176], 0, 1.2826430797576904, 1538732840.2017002], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 31961559, "c": null, "e": [["tile_f", "sp", [6, 1, 16, 4]], ["tile_y", "sp", [1, 4, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 12]], ["tile_ry", "sp", [3, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00018740938764044942], 0, 1.6747934818267822, 1538734623.9567337], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 16468119, "c": null, "e": [["tile_f", "sp", [6, 1, 16, 4]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [48, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00018865015442561206], 0, 1.2870540618896484, 1538735842.1039355], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1492011, "c": null, "e": [["tile_f", "sp", [6, 2, 16, 2]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [64, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00018899474528301887], 0, 1.1503286361694336, 1538737294.7999964], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8302015, "c": null, "e": [["tile_f", "sp", [8, 2, 10, 2]], ["tile_y", "sp", [1, 2, 4, 1]], ["tile_x", "sp", [1, 1, 4, 2]], ["tile_rc", "sp", [64, 20]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00016675274875207986], 0, 9.518572330474854, 1538738244.635086], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13428977, "c": null, "e": [["tile_f", "sp", [3, 4, 16, 1]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 2, 4]], ["tile_rc", "sp", [96, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00018094568173598554], 0, 1.1965785026550293, 1538741184.7944498], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13428971, "c": null, "e": [["tile_f", "sp", [5, 4, 16, 1]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 2, 4]], ["tile_rc", "sp", [96, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00022400241741071428], 0, 6.386166572570801, 1538742171.4319954], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 783342, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0005600653463687151], 0, 1.6617372035980225, 1538744790.6790981], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 337140, "c": null, "e": [["tile_f", "sp", [4, 2, 12, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.000562955595505618], 0, 13.148250341415405, 1538746218.8638897], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 669700, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00044481565777777776], 0, 1.6090643405914307, 1538747566.9998116], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 321089, "c": null, "e": [["tile_f", "sp", [4, 2, 10, 2]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00039756966666666666], 0, 1.2615001201629639, 1538748741.3635602], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 670446, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0004707163849765258], 0, 1.5564532279968262, 1538750318.50236], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 321089, "c": null, "e": [["tile_f", "sp", [4, 2, 10, 2]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00039829519444444444], 0, 1.8980224132537842, 1538751516.9735677], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 342982, "c": null, "e": [["tile_f", "sp", [5, 4, 8, 1]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [96, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0003170707721518987], 0, 11.788979053497314, 1538752849.427508], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 449284, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0003626276859205776], 0, 1.572218656539917, 1538753895.1277938], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 160735, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00031408786249999997], 0, 7.100447654724121, 1538755100.678874], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 450030, "c": null, "e": [["tile_f", "sp", [6, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00038391145977011496], 0, 4.594708442687988, 1538756344.7816327], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 160495, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00029028242363112394], 0, 1.5227651596069336, 1538757551.6295998], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 45535, "c": null, "e": [["tile_f", "sp", [4, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [96, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0003036941212121212], 0, 1.1823575496673584, 1538758601.7473073], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 132180, "c": null, "e": [["tile_f", "sp", [4, 2, 12, 2]], ["tile_y", "sp", [1, 17, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [64, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0003279414786885246], 0, 1.2324786186218262, 1538759576.7236032], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 392487, "c": null, "e": [["tile_f", "sp", [2, 8, 6, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [48, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00047008414084507037], 0, 21.658472061157227, 1538761347.4519036], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2079038, "c": null, "e": [["tile_f", "sp", [12, 2, 8, 2]], ["tile_y", "sp", [1, 1, 1, 17]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [288, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0030419013636363637], 0, 2.1014552116394043, 1538762292.0791416], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 161489, "c": null, "e": [["tile_f", "sp", [1, 4, 6, 2]], ["tile_y", "sp", [7, 5, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00013624694277929156], 0, 1.8275806903839111, 1538763931.0236416], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 900837, "c": null, "e": [["tile_f", "sp", [1, 8, 8, 1]], ["tile_y", "sp", [7, 1, 5, 1]], ["tile_x", "sp", [1, 1, 5, 7]], ["tile_rc", "sp", [36, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002305771013824885], 0, 1.228672742843628, 1538765047.988771], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1739571, "c": null, "e": [["tile_f", "sp", [1, 3, 8, 2]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00015134447734138973], 0, 10.06118893623352, 1538766294.2085798], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 248093, "c": null, "e": [["tile_f", "sp", [2, 4, 4, 2]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00019819382411067195], 0, 5.968884706497192, 1538767651.9267356], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 853510, "c": null, "e": [["tile_f", "sp", [1, 8, 4, 1]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.956719666931006e-05], 0, 10.136065006256104, 1538768634.9824493], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4247211, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 8, 6]], ["tile_x", "sp", [9, 3, 12, 1]], ["tile_rc", "sp", [8, 12]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00050543504], 0, 1.2918884754180908, 1538772108.8917615], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2526009, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 2, 3, 8]], ["tile_x", "sp", [3, 3, 36, 1]], ["tile_rc", "sp", [4, 16]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0003526407711267606], 0, 2.206836223602295, 1538773225.934845], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 4965517, "c": null, "e": [["tile_f", "sp", [2, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 5, 7]], ["tile_rc", "sp", [48, 1]], ["tile_ry", "sp", [1, 5]], ["tile_rx", "sp", [1, 5]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0005109715357142857], 0, 2.0362069606781006, 1538775383.4009142], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1666769, "c": null, "e": [["tile_f", "sp", [1, 4, 6, 2]], ["tile_y", "sp", [7, 5, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[9.65264927815207e-05], 0, 1.7492220401763916, 1538776552.4567373], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 76036, "c": null, "e": [["tile_f", "sp", [2, 8, 4, 1]], ["tile_y", "sp", [7, 1, 1, 5]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00015126252416918428], 0, 12.649672508239746, 1538777934.6820767], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1240584, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [71, 1, 1, 1]], ["tile_x", "sp", [1, 1, 1, 71]], ["tile_rc", "sp", [80, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.004486951652173913], 0, 1.9496736526489258, 1538779494.6178966], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34727, "c": null, "e": [["tile_f", "sp", [4, 20, 1, 1]], ["tile_y", "sp", [73, 1, 1, 1]], ["tile_x", "sp", [1, 1, 73, 1]], ["tile_rc", "sp", [32, 2]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00042778770638297875], 0, 4.8698039054870605, 1538782207.6844997], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15395544, "c": null, "e": [["tile_f", "sp", [1, 4, 16, 1]], ["tile_y", "sp", [147, 1, 1, 1]], ["tile_x", "sp", [1, 7, 7, 3]], ["tile_rc", "sp", [32, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0026279461282051283], 0, 1.4585211277008057, 1538783344.04221], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5944096, "c": null, "e": [["tile_f", "sp", [1, 4, 4, 2]], ["tile_y", "sp", [21, 1, 1, 7]], ["tile_x", "sp", [7, 1, 21, 1]], ["tile_rc", "sp", [32, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.001143458465909091], 0, 1.274094581604004, 1538784938.3402982], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 29168, "c": null, "e": [["tile_f", "sp", [1, 4, 1, 8]], ["tile_y", "sp", [149, 1, 1, 1]], ["tile_x", "sp", [1, 1, 149, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [3, 1]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0003318904620462046], 0, 1.3958790302276611, 1538786276.7700129], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 8750147, "e": [["tile_f", "sp", [4, 4, 16, 1]], ["tile_y", "sp", [1, 2, 7, 1]], ["tile_x", "sp", [2, 7, 1, 1]], ["tile_rc", "sp", [256, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0007959011], 0, 5.501811265945435, 1554974881.5541902], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 675490, "e": [["tile_f", "sp", [8, 1, 32, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 1, 7]], ["tile_rc", "sp", [256, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0010813246], 0, 2.48030948638916, 1554957441.953887], "v": 0.1}
+{"i": ["cuda -model=tx2", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "i": 35599042, "e": [["tile_f", "sp", [2, 1, 16, 4]], ["tile_y", "sp", [14, 1, 1, 2]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [64, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "c": null}], "r": [[0.0008413876000000001], 0, 1.2982349395751953, 1554963324.6987908], "v": 0.1}
diff --git a/tophub/mali_v0.06.log b/tophub/mali_v0.06.log
new file mode 100644
index 0000000..4ae5c37
--- /dev/null
+++ b/tophub/mali_v0.06.log
@@ -0,0 +1,530 @@
+# This is the pre-tuned parameters for mali backend
+# TVM downloaded this during compilation
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 1024], "float32"], ["TENSOR", [1000, 1024], "float32"], ["TENSOR", [1000], "float32"], null], {}, ["dense", [1, 1024, "float32"], [1000, 1024, "float32"], [1000, "float32"], 0], {"i": 407, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [40, 25, 1]], ["c_unroll", "sp", [64, 16]]], "c": null, "t": "direct"}], "r": [[0.0008410332727272727], 0, 0.9504795074462891, 1538219132.0322816], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 7, 7, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 36225, "c": null, "e": [["tile_co", "sp", [32, 2, 2]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012345162650602409], 0, 1.783360242843628, 1538220265.3881273], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6185, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0012251021829268293], 0, 3.442169666290283, 1538220868.9772196], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 7, 7, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14194, "c": null, "e": [["tile_co", "sp", [16, 4, 2]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011703238390804597], 0, 1.7037286758422852, 1538221457.6329682], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 7, 7, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7480, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0011712368295454544], 0, 0.7852833271026611, 1538222207.7694445], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 7, 7, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34902, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011033985934065934], 0, 4.56376838684082, 1538223052.2564673], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 7, 7, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 28377, "c": null, "e": [["tile_co", "sp", [32, 2, 2]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0011010124565217392], 0, 2.8353447914123535, 1538223698.1780994], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 7, 7, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14813, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010267909897959183], 0, 2.883234977722168, 1538224834.7731726], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 7, 7, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 13834, "c": null, "e": [["tile_co", "sp", [16, 4, 2]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009992093168316833], 0, 0.9072630405426025, 1538225309.7437131], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 7, 7, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 33724, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009668028190476191], 0, 1.7385807037353516, 1538226391.2453725], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 7, 7, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26596, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0009314369444444444], 0, 3.3153698444366455, 1538227012.8132076], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 7, 7, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 35576, "c": null, "e": [["tile_co", "sp", [64, 1, 2]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009047837477477477], 0, 7.748708009719849, 1538227676.3741627], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 7, 7, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27946, "c": null, "e": [["tile_co", "sp", [16, 4, 2]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0008836463620689655], 0, 4.92792010307312, 1538228352.9691577], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 7, 7, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 33833, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008398404132231405], 0, 3.9947926998138428, 1538229385.4520714], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15461, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000791507859375], 0, 3.981132984161377, 1538229751.3026388], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 7, 7, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6509, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0007436436518518518], 0, 4.596225738525391, 1538230465.728221], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 7, 7, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 14587, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 128]], ["tile_t2", "sp", [32, 1]], ["c_unroll", "sp", [128, 1]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.00035657818245614034], 0, 2.836203098297119, 1538231356.854444], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7193, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0007103838581560283], 0, 3.733657121658325, 1538232090.0174997], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 492272, "c": null, "e": [["tile_co", "sp", [16, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009498708636363637], 0, 2.043750524520874, 1538232816.5858936], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float32"], ["TENSOR", [128, 992, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 992, 14, 14, "float32"], [128, 992, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 307638, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00310028303030303], 0, 1.6421899795532227, 1538234303.9461672], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float32"], ["TENSOR", [128, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 14, 14, "float32"], [128, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 310554, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002976414205882353], 0, 2.677269458770752, 1538234620.1172462], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float32"], ["TENSOR", [128, 928, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 928, 14, 14, "float32"], [128, 928, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 322217, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002861784742857143], 0, 2.8234505653381348, 1538235442.621727], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float32"], ["TENSOR", [128, 896, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 896, 14, 14, "float32"], [128, 896, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 316385, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027940118333333337], 0, 1.793832540512085, 1538236281.0484364], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float32"], ["TENSOR", [128, 864, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 864, 14, 14, "float32"], [128, 864, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 319302, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026788273684210527], 0, 4.41715145111084, 1538237170.4014416], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float32"], ["TENSOR", [128, 832, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 832, 14, 14, "float32"], [128, 832, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 313469, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026042461025641028], 0, 4.003000736236572, 1538238049.7132258], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float32"], ["TENSOR", [128, 800, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 800, 14, 14, "float32"], [128, 800, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 310553, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0024737304146341464], 0, 0.8378098011016846, 1538238780.1887836], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 14, 14, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 120877, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027737962432432435], 0, 2.659139394760132, 1538239480.5729156], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float32"], ["TENSOR", [128, 736, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 736, 14, 14, "float32"], [128, 736, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 304722, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022991877272727276], 0, 1.9585888385772705, 1538240201.5682926], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float32"], ["TENSOR", [128, 704, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 704, 14, 14, "float32"], [128, 704, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 248353, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0026568285263157895], 0, 5.755255699157715, 1538240895.6482584], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float32"], ["TENSOR", [128, 672, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 672, 14, 14, "float32"], [128, 672, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 316385, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002133825340425532], 0, 1.9175612926483154, 1538242043.072054], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float32"], ["TENSOR", [128, 640, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 640, 14, 14, "float32"], [128, 640, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 325133, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002041318408163265], 0, 5.894289970397949, 1538242938.3826404], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float32"], ["TENSOR", [128, 608, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 608, 14, 14, "float32"], [128, 608, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 129625, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022064258260869564], 0, 6.015295743942261, 1538243829.0504029], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [128, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [128, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 310588, "c": null, "e": [["tile_co", "sp", [16, 2, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018264108545454547], 0, 0.8541479110717773, 1538245025.7590194], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float32"], ["TENSOR", [128, 544, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 544, 14, 14, "float32"], [128, 544, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 132541, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001961382568627451], 0, 0.8074092864990234, 1538245828.2283144], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 325357, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018914905471698114], 0, 1.9953467845916748, 1538246850.82206], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 14, 14, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 304756, "c": null, "e": [["tile_co", "sp", [16, 2, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015727076060606062], 0, 2.9614222049713135, 1538248161.6806047], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 14, 14, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 127033, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016735922459016393], 0, 1.8286712169647217, 1538248567.264625], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 14, 14, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 316385, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013691685733333333], 0, 0.8670430183410645, 1538249801.5412633], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 249001, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001516534552238806], 0, 1.026026964187622, 1538250861.473614], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 14, 14, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 325459, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011761046627906976], 0, 7.373374938964844, 1538251873.172054], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 14, 14, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 304755, "c": null, "e": [["tile_co", "sp", [32, 1, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012063289782608694], 0, 1.0598692893981934, 1538252413.399592], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 14, 14, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 117997, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011222547444444445], 0, 1.6311333179473877, 1538253408.1050537], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 14, 14, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 17651, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [32, 4]], ["tile_t2", "sp", [2, 16]], ["c_unroll", "sp", [32, 4]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.0009855709626168225], 0, 2.25718355178833, 1538254102.5614228], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 307637, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009218025321100917], 0, 1.9833705425262451, 1538255099.5800583], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1590502, "c": null, "e": [["tile_co", "sp", [2, 32, 4]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [14, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0101334866], 0, 4.479913949966431, 1538256652.7894275], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float32"], ["TENSOR", [128, 480, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 480, 28, 28, "float32"], [128, 480, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 543043, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004987010904761905], 0, 0.9149727821350098, 1538257302.1150181], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float32"], ["TENSOR", [128, 448, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 28, 28, "float32"], [128, 448, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1005428, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [4, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00518794495], 0, 1.857978105545044, 1538258394.5055037], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float32"], ["TENSOR", [128, 416, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 416, 28, 28, "float32"], [128, 416, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 554707, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004397177217391304], 0, 5.77596116065979, 1538259367.703746], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float32"], ["TENSOR", [128, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 28, 28, "float32"], [128, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1290188, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00400945984], 0, 5.800011396408081, 1538260616.7798574], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float32"], ["TENSOR", [128, 352, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 352, 28, 28, "float32"], [128, 352, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 543692, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003726450814814815], 0, 5.753964900970459, 1538261776.9915953], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float32"], ["TENSOR", [128, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 28, 28, "float32"], [128, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1266860, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0034079587333333333], 0, 3.8111042976379395, 1538263080.805549], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float32"], ["TENSOR", [128, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 28, 28, "float32"], [128, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 497036, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0030983644242424243], 0, 0.9946346282958984, 1538263669.830712], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1300771, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027880089729729733], 0, 3.930661201477051, 1538265020.223462], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 28, 28, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1251019, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [2, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002671627052631579], 0, 2.776970148086548, 1538266112.4372432], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1242450, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021923815], 0, 5.4998619556427, 1538267618.4328883], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 28, 28, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 554705, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018831142407407407], 0, 0.841820240020752, 1538267827.086937], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11712, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 16]], ["tile_t2", "sp", [2, 16]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0033791028333333335], 0, 9.762807130813599, 1538269143.625517], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1289106, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015545827538461538], 0, 11.107493877410889, 1538270464.2763855], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3420379, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0103995151], 0, 2.7131006717681885, 1538272200.8851922], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float32"], ["TENSOR", [128, 224, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 224, 56, 56, "float32"], [128, 224, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3549979, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009195981363636365], 0, 1.0961134433746338, 1538273341.2677042], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float32"], ["TENSOR", [128, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 56, 56, "float32"], [128, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3577267, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [28, 1, 2]], ["tile_ow", "sp", [2, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008165028153846154], 0, 0.9725632667541504, 1538274802.4149766], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float32"], ["TENSOR", [128, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 56, 56, "float32"], [128, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3416346, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [28, 1, 2]], ["tile_ow", "sp", [1, 28, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007003896333333333], 0, 1.0899279117584229, 1538275230.4090729], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [128, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [128, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2662722, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [14, 1, 4]], ["tile_ow", "sp", [4, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004551849772727273], 0, 1.0746421813964844, 1538276688.247726], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [32, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [32, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 11737, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [4, 32]], ["tile_t2", "sp", [2, 16]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.010623397599999999], 0, 15.461644172668457, 1538277342.094826], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 901946, "c": null, "e": [["tile_co", "sp", [6, 8, 4]], ["tile_oh", "sp", [2, 1, 4]], ["tile_ow", "sp", [1, 8, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003093141393939394], 0, 0.8539721965789795, 1538278723.3299742], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 383911, "c": null, "e": [["tile_co", "sp", [8, 14, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005664537444444445], 0, 4.823970317840576, 1538280064.59884], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 483991, "c": null, "e": [["tile_co", "sp", [12, 8, 4]], ["tile_oh", "sp", [4, 2, 1]], ["tile_ow", "sp", [1, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004452871956521739], 0, 4.555937051773071, 1538281364.5627975], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 896311, "c": null, "e": [["tile_co", "sp", [8, 10, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004279294625], 0, 1.0510468482971191, 1538281943.2283947], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 907907, "c": null, "e": [["tile_co", "sp", [16, 3, 4]], ["tile_oh", "sp", [1, 2, 4]], ["tile_ow", "sp", [8, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019612860392156863], 0, 0.8549942970275879, 1538282888.7653813], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 47157, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 2], ["tile_t1", "sp", [112, 4]], ["tile_t2", "sp", [96, 4]], ["c_unroll", "sp", [224, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.00531834047368421], 0, 14.510762929916382, 1538283841.445443], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 921511, "c": null, "e": [["tile_co", "sp", [8, 14, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035744987857142855], 0, 2.635505437850952, 1538285177.5642416], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 1199600, "c": null, "e": [["tile_co", "sp", [8, 12, 4]], ["tile_oh", "sp", [2, 1, 4]], ["tile_ow", "sp", [8, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001974696568627451], 0, 2.0515048503875732, 1538286483.4281964], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 1152513, "c": null, "e": [["tile_co", "sp", [6, 16, 4]], ["tile_oh", "sp", [4, 2, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0020042177647058824], 0, 4.1319239139556885, 1538287961.675134], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1142898, "c": null, "e": [["tile_co", "sp", [16, 6, 4]], ["tile_oh", "sp", [2, 4, 1]], ["tile_ow", "sp", [1, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028543698333333336], 0, 2.809438705444336, 1538288783.3593895], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 358711, "c": null, "e": [["tile_co", "sp", [8, 10, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027366005789473684], 0, 1.054525375366211, 1538289787.9132423], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 905561, "c": null, "e": [["tile_co", "sp", [2, 24, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [1, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026406593947368422], 0, 11.853938817977905, 1538291270.600422], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 897151, "c": null, "e": [["tile_co", "sp", [8, 10, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [1, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0036642411724137933], 0, 5.887054443359375, 1538291886.969703], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 63558, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.011833980777777777], 0, 6.884978771209717, 1538293164.5629773], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 16951, "c": null, "e": [["tile_co", "sp", [1, 12, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.010820508400000001], 0, 1.0737414360046387, 1538293752.6717374], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 80946, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009603040545454546], 0, 4.703333377838135, 1538294712.964426], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 62031, "c": null, "e": [["tile_co", "sp", [4, 5, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008599302166666666], 0, 5.004313230514526, 1538295442.024167], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 83970, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.009755582363636363], 0, 1.136061191558838, 1538296097.242431], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 26675, "c": null, "e": [["tile_co", "sp", [5, 8, 4]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00843149725], 0, 3.9440596103668213, 1538296741.1481023], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 45401, "c": null, "e": [["tile_co", "sp", [1, 20, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.006762913266666667], 0, 1.6746673583984375, 1538297509.9309263], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 62046, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007697260153846154], 0, 5.472573757171631, 1538298279.3446426], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 6617, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005569351944444444], 0, 5.74540638923645, 1538299033.025988], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 63558, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.007915224307692308], 0, 4.111681938171387, 1538299697.842241], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 6293, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005384252052631579], 0, 3.2678561210632324, 1538300477.874018], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6293, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0049474270476190475], 0, 3.3933444023132324, 1538301581.5805511], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34337, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006945903133333334], 0, 1.8829166889190674, 1538302637.020568], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 62971, "c": null, "e": [["tile_co", "sp", [2, 12, 4]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003962862692307692], 0, 6.376277685165405, 1538302959.3711205], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 44774, "c": null, "e": [["tile_co", "sp", [6, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.041978142], 0, 7.7385413646698, 1538303670.8653693], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 384373, "c": null, "e": [["tile_co", "sp", [1, 12, 4]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0022062572391304347], 0, 4.3278889656066895, 1538304539.73653], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 93264, "c": null, "e": [["tile_co", "sp", [1, 4, 16]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 5, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0041348406], 0, 2.6646549701690674, 1538305463.1177087], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 172963, "c": null, "e": [["tile_co", "sp", [1, 12, 4]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019914488823529414], 0, 0.9817342758178711, 1538306583.0554066], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 245969, "c": null, "e": [["tile_co", "sp", [1, 16, 4]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002416825666666667], 0, 4.82902193069458, 1538307282.9633918], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 179584, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [7, 5, 1]], ["tile_ow", "sp", [1, 7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001231293373493976], 0, 6.851509094238281, 1538308378.8596547], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 59786, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [24, 4]], ["tile_t2", "sp", [6, 16]], ["c_unroll", "sp", [16, 6]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.00632436325], 0, 3.6381335258483887, 1538309043.413901], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 20187, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 4]], ["tile_t2", "sp", [6, 16]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0040734974], 0, 1.1747214794158936, 1538309857.7112935], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 252771, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007183659214285714], 0, 7.255812644958496, 1538311104.459706], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 359300, "c": null, "e": [["tile_co", "sp", [4, 4, 3]], ["tile_oh", "sp", [7, 5, 1]], ["tile_ow", "sp", [1, 7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.002339198162790698], 0, 1.0177745819091797, 1538311707.5983207], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 52440, "c": null, "e": [["tile_co", "sp", [1, 4, 16]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 5, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0027788449444444445], 0, 3.2167224884033203, 1538312596.7125585], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 83225, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [71, 1, 1]], ["tile_ow", "sp", [71, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.098159602], 0, 13.18534803390503, 1538313661.8016596], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 7727, "c": null, "e": [["tile_co", "sp", [1, 10, 8]], ["tile_oh", "sp", [73, 1, 1]], ["tile_ow", "sp", [73, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.006182748], 0, 5.448895454406738, 1538314361.632059], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 429926, "c": null, "e": [["tile_co", "sp", [8, 2, 4]], ["tile_oh", "sp", [147, 1, 1]], ["tile_ow", "sp", [7, 7, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0448735845], 0, 5.395364999771118, 1538316332.9785912], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 757904, "c": null, "e": [["tile_co", "sp", [1, 8, 4]], ["tile_oh", "sp", [147, 1, 1]], ["tile_ow", "sp", [7, 7, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.019910080333333333], 0, 1.1186540126800537, 1538317451.7706919], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8711, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [149, 1, 1]], ["tile_ow", "sp", [149, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004713186727272728], 0, 1.125840663909912, 1538318266.8900588], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float32"], ["TENSOR", [1000, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 1, 1, "float32"], [1000, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 10728, "c": null, "e": [["tile_co", "sp", [250, 1, 4]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011362910222222223], 0, 0.8758072853088379, 1538318910.1584704], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float32"], ["TENSOR", [1280, 320, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 320, 7, 7, "float32"], [1280, 320, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 129501, "c": null, "e": [["tile_co", "sp", [80, 4, 4]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002977701382352941], 0, 3.4135563373565674, 1538319564.1158729], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [320, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [320, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 33644, "c": null, "e": [["tile_co", "sp", [20, 2, 8]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002764596108108108], 0, 4.536156415939331, 1538320266.65954], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [160, 960, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 960, 7, 7, "float32"], [160, 960, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 21762, "c": null, "e": [["tile_co", "sp", [4, 10, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.00168144845], 0, 4.475642442703247, 1538320866.5682778], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float32"], ["TENSOR", [960, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float32"], [960, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 17673, "c": null, "e": [["tile_c", "sp", [80, 6, 2]], ["tile_y", "sp", [1, 7, 1]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0003460237542662116], 0, 0.9481427669525146, 1538321455.6322298], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float32"], ["TENSOR", [960, 160, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 7, 7, "float32"], [960, 160, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 103641, "c": null, "e": [["tile_co", "sp", [60, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013231910263157894], 0, 3.4934847354888916, 1538322329.2898219], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float32"], ["TENSOR", [160, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 7, 7, "float32"], [160, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8721, "c": null, "e": [["tile_co", "sp", [4, 10, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "unroll"]]], "t": "direct"}], "r": [[0.0010643950736842105], 0, 9.561098337173462, 1538323070.1605473], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 3868, "c": null, "e": [["tile_c", "sp", [96, 6, 1]], ["tile_y", "sp", [1, 1, 7]], ["tile_x", "sp", [1, 7, 1]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0004189379958677686], 0, 2.461210250854492, 1538323770.3301597], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [96, 576, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 576, 14, 14, "float32"], [96, 576, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 558839, "c": null, "e": [["tile_co", "sp", [12, 2, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014517212898550725], 0, 5.802470684051514, 1538325082.4654295], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float32"], ["TENSOR", [576, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float32"], [576, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 264772, "c": null, "e": [["tile_c", "sp", [96, 6, 1]], ["tile_y", "sp", [7, 2, 1]], ["tile_x", "sp", [7, 1, 2]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0005632651945945946], 0, 0.8403832912445068, 1538325297.6404765], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float32"], ["TENSOR", [576, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 14, 14, "float32"], [576, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1476440, "c": null, "e": [["tile_co", "sp", [36, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012728439873417722], 0, 3.9396111965179443, 1538326108.6378531], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [96, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [96, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 543530, "c": null, "e": [["tile_co", "sp", [12, 2, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010562107216494847], 0, 0.9871692657470703, 1538327366.8222942], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 14, 14, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 91831, "c": null, "e": [["tile_co", "sp", [4, 2, 8]], ["tile_oh", "sp", [1, 14, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008975298947368421], 0, 1.9642689228057861, 1538328275.460502], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float32"], ["TENSOR", [384, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float32"], [384, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 108976, "c": null, "e": [["tile_c", "sp", [64, 6, 1]], ["tile_y", "sp", [7, 2, 1]], ["tile_x", "sp", [7, 1, 2]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0003887394031007752], 0, 1.04939603805542, 1538329208.6015944], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float32"], ["TENSOR", [384, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 14, 14, "float32"], [384, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 975389, "c": null, "e": [["tile_co", "sp", [24, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007483880223880597], 0, 0.9486618041992188, 1538330204.8111427], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 14, 14, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 239275, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004616047863636364], 0, 3.9898664951324463, 1538331242.3884757], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 105172, "c": null, "e": [["tile_c", "sp", [32, 6, 1]], ["tile_y", "sp", [7, 2, 1]], ["tile_x", "sp", [7, 1, 2]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.00048510188888888886], 0, 0.8756880760192871, 1538332014.3610058], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 28, 28, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 744043, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [14, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009568258679245284], 0, 4.936660528182983, 1538332783.0402787], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float32"], ["TENSOR", [192, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float32"], [192, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 48402, "c": null, "e": [["tile_c", "sp", [16, 6, 2]], ["tile_y", "sp", [28, 1, 1]], ["tile_x", "sp", [1, 4, 7]], ["ann_spatial", "an", ["none", "none", "unroll"]]], "t": "direct"}], "r": [[0.0007704952538461538], 0, 3.2735280990600586, 1538333856.746618], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float32"], ["TENSOR", [192, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 28, 28, "float32"], [192, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2238636, "c": null, "e": [["tile_co", "sp", [12, 4, 4]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [2, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0008696808434782609], 0, 0.8544929027557373, 1538334844.0878804], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float32"], ["TENSOR", [32, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 28, 28, "float32"], [32, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 296426, "c": null, "e": [["tile_co", "sp", [8, 1, 4]], ["tile_oh", "sp", [4, 7, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007208077785714285], 0, 3.8276727199554443, 1538336095.7105505], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 132934, "c": null, "e": [["tile_c", "sp", [24, 6, 1]], ["tile_y", "sp", [14, 2, 1]], ["tile_x", "sp", [7, 1, 4]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0012693614625], 0, 1.0440986156463623, 1538337179.0411355], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [24, 144, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 144, 56, 56, "float32"], [24, 144, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1120550, "c": null, "e": [["tile_co", "sp", [2, 3, 4]], ["tile_oh", "sp", [28, 2, 1]], ["tile_ow", "sp", [7, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0020960426041666666], 0, 2.749624013900757, 1538338028.0462437], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float32"], ["TENSOR", [144, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float32"], [144, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 364504, "c": null, "e": [["tile_c", "sp", [24, 6, 1]], ["tile_y", "sp", [56, 1, 1]], ["tile_x", "sp", [7, 2, 4]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0016924053833333335], 0, 1.1420512199401855, 1538339330.142449], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float32"], ["TENSOR", [144, 24, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 24, 56, 56, "float32"], [144, 24, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3690940, "c": null, "e": [["tile_co", "sp", [9, 4, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001836336509090909], 0, 0.9741520881652832, 1538340477.3736506], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float32"], ["TENSOR", [24, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 56, 56, "float32"], [24, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1203321, "c": null, "e": [["tile_co", "sp", [1, 6, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014784792647058822], 0, 11.301132440567017, 1538341957.5027587], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float32"], ["TENSOR", [96, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float32"], [96, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 708754, "c": null, "e": [["tile_c", "sp", [16, 6, 1]], ["tile_y", "sp", [56, 1, 1]], ["tile_x", "sp", [7, 2, 4]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0030467906363636364], 0, 10.850241899490356, 1538343530.652503], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float32"], ["TENSOR", [96, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 112, 112, "float32"], [96, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 14048590, "c": null, "e": [["tile_co", "sp", [6, 4, 4]], ["tile_oh", "sp", [28, 1, 4]], ["tile_ow", "sp", [8, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004229807958333333], 0, 6.609508752822876, 1538344776.175978], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [16, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [16, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3295361, "c": null, "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [2, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016980724745762712], 0, 3.640692949295044, 1538346282.499106], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [32, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1721803, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [2, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002463403], 0, 2.841625213623047, 1538347943.6641102], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 1024], "float32"], ["TENSOR", [1000, 1024], "float32"], null, null], {}, ["dense", [1, 1024, "float32"], [1000, 1024, "float32"], 0, 0], {"i": 407, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [40, 25, 1]], ["c_unroll", "sp", [64, 16]]], "c": null, "t": "direct"}], "r": [[0.0008327435573770492], 0, 3.116755485534668, 1538348112.5131621], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 63977, "c": null, "e": [["tile_co", "sp", [64, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007325475600000001], 0, 1.8138630390167236, 1535374999.9505277], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 7564, "c": null, "e": [["tile_c", "sp", [16, 4, 16]], ["tile_y", "sp", [7, 1, 1]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.000412161125], 0, 0.7525043487548828, 1534782760.153166], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25367, "c": null, "e": [["tile_co", "sp", [64, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0036721981000000002], 0, 3.881497859954834, 1535375736.392116], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 8527, "c": null, "e": [["tile_c", "sp", [128, 4, 1]], ["tile_y", "sp", [1, 1, 7]], ["tile_x", "sp", [7, 1, 1]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.0003942965], 0, 0.432330846786499, 1534782059.079115], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 497274, "c": null, "e": [["tile_co", "sp", [128, 1, 4]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004716177], 0, 2.6632957458496094, 1535376695.0092752], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 47082, "c": null, "e": [["tile_c", "sp", [128, 4, 1]], ["tile_y", "sp", [7, 2, 1]], ["tile_x", "sp", [1, 7, 2]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005478352], 0, 3.827186107635498, 1535377399.6289985], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 478906, "c": null, "e": [["tile_co", "sp", [32, 4, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0024410087], 0, 3.8536217212677, 1535377744.7710824], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 56342, "c": null, "e": [["tile_c", "sp", [64, 4, 1]], ["tile_y", "sp", [7, 2, 1]], ["tile_x", "sp", [7, 1, 2]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0006491035], 0, 0.49543166160583496, 1534780068.517396], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 650452, "c": null, "e": [["tile_co", "sp", [2, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004925248375], 0, 0.5273537635803223, 1534779379.078077], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 110207, "c": null, "e": [["tile_c", "sp", [64, 4, 1]], ["tile_y", "sp", [14, 2, 1]], ["tile_x", "sp", [7, 1, 4]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0008478718], 0, 1.851642370223999, 1535378841.8770516], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1567640, "c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026343824], 0, 0.6797151565551758, 1535379471.8211994], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 228134, "c": null, "e": [["tile_c", "sp", [32, 4, 1]], ["tile_y", "sp", [14, 2, 1]], ["tile_x", "sp", [7, 1, 4]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0011536251000000002], 0, 0.6222481727600098, 1535380022.245833], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1508778, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00551769375], 0, 0.48728442192077637, 1534777149.812052], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 413714, "c": null, "e": [["tile_c", "sp", [32, 4, 1]], ["tile_y", "sp", [14, 4, 1]], ["tile_x", "sp", [1, 7, 8]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0014407702], 0, 3.0985889434814453, 1535381111.3575757], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3582378, "c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0031562310000000002], 0, 2.764233112335205, 1535381578.816183], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 390602, "c": null, "e": [["tile_c", "sp", [16, 4, 1]], ["tile_y", "sp", [56, 1, 1]], ["tile_x", "sp", [7, 2, 4]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0021837009], 0, 4.606369733810425, 1535382047.4107864], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2522535, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [2, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004092848125], 0, 0.5883266925811768, 1534774593.4489887], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 667172, "c": null, "e": [["tile_c", "sp", [8, 4, 1]], ["tile_y", "sp", [112, 1, 1]], ["tile_x", "sp", [2, 7, 8]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0014363367], 0, 4.407003164291382, 1535384067.9484398], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4528453, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [2, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00200064425], 0, 0.7528524398803711, 1534773368.5821507], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27993, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000986561625], 0, 0.4286386966705322, 1534790875.5712688], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 13, 13, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 45124, "c": null, "e": [["tile_co", "sp", [3, 2, 8]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [1, 13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000964101125], 0, 0.45015430450439453, 1534790533.7602704], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 65809, "c": null, "e": [["tile_co", "sp", [8, 6, 4]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0022707995], 0, 0.7499988079071045, 1534790026.0973186], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 13, 13, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 81715, "c": null, "e": [["tile_co", "sp", [1, 12, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000471697625], 0, 0.5158183574676514, 1534789467.3603568], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 13, 13, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 34909, "c": null, "e": [["tile_co", "sp", [3, 2, 8]], ["tile_oh", "sp", [1, 13, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000679179875], 0, 0.4952850341796875, 1534789066.4712656], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 27, 27, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 99553, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [9, 1, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000670541125], 0, 0.45807671546936035, 1534786969.579846], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float32"], ["TENSOR", [16, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 55, 55, "float32"], [16, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 130556, "c": null, "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000894939625], 0, 0.43424487113952637, 1534785262.3156724], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2465835, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [2, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003503452], 0, 0.631648063659668, 1534784665.5832863], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [1000, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [1000, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 39429, "c": null, "e": [["tile_co", "sp", [125, 2, 4]], ["tile_oh", "sp", [1, 13, 1]], ["tile_ow", "sp", [1, 1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012538046625], 0, 0.7728147506713867, 1534792723.6829607], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 13112, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 64]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.001856853125], 0, 0.8097231388092041, 1534791707.3338065], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 13, 13, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 33244, "c": null, "e": [["tile_co", "sp", [1, 16, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000641737375], 0, 0.44287848472595215, 1534791308.1715407], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float32"], ["TENSOR", [64, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 13, 13, "float32"], [64, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 27741, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00126302175], 0, 0.4725668430328369, 1534792188.234129], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 18787, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 4]], ["tile_t2", "sp", [32, 8]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.005588395944444445], 0, 6.1490254402160645, 1538376068.1898043], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 27, 27, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 369199, "c": null, "e": [["tile_co", "sp", [16, 4, 4]], ["tile_oh", "sp", [9, 1, 3]], ["tile_ow", "sp", [27, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0018697662962962962], 0, 1.701404333114624, 1538376841.6628475], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [64, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [64, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 127697, "c": null, "e": [["tile_co", "sp", [1, 16, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026870259210526313], 0, 0.8440773487091064, 1538377712.0940447], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float32"], ["TENSOR", [48, 384, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 27, 27, "float32"], [48, 384, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 205228, "c": null, "e": [["tile_co", "sp", [1, 12, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021594435957446806], 0, 3.4219133853912354, 1538379032.4608843], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 357879, "c": null, "e": [["tile_co", "sp", [4, 12, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00514746175], 0, 4.5888519287109375, 1538380304.8954253], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float32"], ["TENSOR", [192, 48, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 27, 27, "float32"], [192, 48, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 380585, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 27, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013880688630136986], 0, 0.8284728527069092, 1538381483.2091045], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 191728, "c": null, "e": [["tile_co", "sp", [1, 12, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015026609701492537], 0, 0.8274130821228027, 1538382583.1104581], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 387379, "c": null, "e": [["tile_co", "sp", [2, 16, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002533298875], 0, 0.6990301609039307, 1534787829.1735094], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 27, 27, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 156257, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [9, 1, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000720014875], 0, 0.44545745849609375, 1534787339.650611], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float32"], ["TENSOR", [32, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 27, 27, "float32"], [32, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 99553, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [9, 1, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00112849375], 0, 0.4669630527496338, 1534788519.350505], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 324989, "c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00979368418181818], 0, 2.642887592315674, 1538386305.035], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float32"], ["TENSOR", [128, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 55, 55, "float32"], [128, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 132540, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002558397], 0, 2.0922298431396484, 1538387352.3431842], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [32, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [32, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 142981, "c": null, "e": [["tile_co", "sp", [2, 4, 4]], ["tile_oh", "sp", [11, 1, 5]], ["tile_ow", "sp", [55, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0024637076585365854], 0, 1.844597578048706, 1538387843.6667852], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float32"], ["TENSOR", [16, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 55, 55, "float32"], [16, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 135416, "c": null, "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001490051], 0, 0.4474771022796631, 1534786546.7636065], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 252771, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00299680025], 0, 0.9511630535125732, 1534785996.6508455], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float32"], ["TENSOR", [64, 16, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 16, 55, 55, "float32"], [64, 16, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 107619, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008865205], 0, 0.48459386825561523, 1534785643.458716], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float32"], ["TENSOR", [16, 96, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 55, 55, "float32"], [16, 96, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 54011, "c": null, "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011997278452380953], 0, 0.8168883323669434, 1538391589.1191638], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [96, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [96, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 14039582, "c": null, "e": [["tile_co", "sp", [4, 6, 4]], ["tile_oh", "sp", [56, 1, 2]], ["tile_ow", "sp", [28, 4, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.019280515], 0, 2.3462038040161133, 1538393197.0185685], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 4096], "float32"], ["TENSOR", [1000, 4096], "float32"], ["TENSOR", [1000], "float32"], null], {}, ["dense", [1, 4096, "float32"], [1000, 4096, "float32"], [1000, "float32"], 0], {"i": 301, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [500, 2, 1]], ["c_unroll", "sp", [512, 8]]], "c": null, "t": "direct"}], "r": [[0.003322406741935484], 0, 1.7017889022827148, 1538393567.5084052], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 4096], "float32"], ["TENSOR", [4096, 4096], "float32"], ["TENSOR", [4096], "float32"], null], {}, ["dense", [1, 4096, "float32"], [4096, 4096, "float32"], [4096, "float32"], 0], {"i": 371, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [32, 128, 1]], ["c_unroll", "sp", [256, 16]]], "c": null, "t": "direct"}], "r": [[0.012385585777777777], 0, 7.105235576629639, 1538394518.237648], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 25088], "float32"], ["TENSOR", [4096, 25088], "float32"], ["TENSOR", [4096], "float32"], null], {}, ["dense", [1, 25088, "float32"], [4096, 25088, "float32"], [4096, "float32"], 0], {"i": 734, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [64, 64, 1]], ["c_unroll", "sp", [784, 32]]], "c": null, "t": "direct"}], "r": [[0.07633933325], 0, 4.170381546020508, 1538395670.6374521], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15562, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 64]], ["tile_t2", "sp", [16, 32]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.017874490833333333], 0, 1.290653944015503, 1538396945.6981122], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15187, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [4, 128]], ["tile_t2", "sp", [64, 8]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.04691042125], 0, 20.44170594215393, 1538397677.1434402], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15662, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [64, 4]], ["tile_t2", "sp", [8, 64]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0242106638], 0, 15.754741668701172, 1538398988.6440294], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 21262, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [64, 4]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.05234300175], 0, 7.359251260757446, 1538400291.1852999], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 22062, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 4]], ["tile_t2", "sp", [4, 64]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.0289213615], 0, 9.471327066421509, 1538401754.8965714], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 21387, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 128]], ["tile_t2", "sp", [32, 4]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.074520321], 0, 7.8939855098724365, 1538403014.524347], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 13012, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 4]], ["tile_t2", "sp", [32, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.042412352], 0, 13.567544221878052, 1538404298.4877222], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 16437, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 64]], ["tile_t2", "sp", [16, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.1168684115], 0, 4.753565549850464, 1538405808.8227706], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 12254859, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [112, 1, 2]], ["tile_ow", "sp", [8, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011571153333333334], 0, 7.192219972610474, 1538407820.9872587], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 512], "float32"], ["TENSOR", [1000, 512], "float32"], ["TENSOR", [1000], "float32"], null], {}, ["dense", [1, 512, "float32"], [1000, 512, "float32"], [1000, "float32"], 0], {"i": 403, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [200, 5, 1]], ["c_unroll", "sp", [32, 16]]], "c": null, "t": "direct"}], "r": [[0.0004548072443438914], 0, 1.7648944854736328, 1538408504.6271598], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9598, "c": null, "e": [["tile_co", "sp", [32, 2, 8]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0012421755853658538], 0, 1.6651322841644287, 1538409729.06011], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 23174, "c": null, "e": [["tile_co", "sp", [128, 1, 4]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00782902476923077], 0, 5.621083498001099, 1538410429.7944782], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 401604, "c": null, "e": [["tile_co", "sp", [32, 1, 8]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011055100322580644], 0, 2.768322229385376, 1538411424.6175883], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 391835, "c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006033394117647059], 0, 8.492167472839355, 1538412409.9492764], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 472416, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012038617023809523], 0, 2.707627534866333, 1538413684.605693], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1255196, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005840314666666666], 0, 2.1762707233428955, 1538415094.5461981], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 2048], "float32"], ["TENSOR", [1000, 2048], "float32"], ["TENSOR", [1000], "float32"], null], {}, ["dense", [1, 2048, "float32"], [1000, 2048, "float32"], [1000, "float32"], 0], {"i": 303, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [200, 5, 1]], ["c_unroll", "sp", [256, 8]]], "c": null, "t": "direct"}], "r": [[0.0016054948412698414], 0, 5.270214080810547, 1538415378.4916327], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 20153, "c": null, "e": [["tile_co", "sp", [8, 16, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008446886416666667], 0, 1.759467601776123, 1538416829.7712595], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 75607, "c": null, "e": [["tile_co", "sp", [128, 4, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007194806571428571], 0, 7.711011171340942, 1538417666.3943324], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 75469, "c": null, "e": [["tile_co", "sp", [64, 2, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.015076324857142857], 0, 0.9855766296386719, 1538418776.0274906], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15112, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 16]], ["tile_t2", "sp", [64, 8]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.005999681411764705], 0, 1.2697763442993164, 1538419116.7984116], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 22859, "c": null, "e": [["tile_co", "sp", [32, 1, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004524040565217391], 0, 3.3607022762298584, 1538419966.8548815], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 402770, "c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0053092713157894735], 0, 2.1933248043060303, 1538420672.554335], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 580032, "c": null, "e": [["tile_co", "sp", [32, 8, 4]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004549202772727272], 0, 1.7156963348388672, 1538421768.1659548], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 232288, "c": null, "e": [["tile_co", "sp", [8, 16, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012320971777777778], 0, 3.5915160179138184, 1538422680.910124], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15287, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 8]], ["tile_t2", "sp", [16, 16]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.00510876295], 0, 3.906111001968384, 1538423320.169262], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 83868, "c": null, "e": [["tile_co", "sp", [2, 8, 16]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [14, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0043503498260869565], 0, 1.778557538986206, 1538424080.3775072], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1220204, "c": null, "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005311743842105263], 0, 1.872971773147583, 1538425377.5546498], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1916661, "c": null, "e": [["tile_co", "sp", [32, 4, 4]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0050233419000000005], 0, 2.772636651992798, 1538426054.7468307], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1931521, "c": null, "e": [["tile_co", "sp", [4, 16, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 2, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.012233852777777778], 0, 0.956183671951294, 1538426799.6929834], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 14962, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [2, 64]], ["tile_t2", "sp", [32, 4]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.005460027157894737], 0, 12.395811796188354, 1538427909.0246282], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 528793, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [14, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003750817925925926], 0, 0.8128612041473389, 1538429139.6416435], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2660297, "c": null, "e": [["tile_co", "sp", [1, 16, 4]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006164332941176471], 0, 5.743805646896362, 1538430418.56637], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4462400, "c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [14, 1, 4]], ["tile_ow", "sp", [4, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006005954], 0, 2.00496244430542, 1538431724.6584105], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 16437, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 64]], ["tile_t2", "sp", [16, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.007457414428571429], 0, 9.003478765487671, 1538432653.326788], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1020643, "c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [28, 2, 1]], ["tile_ow", "sp", [7, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001966272725490196], 0, 4.835139274597168, 1538434069.6436212], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 2632156, "c": null, "e": [["tile_co", "sp", [2, 8, 4]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [14, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0102994957], 0, 11.073105812072754, 1538435651.6961672], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 1024], "float16"], ["TENSOR", [1000, 1024], "float16"], ["TENSOR", [1000], "float16"], null], {}, ["dense", [1, 1024, "float16"], [1000, 1024, "float16"], [1000, "float16"], 0], {"i": 335, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [1, 250, 4]], ["c_unroll", "sp", [128, 8]]], "t": "direct"}], "v": 0.1, "r": [[0.0005802140900473934], 0, 0.8369655609130859, 1538632379.9443264]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 992, 7, 7], "float16"], ["TENSOR", [128, 992, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 992, 7, 7, "float16"], [128, 992, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 34696, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008302515123966943], 0, 4.633612155914307, 1538633550.8199935], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float16"], ["TENSOR", [128, 960, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 960, 7, 7, "float16"], [128, 960, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 6617, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0008077208548387096], 0, 3.4453001022338867, 1538634393.4249732], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 928, 7, 7], "float16"], ["TENSOR", [128, 928, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 928, 7, 7, "float16"], [128, 928, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 13745, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007962906614173229], 0, 0.7993981838226318, 1538634906.7869782], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 896, 7, 7], "float16"], ["TENSOR", [128, 896, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 896, 7, 7, "float16"], [128, 896, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 12989, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007680810977443609], 0, 4.5978844165802, 1538635737.760356], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 864, 7, 7], "float16"], ["TENSOR", [128, 864, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 864, 7, 7, "float16"], [128, 864, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 26381, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0007389526838235293], 0, 4.7079918384552, 1538636422.95606], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 832, 7, 7], "float16"], ["TENSOR", [128, 832, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 832, 7, 7, "float16"], [128, 832, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 28216, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0007255804604316547], 0, 5.52863883972168, 1538637092.967324], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 800, 7, 7], "float16"], ["TENSOR", [128, 800, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 800, 7, 7, "float16"], [128, 800, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 6545, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0006924141172413794], 0, 0.7971675395965576, 1538637777.038607], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 7, 7], "float16"], ["TENSOR", [128, 768, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 768, 7, 7, "float16"], [128, 768, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 33833, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006686647533333333], 0, 5.175330877304077, 1538638894.0770178], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 736, 7, 7], "float16"], ["TENSOR", [128, 736, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 736, 7, 7, "float16"], [128, 736, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 35668, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006447821898734178], 0, 1.7928307056427002, 1538639375.5196114], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 704, 7, 7], "float16"], ["TENSOR", [128, 704, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 704, 7, 7, "float16"], [128, 704, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 5896, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0006425848238993711], 0, 0.7784490585327148, 1538640220.8947418], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 672, 7, 7], "float16"], ["TENSOR", [128, 672, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 672, 7, 7, "float16"], [128, 672, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 27929, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0006156154909090909], 0, 5.3376617431640625, 1538641047.953468], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 640, 7, 7], "float16"], ["TENSOR", [128, 640, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 640, 7, 7, "float16"], [128, 640, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 5645, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0005931136529411764], 0, 0.7876698970794678, 1538641558.9957812], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 608, 7, 7], "float16"], ["TENSOR", [128, 608, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 608, 7, 7, "float16"], [128, 608, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 13312, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005629246123595506], 0, 5.373538970947266, 1538642196.1347837], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float16"], ["TENSOR", [128, 576, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 576, 7, 7, "float16"], [128, 576, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 13097, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005416425401069518], 0, 0.801605224609375, 1538643052.4662702], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 544, 7, 7], "float16"], ["TENSOR", [128, 544, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 544, 7, 7, "float16"], [128, 544, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 35668, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0005188901761658031], 0, 5.181230545043945, 1538643565.1976287], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float16"], ["TENSOR", [128, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 7, 7, "float16"], [128, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 34085, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004926147142857143], 0, 2.7008564472198486, 1538644762.721485], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float16"], ["TENSOR", [512, 1024, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1024, 14, 14, "float16"], [512, 1024, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 492279, "c": null, "e": [["tile_co", "sp", [16, 4, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004981497666666666], 0, 4.495107889175415, 1538645714.611753], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 992, 14, 14], "float16"], ["TENSOR", [128, 992, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 992, 14, 14, "float16"], [128, 992, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 118070, "c": null, "e": [["tile_co", "sp", [8, 1, 16]], ["tile_oh", "sp", [1, 14, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019195990754716983], 0, 1.6776905059814453, 1538646630.5020468], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 14, 14], "float16"], ["TENSOR", [128, 960, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 960, 14, 14, "float16"], [128, 960, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 325462, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001609290984126984], 0, 6.33248233795166, 1538647483.9055922], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 928, 14, 14], "float16"], ["TENSOR", [128, 928, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 928, 14, 14, "float16"], [128, 928, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 124191, "c": null, "e": [["tile_co", "sp", [4, 2, 16]], ["tile_oh", "sp", [2, 7, 1]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018524190555555555], 0, 0.8581976890563965, 1538648945.8736632], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 896, 14, 14], "float16"], ["TENSOR", [128, 896, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 896, 14, 14, "float16"], [128, 896, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 303794, "c": null, "e": [["tile_co", "sp", [8, 1, 16]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018330000909090908], 0, 2.708740711212158, 1538649806.9367385], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 864, 14, 14], "float16"], ["TENSOR", [128, 864, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 864, 14, 14, "float16"], [128, 864, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 319632, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001472205811594203], 0, 6.499777555465698, 1538651080.400862], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 832, 14, 14], "float16"], ["TENSOR", [128, 832, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 832, 14, 14, "float16"], [128, 832, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 307645, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0015926148714285715], 0, 1.0704965591430664, 1538651492.1664598], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 800, 14, 14], "float16"], ["TENSOR", [128, 800, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 800, 14, 14, "float16"], [128, 800, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 233776, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0016311687419354839], 0, 3.8076772689819336, 1538652486.6498053], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 14, 14], "float16"], ["TENSOR", [128, 768, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 768, 14, 14, "float16"], [128, 768, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 309265, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [14, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018593849814814813], 0, 0.8423154354095459, 1538653456.6907592], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 736, 14, 14], "float16"], ["TENSOR", [128, 736, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 736, 14, 14, "float16"], [128, 736, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 313475, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012674096543209877], 0, 3.742419481277466, 1538654370.4883428], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 704, 14, 14], "float16"], ["TENSOR", [128, 704, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 704, 14, 14, "float16"], [128, 704, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 322546, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012305131204819277], 0, 1.1165590286254883, 1538655487.8192883], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 672, 14, 14], "float16"], ["TENSOR", [128, 672, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 672, 14, 14, "float16"], [128, 672, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 307678, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001196301130952381], 0, 1.0795130729675293, 1538656719.8262758], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 640, 14, 14], "float16"], ["TENSOR", [128, 640, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 640, 14, 14, "float16"], [128, 640, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 322258, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011496765227272726], 0, 5.9161293506622314, 1538658196.51366], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 608, 14, 14], "float16"], ["TENSOR", [128, 608, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 608, 14, 14, "float16"], [128, 608, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 120881, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012372813703703705], 0, 0.8111200332641602, 1538658779.6278653], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float16"], ["TENSOR", [128, 576, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 576, 14, 14, "float16"], [128, 576, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 310594, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00105075546875], 0, 1.9004120826721191, 1538660174.2251446], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 544, 14, 14], "float16"], ["TENSOR", [128, 544, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 544, 14, 14, "float16"], [128, 544, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 305050, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000995429891089109], 0, 4.815294981002808, 1538660803.2803128], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float16"], ["TENSOR", [128, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 14, 14, "float16"], [128, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 324820, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010821143333333333], 0, 1.802292823791504, 1538662227.8175306], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 480, 14, 14], "float16"], ["TENSOR", [128, 480, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 480, 14, 14, "float16"], [128, 480, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 309629, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [1, 7, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010292206565656566], 0, 3.741802215576172, 1538663154.4683447], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 14, 14], "float16"], ["TENSOR", [128, 448, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 448, 14, 14, "float16"], [128, 448, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 319307, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008734144695652174], 0, 0.9172201156616211, 1538663622.3438056], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 416, 14, 14], "float16"], ["TENSOR", [128, 416, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 416, 14, 14, "float16"], [128, 416, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 322154, "c": null, "e": [["tile_co", "sp", [8, 1, 16]], ["tile_oh", "sp", [2, 7, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008927266548672567], 0, 0.8589274883270264, 1538664804.1071112], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float16"], ["TENSOR", [128, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 14, 14, "float16"], [128, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 319631, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007928898031496064], 0, 1.061507225036621, 1538665502.478221], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 352, 14, 14], "float16"], ["TENSOR", [128, 352, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 352, 14, 14, "float16"], [128, 352, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 325462, "c": null, "e": [["tile_co", "sp", [8, 2, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000735102138888889], 0, 1.0425701141357422, 1538666397.1205866], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 14, 14], "float16"], ["TENSOR", [128, 320, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 320, 14, 14, "float16"], [128, 320, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 323079, "c": null, "e": [["tile_co", "sp", [32, 1, 4]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [1, 2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008657809915254237], 0, 2.9200665950775146, 1538667832.419738], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 14, 14], "float16"], ["TENSOR", [128, 288, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 288, 14, 14, "float16"], [128, 288, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 322223, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006014229940828402], 0, 2.0964159965515137, 1538668238.9527838], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [128, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [128, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 130659, "c": null, "e": [["tile_co", "sp", [32, 1, 4]], ["tile_oh", "sp", [2, 7, 1]], ["tile_ow", "sp", [1, 2, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007143878137931035], 0, 12.980849504470825, 1538670521.7877269], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float16"], ["TENSOR", [256, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 28, 28, "float16"], [256, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 649647, "c": null, "e": [["tile_co", "sp", [4, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00505128535], 0, 5.5100438594818115, 1538671498.4962397], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 480, 28, 28], "float16"], ["TENSOR", [128, 480, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 480, 28, 28, "float16"], [128, 480, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 508056, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0028251311666666664], 0, 5.226840019226074, 1538672466.5484972], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 28, 28], "float16"], ["TENSOR", [128, 448, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 448, 28, 28, "float16"], [128, 448, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 543048, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002648561894736842], 0, 0.8817195892333984, 1538673905.2109208], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 416, 28, 28], "float16"], ["TENSOR", [128, 416, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 416, 28, 28, "float16"], [128, 416, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1289544, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00250012945], 0, 1.8410038948059082, 1538674473.4927793], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 28, 28], "float16"], ["TENSOR", [128, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 28, 28, "float16"], [128, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 484728, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0023537088372093022], 0, 3.7291653156280518, 1538675388.1367772], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 352, 28, 28], "float16"], ["TENSOR", [128, 352, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 352, 28, 28, "float16"], [128, 352, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 508057, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0021763842978723405], 0, 1.802276611328125, 1538676877.57384], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 28, 28], "float16"], ["TENSOR", [128, 320, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 320, 28, 28, "float16"], [128, 320, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1277233, "c": null, "e": [["tile_co", "sp", [1, 16, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025299878000000003], 0, 3.7298665046691895, 1538677801.5590632], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 28, 28], "float16"], ["TENSOR", [128, 288, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 288, 28, 28, "float16"], [128, 288, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 473064, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018253179454545455], 0, 2.899953603744507, 1538678589.712446], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float16"], ["TENSOR", [128, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 28, 28, "float16"], [128, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1286664, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [1, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001793099767857143], 0, 4.618740081787109, 1538679869.496493], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 224, 28, 28], "float16"], ["TENSOR", [128, 224, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 224, 28, 28, "float16"], [128, 224, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1266216, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014990044782608695], 0, 3.792128801345825, 1538680902.1791499], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float16"], ["TENSOR", [128, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 28, 28, "float16"], [128, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1017743, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [2, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0014277814788732394], 0, 1.0910990238189697, 1538682348.2482889], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 28, 28], "float16"], ["TENSOR", [128, 160, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 28, 28, "float16"], [128, 160, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1251671, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [1, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012176994939759036], 0, 2.757216453552246, 1538683153.6538024], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [32, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [32, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 752636, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [14, 2, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0019843035384615386], 0, 9.879336595535278, 1538684095.7481098], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [128, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [128, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1231223, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010005965196078432], 0, 5.070068836212158, 1538685131.6204147], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float16"], ["TENSOR", [128, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 56, 56, "float16"], [128, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1476384, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005640069388888888], 0, 5.515628337860107, 1538686486.385828], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 224, 56, 56], "float16"], ["TENSOR", [128, 224, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 224, 56, 56, "float16"], [128, 224, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3387984, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0050315225], 0, 8.301894903182983, 1538687590.6039946], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 56, 56], "float16"], ["TENSOR", [128, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 56, 56, "float16"], [128, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3582384, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0044253551304347825], 0, 0.9474034309387207, 1538689006.4165893], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 56, 56], "float16"], ["TENSOR", [128, 160, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 56, 56, "float16"], [128, 160, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3514344, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [14, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003903540653846154], 0, 2.0458579063415527, 1538690227.782512], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float16"], ["TENSOR", [128, 96, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 56, 56, "float16"], [128, 96, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3544589, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [2, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002811924054054054], 0, 5.276049613952637, 1538691477.5678067], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float16"], ["TENSOR", [192, 2048, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 2048, 8, 8, "float16"], [192, 2048, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 673567, "c": null, "e": [["tile_co", "sp", [2, 12, 8]], ["tile_oh", "sp", [1, 2, 4]], ["tile_ow", "sp", [4, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.00156525590625], 0, 4.502837657928467, 1538693492.23403], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float16"], ["TENSOR", [448, 2048, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 2048, 8, 8, "float16"], [448, 2048, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 914789, "c": null, "e": [["tile_co", "sp", [16, 7, 4]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [1, 1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003960999538461538], 0, 1.0708422660827637, 1538694656.7070105], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float16"], ["TENSOR", [384, 2048, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 2048, 8, 8, "float16"], [384, 2048, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 865899, "c": null, "e": [["tile_co", "sp", [4, 12, 8]], ["tile_oh", "sp", [2, 1, 4]], ["tile_ow", "sp", [4, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0028247660000000003], 0, 0.9769062995910645, 1538696279.3560545], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float16"], ["TENSOR", [320, 2048, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 2048, 8, 8, "float16"], [320, 2048, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 723118, "c": null, "e": [["tile_co", "sp", [8, 5, 8]], ["tile_oh", "sp", [1, 2, 4]], ["tile_ow", "sp", [8, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.002758920108108108], 0, 0.95157790184021, 1538697565.3885162], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float16"], ["TENSOR", [192, 1280, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1280, 8, 8, "float16"], [192, 1280, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 920425, "c": null, "e": [["tile_co", "sp", [8, 6, 4]], ["tile_oh", "sp", [2, 1, 4]], ["tile_ow", "sp", [2, 2, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014201828591549296], 0, 1.1706061363220215, 1538697608.2999916], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float16"], ["TENSOR", [448, 1280, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1280, 8, 8, "float16"], [448, 1280, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 889841, "c": null, "e": [["tile_co", "sp", [16, 7, 4]], ["tile_oh", "sp", [1, 8, 1]], ["tile_ow", "sp", [1, 1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002631982641025641], 0, 4.897541761398315, 1538700416.7615688], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float16"], ["TENSOR", [384, 384, 3, 1], "float16"], [1, 1], [1, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 8, 8, "float16"], [384, 384, 3, 1, "float16"], [1, 1], [1, 0], [1, 1], "NCHW", "float16"], {"i": 504530, "c": null, "e": [["tile_co", "sp", [6, 8, 8]], ["tile_oh", "sp", [4, 2, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013949801666666667], 0, 4.736252307891846, 1538701437.0085294], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float16"], ["TENSOR", [384, 384, 1, 3], "float16"], [1, 1], [0, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 8, 8, "float16"], [384, 384, 1, 3, "float16"], [1, 1], [0, 1], [1, 1], "NCHW", "float16"], {"i": 941280, "c": null, "e": [["tile_co", "sp", [12, 4, 8]], ["tile_oh", "sp", [2, 2, 2]], ["tile_ow", "sp", [4, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0016728482], 0, 3.119023084640503, 1538703076.7012894], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float16"], ["TENSOR", [384, 1280, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1280, 8, 8, "float16"], [384, 1280, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1327869, "c": null, "e": [["tile_co", "sp", [12, 32, 1]], ["tile_oh", "sp", [2, 2, 2]], ["tile_ow", "sp", [1, 1, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.00271998], 0, 4.614370346069336, 1538703876.2708237], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float16"], ["TENSOR", [320, 1280, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1280, 8, 8, "float16"], [320, 1280, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 887928, "c": null, "e": [["tile_co", "sp", [4, 10, 8]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017871249107142858], 0, 6.271266222000122, 1538704984.5812657], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float16"], ["TENSOR", [192, 192, 3, 3], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 17, 17, "float16"], [192, 192, 3, 3, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 392587, "c": null, "e": [["tile_co", "sp", [2, 12, 8]], ["tile_oh", "sp", [1, 8, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012110919176470587], 0, 1.3926517963409424, 1538706847.6379693], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float16"], ["TENSOR", [320, 192, 3, 3], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 17, 17, "float16"], [320, 192, 3, 3, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 400726, "c": null, "e": [["tile_co", "sp", [8, 5, 8]], ["tile_oh", "sp", [8, 1, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00204030774], 0, 1.317357063293457, 1538707926.608499], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float16"], ["TENSOR", [192, 192, 7, 1], "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 17, 17, "float16"], [192, 192, 7, 1, "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {"i": 35082, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006124790411764705], 0, 1.9260151386260986, 1538708652.0809834], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float16"], ["TENSOR", [192, 192, 1, 7], "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 17, 17, "float16"], [192, 192, 1, 7, "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {"i": 81198, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006117155764705883], 0, 1.1298480033874512, 1538709216.6159678], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float16"], ["TENSOR", [192, 160, 1, 7], "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 17, 17, "float16"], [192, 160, 1, 7, "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {"i": 14681, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.005114095149999999], 0, 2.314068555831909, 1538710053.3785784], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float16"], ["TENSOR", [160, 160, 7, 1], "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 17, 17, "float16"], [160, 160, 7, 1, "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {"i": 24988, "c": null, "e": [["tile_co", "sp", [2, 10, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004500629], 0, 2.0046944618225098, 1538710691.0100608], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float16"], ["TENSOR", [192, 160, 7, 1], "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 17, 17, "float16"], [192, 160, 7, 1, "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {"i": 65333, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.005124026], 0, 3.359689712524414, 1538711429.3974311], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float16"], ["TENSOR", [160, 160, 1, 7], "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 17, 17, "float16"], [160, 160, 1, 7, "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {"i": 23840, "c": null, "e": [["tile_co", "sp", [5, 8, 4]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00652091475], 0, 1.0011231899261475, 1538712367.8094842], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float16"], ["TENSOR", [160, 768, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 768, 17, 17, "float16"], [160, 768, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 59081, "c": null, "e": [["tile_co", "sp", [1, 10, 16]], ["tile_oh", "sp", [1, 17, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0037144464444444445], 0, 3.4059293270111084, 1538713272.882497], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float16"], ["TENSOR", [192, 128, 1, 7], "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 17, 17, "float16"], [192, 128, 1, 7, "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {"i": 65826, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004182826916666667], 0, 6.204594612121582, 1538714416.5394592], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float16"], ["TENSOR", [128, 128, 7, 1], "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 17, 17, "float16"], [128, 128, 7, 1, "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {"i": 28212, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.003033843088235294], 0, 1.1723549365997314, 1538714898.046283], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float16"], ["TENSOR", [192, 128, 7, 1], "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 17, 17, "float16"], [192, 128, 7, 1, "float16"], [1, 1], [3, 0], [1, 1], "NCHW", "float16"], {"i": 82723, "c": null, "e": [["tile_co", "sp", [1, 12, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0040721012], 0, 3.0051791667938232, 1538716051.327624], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float16"], ["TENSOR", [128, 128, 1, 7], "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 17, 17, "float16"], [128, 128, 1, 7, "float16"], [1, 1], [0, 3], [1, 1], "NCHW", "float16"], {"i": 35664, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0030309477647058824], 0, 0.8793728351593018, 1538716342.3459656], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float16"], ["TENSOR", [128, 768, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 768, 17, 17, "float16"], [128, 768, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 34805, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002980506411764706], 0, 0.8260703086853027, 1538716956.1994133], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float16"], ["TENSOR", [192, 768, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 768, 17, 17, "float16"], [192, 768, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 30559, "c": null, "e": [["tile_co", "sp", [1, 12, 16]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [1, 17, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0040405340769230766], 0, 0.8526546955108643, 1538717462.0254877], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float16"], ["TENSOR", [96, 96, 3, 3], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 35, 35, "float16"], [96, 96, 3, 3, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 26695, "c": null, "e": [["tile_co", "sp", [1, 12, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00206387604], 0, 2.3889575004577637, 1538718189.4889345], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float16"], ["TENSOR", [384, 288, 3, 3], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 288, 35, 35, "float16"], [384, 288, 3, 3, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 107954, "c": null, "e": [["tile_co", "sp", [6, 8, 8]], ["tile_oh", "sp", [17, 1, 1]], ["tile_ow", "sp", [17, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0193379805], 0, 2.159578800201416, 1538718726.7771864], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float16"], ["TENSOR", [48, 288, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 288, 35, 35, "float16"], [48, 288, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 147860, "c": null, "e": [["tile_co", "sp", [2, 3, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [1, 7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014824983088235295], 0, 1.6684863567352295, 1538719662.345594], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float16"], ["TENSOR", [64, 288, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 288, 35, 35, "float16"], [64, 288, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 107625, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016819827666666666], 0, 0.9691123962402344, 1538720768.2531798], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float16"], ["TENSOR", [48, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 35, 35, "float16"], [48, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 380735, "c": null, "e": [["tile_co", "sp", [2, 3, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016540946769230769], 0, 0.9704475402832031, 1538721879.5558543], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float16"], ["TENSOR", [64, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 35, 35, "float16"], [64, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 241436, "c": null, "e": [["tile_co", "sp", [2, 4, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001587969712121212], 0, 2.090437173843384, 1538722629.4277573], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float16"], ["TENSOR", [32, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 35, 35, "float16"], [32, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 148464, "c": null, "e": [["tile_co", "sp", [4, 1, 8]], ["tile_oh", "sp", [7, 1, 5]], ["tile_ow", "sp", [5, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0012750849874999999], 0, 0.9752547740936279, 1538723578.9492583], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float16"], ["TENSOR", [64, 48, 5, 5], "float16"], [1, 1], [2, 2], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 48, 35, 35, "float16"], [64, 48, 5, 5, "float16"], [1, 1], [2, 2], [1, 1], "NCHW", "float16"], {"i": 107624, "c": null, "e": [["tile_co", "sp", [2, 4, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [7, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0044686349130434785], 0, 3.9678690433502197, 1538726322.2705054], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float16"], ["TENSOR", [48, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 35, 35, "float16"], [48, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 169729, "c": null, "e": [["tile_co", "sp", [3, 2, 8]], ["tile_oh", "sp", [35, 1, 1]], ["tile_ow", "sp", [1, 7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011040781086956522], 0, 4.044262886047363, 1538727775.058987], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float16"], ["TENSOR", [64, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 35, 35, "float16"], [64, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 241743, "c": null, "e": [["tile_co", "sp", [4, 2, 8]], ["tile_oh", "sp", [5, 7, 1]], ["tile_ow", "sp", [1, 7, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012060438674698795], 0, 3.7417120933532715, 1538728691.673738], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float16"], ["TENSOR", [192, 80, 3, 3], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 80, 73, 73, "float16"], [192, 80, 3, 3, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 16686, "c": null, "e": [["tile_co", "sp", [3, 8, 8]], ["tile_oh", "sp", [71, 1, 1]], ["tile_ow", "sp", [71, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0476860315], 0, 4.525280714035034, 1538729478.262146], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float16"], ["TENSOR", [80, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 73, 73, "float16"], [80, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 43507, "c": null, "e": [["tile_co", "sp", [5, 1, 16]], ["tile_oh", "sp", [73, 1, 1]], ["tile_ow", "sp", [1, 73, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003745754962962963], 0, 5.755050182342529, 1538730490.829217], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float16"], ["TENSOR", [64, 32, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 147, 147, "float16"], [64, 32, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 1000965, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [147, 1, 1]], ["tile_ow", "sp", [49, 1, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.020902336], 0, 1.3168449401855469, 1538731537.1561825], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float16"], ["TENSOR", [32, 32, 3, 3], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 149, 149, "float16"], [32, 32, 3, 3, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 315647, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [147, 1, 1]], ["tile_ow", "sp", [7, 7, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.011749658333333335], 0, 2.4849655628204346, 1538732457.0543551], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float16"], ["TENSOR", [32, 3, 3, 3], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 299, 299, "float16"], [32, 3, 3, 3, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 20996, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [149, 1, 1]], ["tile_ow", "sp", [149, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0033477972999999997], 0, 1.0487494468688965, 1538733171.5421941], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1280, 1, 1], "float16"], ["TENSOR", [1000, 1280, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1280, 1, 1, "float16"], [1000, 1280, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 12128, "c": null, "e": [["tile_co", "sp", [250, 1, 4]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.00082735309375], 0, 5.991708755493164, 1538733607.5055413], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 320, 7, 7], "float16"], ["TENSOR", [1280, 320, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 320, 7, 7, "float16"], [1280, 320, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 100117, "c": null, "e": [["tile_co", "sp", [8, 10, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0018399709454545456], 0, 3.401978015899658, 1538734459.3327923], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float16"], ["TENSOR", [320, 960, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 960, 7, 7, "float16"], [320, 960, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 12396, "c": null, "e": [["tile_co", "sp", [4, 10, 8]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0016991899830508474], 0, 2.502788782119751, 1538735166.0346525], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 960, 7, 7], "float16"], ["TENSOR", [160, 960, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 960, 7, 7, "float16"], [160, 960, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 23654, "c": null, "e": [["tile_co", "sp", [1, 40, 4]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001092880652173913], 0, 1.727158784866333, 1538735776.811657], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 960, 7, 7], "float16"], ["TENSOR", [960, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 960, 7, 7, "float16"], [960, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 44889, "c": null, "e": [["tile_c", "sp", [80, 6, 2]], ["tile_y", "sp", [1, 7, 1]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0002945057609329446], 0, 3.6799509525299072, 1538736403.816761], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 160, 7, 7], "float16"], ["TENSOR", [960, 160, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 160, 7, 7, "float16"], [960, 160, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 246271, "c": null, "e": [["tile_co", "sp", [120, 2, 4]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008717331083333334], 0, 5.329559564590454, 1538737297.750011], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 7, 7], "float16"], ["TENSOR", [160, 576, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 576, 7, 7, "float16"], [160, 576, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 61842, "c": null, "e": [["tile_co", "sp", [4, 5, 8]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007742741145038168], 0, 1.712576150894165, 1538737988.3486717], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float16"], ["TENSOR", [576, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float16"], [576, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 11596, "c": null, "e": [["tile_c", "sp", [96, 6, 1]], ["tile_y", "sp", [7, 1, 1]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0003475621736111111], 0, 1.2191541194915771, 1538738540.7881973], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 576, 14, 14], "float16"], ["TENSOR", [96, 576, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 576, 14, 14, "float16"], [96, 576, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 559360, "c": null, "e": [["tile_co", "sp", [1, 12, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009930185294117647], 0, 4.569958209991455, 1538739932.418995], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 576, 14, 14], "float16"], ["TENSOR", [576, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 576, 14, 14, "float16"], [576, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 105172, "c": null, "e": [["tile_c", "sp", [96, 6, 1]], ["tile_y", "sp", [1, 7, 2]], ["tile_x", "sp", [2, 1, 7]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.00046069386238532106], 0, 3.2909491062164307, 1538740311.9527662], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 14, 14], "float16"], ["TENSOR", [576, 96, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 14, 14, "float16"], [576, 96, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1470728, "c": null, "e": [["tile_co", "sp", [36, 4, 4]], ["tile_oh", "sp", [2, 1, 7]], ["tile_ow", "sp", [14, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0011011082197802198], 0, 1.1000511646270752, 1538741092.6279938], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float16"], ["TENSOR", [96, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 14, 14, "float16"], [96, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 553559, "c": null, "e": [["tile_co", "sp", [12, 1, 8]], ["tile_oh", "sp", [2, 7, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00081641196], 0, 5.635289192199707, 1538742279.2084978], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 14, 14], "float16"], ["TENSOR", [64, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 14, 14, "float16"], [64, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 237265, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00048669454028436014], 0, 4.8611414432525635, 1538743292.0175557], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 384, 14, 14], "float16"], ["TENSOR", [384, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 384, 14, 14, "float16"], [384, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 112664, "c": null, "e": [["tile_c", "sp", [32, 6, 2]], ["tile_y", "sp", [1, 1, 14]], ["tile_x", "sp", [1, 2, 7]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.00033090298407643313], 0, 5.77230429649353, 1538744513.7933187], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 14, 14], "float16"], ["TENSOR", [384, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 14, 14, "float16"], [384, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 974453, "c": null, "e": [["tile_co", "sp", [4, 6, 16]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0004884831512195121], 0, 0.855797529220581, 1538745469.270062], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 14, 14], "float16"], ["TENSOR", [64, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 14, 14, "float16"], [64, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 243736, "c": null, "e": [["tile_co", "sp", [1, 4, 16]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0003364673222591362], 0, 3.832648515701294, 1538746433.3093793], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float16"], ["TENSOR", [192, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float16"], [192, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 120208, "c": null, "e": [["tile_c", "sp", [32, 6, 1]], ["tile_y", "sp", [14, 1, 1]], ["tile_x", "sp", [2, 1, 7]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "t": "direct"}], "r": [[0.0004406171754385965], 0, 2.9929206371307373, 1538746761.5252395], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 192, 28, 28], "float16"], ["TENSOR", [32, 192, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 192, 28, 28, "float16"], [32, 192, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 752636, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [14, 2, 1]], ["tile_ow", "sp", [1, 7, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006075025212121212], 0, 3.036311626434326, 1538747566.5884657], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 192, 28, 28], "float16"], ["TENSOR", [192, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 192, 28, 28, "float16"], [192, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 47077, "c": null, "e": [["tile_c", "sp", [8, 6, 4]], ["tile_y", "sp", [7, 4, 1]], ["tile_x", "sp", [2, 2, 7]], ["ann_spatial", "an", ["none", "none", "unroll"]]], "t": "direct"}], "r": [[0.0006215923086419754], 0, 1.015758752822876, 1538748241.2513597], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 28, 28], "float16"], ["TENSOR", [192, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 28, 28, "float16"], [192, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 2238652, "c": null, "e": [["tile_co", "sp", [6, 4, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [2, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0005885712369942197], 0, 1.906054973602295, 1538749109.210735], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 144, 28, 28], "float16"], ["TENSOR", [32, 144, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 144, 28, 28, "float16"], [32, 144, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 738397, "c": null, "e": [["tile_co", "sp", [2, 2, 8]], ["tile_oh", "sp", [7, 2, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000487724985645933], 0, 1.93284010887146, 1538750710.8995194], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float16"], ["TENSOR", [144, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float16"], [144, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 162094, "c": null, "e": [["tile_c", "sp", [24, 6, 1]], ["tile_y", "sp", [14, 2, 1]], ["tile_x", "sp", [7, 1, 4]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001020423581632653], 0, 1.0008134841918945, 1538751379.7839434], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 144, 56, 56], "float16"], ["TENSOR", [24, 144, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 144, 56, 56, "float16"], [24, 144, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 2901679, "c": null, "e": [["tile_co", "sp", [3, 2, 4]], ["tile_oh", "sp", [14, 4, 1]], ["tile_ow", "sp", [14, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0017636626949152543], 0, 1.0554084777832031, 1538752368.2719805], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 144, 56, 56], "float16"], ["TENSOR", [144, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 144, 56, 56, "float16"], [144, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 629284, "c": null, "e": [["tile_c", "sp", [24, 6, 1]], ["tile_y", "sp", [14, 4, 1]], ["tile_x", "sp", [1, 7, 8]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0012480232222222224], 0, 1.1614108085632324, 1538753785.7582777], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 24, 56, 56], "float16"], ["TENSOR", [144, 24, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 24, 56, 56, "float16"], [144, 24, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 8858845, "c": null, "e": [["tile_co", "sp", [9, 2, 8]], ["tile_oh", "sp", [28, 2, 1]], ["tile_ow", "sp", [4, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001882623625], 0, 1.0276691913604736, 1538754397.3081462], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 56, 56], "float16"], ["TENSOR", [24, 96, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 56, 56, "float16"], [24, 96, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1721703, "c": null, "e": [["tile_co", "sp", [6, 4, 1]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 7, 8]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0015297621060606059], 0, 0.8233692646026611, 1538755301.6918397], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 96, 112, 112], "float16"], ["TENSOR", [96, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 96, 112, 112, "float16"], [96, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 722047, "c": null, "e": [["tile_c", "sp", [16, 6, 1]], ["tile_y", "sp", [28, 2, 1]], ["tile_x", "sp", [7, 1, 8]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.002447977804878049], 0, 1.1440792083740234, 1538756012.463246], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 112, 112], "float16"], ["TENSOR", [96, 16, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 16, 112, 112, "float16"], [96, 16, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 5806124, "c": null, "e": [["tile_co", "sp", [3, 4, 8]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [1, 28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027187510810810812], 0, 3.9679555892944336, 1538757073.2135222], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float16"], ["TENSOR", [16, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 112, 112, "float16"], [16, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3164038, "c": null, "e": [["tile_co", "sp", [1, 2, 8]], ["tile_oh", "sp", [7, 4, 4]], ["tile_ow", "sp", [4, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013913827638888888], 0, 0.8730649948120117, 1538758577.6888833], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float16"], ["TENSOR", [32, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 112, 112, "float16"], [32, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 3577220, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [28, 1, 4]], ["tile_ow", "sp", [8, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0018517479272727273], 0, 0.8693032264709473, 1538760003.7607224], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 1024], "float16"], ["TENSOR", [1000, 1024], "float16"], null, null], {}, ["dense", [1, 1024, "float16"], [1000, 1024, "float16"], 0, 0], {"i": 604, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [125, 8, 1]], ["c_unroll", "sp", [16, 64]]], "t": "direct"}], "v": 0.1, "r": [[0.0004800026208530805], 0, 5.0448877811431885, 1538760490.245731]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float16"], ["TENSOR", [1024, 1024, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1024, 7, 7, "float16"], [1024, 1024, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 26565, "c": null, "e": [["tile_co", "sp", [16, 8, 8]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0040215541923076924], 0, 4.570726633071899, 1538761787.787628], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float16"], ["TENSOR", [1024, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float16"], [1024, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 7526, "c": null, "e": [["tile_c", "sp", [256, 4, 1]], ["tile_y", "sp", [7, 1, 1]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0003215932307692308], 0, 2.378788948059082, 1538762257.7709763], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float16"], ["TENSOR", [1024, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 7, 7, "float16"], [1024, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 49467, "c": null, "e": [["tile_co", "sp", [16, 8, 8]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0021518661489361704], 0, 2.1718955039978027, 1538762844.4388447], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float16"], ["TENSOR", [512, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float16"], [512, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 9857, "c": null, "e": [["tile_c", "sp", [64, 4, 2]], ["tile_y", "sp", [1, 1, 7]], ["tile_x", "sp", [1, 1, 7]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.00033503882450331123], 0, 1.321056604385376, 1538763298.0003998], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float16"], ["TENSOR", [512, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 14, 14, "float16"], [512, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 189181, "c": null, "e": [["tile_co", "sp", [8, 4, 16]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0030005106470588237], 0, 5.772088050842285, 1538764301.7045808], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float16"], ["TENSOR", [512, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float16"], [512, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 70082, "c": null, "e": [["tile_c", "sp", [64, 4, 2]], ["tile_y", "sp", [1, 7, 2]], ["tile_x", "sp", [2, 1, 7]], ["ann_spatial", "an", ["unroll", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0004434800398230089], 0, 4.477323532104492, 1538765294.8901157], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [512, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [512, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 492280, "c": null, "e": [["tile_co", "sp", [8, 8, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014869154117647058], 0, 1.7835874557495117, 1538766275.6836104], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float16"], ["TENSOR", [256, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float16"], [256, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 35372, "c": null, "e": [["tile_c", "sp", [64, 4, 1]], ["tile_y", "sp", [1, 14, 1]], ["tile_x", "sp", [2, 1, 7]], ["ann_spatial", "an", ["unroll", "none", "unroll"]]], "t": "direct"}], "r": [[0.0005605422458100559], 0, 3.125685453414917, 1538766849.3921683], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float16"], ["TENSOR", [256, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 28, 28, "float16"], [256, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1200087, "c": null, "e": [["tile_co", "sp", [4, 8, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [1, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0028653855714285714], 0, 4.644031524658203, 1538768024.2518125], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float16"], ["TENSOR", [256, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float16"], [256, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 112616, "c": null, "e": [["tile_c", "sp", [8, 4, 8]], ["tile_y", "sp", [28, 1, 1]], ["tile_x", "sp", [2, 2, 7]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.000808678], 0, 1.0405917167663574, 1538768708.4940844], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [256, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [256, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1257596, "c": null, "e": [["tile_co", "sp", [8, 4, 8]], ["tile_oh", "sp", [7, 1, 4]], ["tile_ow", "sp", [2, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0016137771746031746], 0, 1.1051568984985352, 1538769875.6562896], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float16"], ["TENSOR", [128, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float16"], [128, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 65465, "c": null, "e": [["tile_c", "sp", [8, 4, 4]], ["tile_y", "sp", [28, 1, 1]], ["tile_x", "sp", [1, 7, 4]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010351964020618558], 0, 4.849895715713501, 1538770614.7380552], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float16"], ["TENSOR", [128, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 56, 56, "float16"], [128, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1443983, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00312973334375], 0, 0.9203085899353027, 1538771910.0466425], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float16"], ["TENSOR", [128, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float16"], [128, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 251715, "c": null, "e": [["tile_c", "sp", [16, 8, 1]], ["tile_y", "sp", [14, 4, 1]], ["tile_x", "sp", [1, 7, 8]], ["ann_spatial", "an", ["none", "vec", "unroll"]]], "t": "direct"}], "r": [[0.0011625055632183909], 0, 1.1934692859649658, 1538772666.4781966], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [128, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [128, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 2696208, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [14, 1, 4]], ["tile_ow", "sp", [2, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0020184200399999998], 0, 2.779381275177002, 1538773937.606552], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float16"], ["TENSOR", [64, 1, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float16"], [64, 1, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "float16"], {"i": 498795, "c": null, "e": [["tile_c", "sp", [8, 8, 1]], ["tile_y", "sp", [4, 1, 14]], ["tile_x", "sp", [1, 7, 8]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0019357952115384614], 0, 3.128297805786133, 1538774973.4498615], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float16"], ["TENSOR", [64, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 112, 112, "float16"], [64, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 6132945, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [28, 1, 4]], ["tile_ow", "sp", [4, 28, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0027895911111111113], 0, 1.6653130054473877, 1538775788.077858], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float16"], ["TENSOR", [32, 1, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "float16"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float16"], [32, 1, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "float16"], {"i": 418680, "c": null, "e": [["tile_c", "sp", [4, 8, 1]], ["tile_y", "sp", [28, 4, 1]], ["tile_x", "sp", [1, 7, 16]], ["ann_spatial", "an", ["unroll", "none", "unroll"]]], "t": "direct"}], "r": [[0.0010919210322580645], 0, 1.4519503116607666, 1538777121.3120685], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float16"], ["TENSOR", [32, 3, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 224, 224, "float16"], [32, 3, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {"i": 4695743, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [56, 2, 1]], ["tile_ow", "sp", [14, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014313319000000002], 0, 1.5722167491912842, 1538777904.6131725], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float16"], ["TENSOR", [64, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 13, 13, "float16"], [64, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 10944, "c": null, "e": [["tile_co", "sp", [1, 4, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [1, 13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007159475428571429], 0, 4.324434995651245, 1538778722.4897132], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 13, 13], "float16"], ["TENSOR", [48, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 13, 13, "float16"], [48, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 41177, "c": null, "e": [["tile_co", "sp", [16, 3, 1]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [1, 1, 13]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "unroll"]]], "t": "direct"}], "r": [[0.0009162163008849558], 0, 6.054295539855957, 1538779373.250644], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float16"], ["TENSOR", [192, 48, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 48, 13, 13, "float16"], [192, 48, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 35584, "c": null, "e": [["tile_co", "sp", [6, 4, 8]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012065775294117648], 0, 2.5178112983703613, 1538780133.707966], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 13, 13], "float16"], ["TENSOR", [192, 48, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 48, 13, 13, "float16"], [192, 48, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 34085, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00033492270903010034], 0, 4.900341510772705, 1538780634.2407472], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 13, 13], "float16"], ["TENSOR", [48, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 13, 13, "float16"], [48, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 33651, "c": null, "e": [["tile_co", "sp", [1, 6, 8]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0004829146220095694], 0, 0.9158263206481934, 1538781297.9890645], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 27, 27], "float16"], ["TENSOR", [32, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 27, 27, "float16"], [32, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 223707, "c": null, "e": [["tile_co", "sp", [4, 1, 8]], ["tile_oh", "sp", [3, 9, 1]], ["tile_ow", "sp", [3, 3, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000490522570754717], 0, 0.9912846088409424, 1538781757.6977046], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 55, 55], "float16"], ["TENSOR", [16, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 55, 55, "float16"], [16, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 135568, "c": null, "e": [["tile_co", "sp", [1, 2, 8]], ["tile_oh", "sp", [11, 5, 1]], ["tile_ow", "sp", [1, 11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0006113099390243902], 0, 1.9266202449798584, 1538782611.679042], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float16"], ["TENSOR", [64, 3, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 224, 224, "float16"], [64, 3, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {"i": 6243601, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [56, 1, 2]], ["tile_ow", "sp", [8, 14, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002612249225], 0, 6.574219465255737, 1538784063.3223617], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float16"], ["TENSOR", [1000, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 13, 13, "float16"], [1000, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 95633, "c": null, "e": [["tile_co", "sp", [5, 50, 4]], ["tile_oh", "sp", [1, 1, 13]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008705783666666666], 0, 3.4292759895324707, 1538785071.9670231], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float16"], ["TENSOR", [256, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 13, 13, "float16"], [256, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 6919, "c": null, "e": [["tile_co", "sp", [1, 16, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [13, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "none", "vec"]]], "t": "direct"}], "r": [[0.0004510531592920354], 0, 5.404657602310181, 1538786886.7107954], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 13, 13], "float16"], ["TENSOR", [64, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 13, 13, "float16"], [64, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 26818, "c": null, "e": [["tile_co", "sp", [4, 1, 16]], ["tile_oh", "sp", [13, 1, 1]], ["tile_ow", "sp", [1, 13, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0009014861261261261], 0, 1.7153654098510742, 1538787824.1501515], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float16"], ["TENSOR", [256, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 27, 27, "float16"], [256, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 373707, "c": null, "e": [["tile_co", "sp", [4, 8, 8]], ["tile_oh", "sp", [9, 1, 3]], ["tile_ow", "sp", [27, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0012576545124999999], 0, 1.905381441116333, 1538789637.6545336], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float16"], ["TENSOR", [64, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 27, 27, "float16"], [64, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 292901, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001597848953125], 0, 1.8963291645050049, 1538790556.5519545], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 384, 27, 27], "float16"], ["TENSOR", [48, 384, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 384, 27, 27, "float16"], [48, 384, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 215213, "c": null, "e": [["tile_co", "sp", [12, 1, 4]], ["tile_oh", "sp", [3, 9, 1]], ["tile_ow", "sp", [1, 3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0013533014459459459], 0, 4.6253931522369385, 1538791741.6260657], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float16"], ["TENSOR", [192, 48, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 48, 27, 27, "float16"], [192, 48, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 928239, "c": null, "e": [["tile_co", "sp", [4, 12, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [3, 3, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.004417422304347826], 0, 1.3400623798370361, 1538792372.5325084], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 48, 27, 27], "float16"], ["TENSOR", [192, 48, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 48, 27, 27, "float16"], [192, 48, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 708185, "c": null, "e": [["tile_co", "sp", [3, 4, 16]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [1, 27, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.000951520820754717], 0, 1.8304121494293213, 1538793070.1763353], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float16"], ["TENSOR", [48, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 27, 27, "float16"], [48, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 195327, "c": null, "e": [["tile_co", "sp", [2, 6, 4]], ["tile_oh", "sp", [27, 1, 1]], ["tile_ow", "sp", [9, 1, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0012988313544303797], 0, 1.0114781856536865, 1538793985.8863382], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float16"], ["TENSOR", [128, 32, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 27, 27, "float16"], [128, 32, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 171420, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [9, 3, 1]], ["tile_ow", "sp", [1, 9, 3]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0014980752647058822], 0, 1.2826087474822998, 1538795305.753612], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 27, 27], "float16"], ["TENSOR", [128, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 27, 27, "float16"], [128, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 378167, "c": null, "e": [["tile_co", "sp", [4, 4, 8]], ["tile_oh", "sp", [9, 1, 3]], ["tile_ow", "sp", [27, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000549569737704918], 0, 1.0280110836029053, 1538796224.8092782], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 27, 27], "float16"], ["TENSOR", [32, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 27, 27, "float16"], [32, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 100418, "c": null, "e": [["tile_co", "sp", [1, 4, 8]], ["tile_oh", "sp", [9, 3, 1]], ["tile_ow", "sp", [1, 3, 9]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0008789584086956522], 0, 4.236295938491821, 1538797326.8957155], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float16"], ["TENSOR", [128, 32, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 55, 55, "float16"], [128, 32, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 123792, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.005310756526315789], 0, 1.1623306274414062, 1538798048.1751225], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 32, 55, 55], "float16"], ["TENSOR", [128, 32, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 32, 55, 55, "float16"], [128, 32, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 254692, "c": null, "e": [["tile_co", "sp", [2, 4, 16]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [1, 55, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.001927231576923077], 0, 1.7862107753753662, 1538798922.9389913], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float16"], ["TENSOR", [32, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 55, 55, "float16"], [32, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 179569, "c": null, "e": [["tile_co", "sp", [1, 2, 16]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [1, 11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0016680017833333334], 0, 5.452693700790405, 1538800032.0433037], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 55, 55], "float16"], ["TENSOR", [16, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 55, 55, "float16"], [16, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 56473, "c": null, "e": [["tile_co", "sp", [1, 2, 8]], ["tile_oh", "sp", [5, 11, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0010725631063829787], 0, 3.7978835105895996, 1538800324.682729], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float16"], ["TENSOR", [64, 16, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 16, 55, 55, "float16"], [64, 16, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 241436, "c": null, "e": [["tile_co", "sp", [2, 4, 8]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [11, 1, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0018171205892857142], 0, 1.3430955410003662, 1538801270.3405948], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 16, 55, 55], "float16"], ["TENSOR", [64, 16, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 16, 55, 55, "float16"], [64, 16, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 242809, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [11, 1, 5]], ["tile_ow", "sp", [55, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007298120413793103], 0, 1.1111550331115723, 1538802053.9266298], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 55, 55], "float16"], ["TENSOR", [16, 96, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 55, 55, "float16"], [16, 96, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 127048, "c": null, "e": [["tile_co", "sp", [1, 2, 8]], ["tile_oh", "sp", [55, 1, 1]], ["tile_ow", "sp", [1, 11, 5]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000804700848], 0, 4.661175012588501, 1538802778.8855352], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float16"], ["TENSOR", [96, 3, 7, 7], "float16"], [2, 2], [3, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 224, 224, "float16"], [96, 3, 7, 7, "float16"], [2, 2], [3, 3], [1, 1], "NCHW", "float16"], {"i": 6049997, "c": null, "e": [["tile_co", "sp", [3, 4, 8]], ["tile_oh", "sp", [56, 2, 1]], ["tile_ow", "sp", [14, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.008513244], 0, 12.439008474349976, 1538804403.9426947], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 4096], "float16"], ["TENSOR", [1000, 4096], "float16"], ["TENSOR", [1000], "float16"], null], {}, ["dense", [1, 4096, "float16"], [1000, 4096, "float16"], [1000, "float16"], 0], {"i": 605, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [100, 10, 1]], ["c_unroll", "sp", [64, 64]]], "t": "direct"}], "v": 0.1, "r": [[0.0016756960166666665], 0, 3.6759862899780273, 1538805322.1950846]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 4096], "float16"], ["TENSOR", [4096, 4096], "float16"], ["TENSOR", [4096], "float16"], null], {}, ["dense", [1, 4096, "float16"], [4096, 4096, "float16"], [4096, "float16"], 0], {"i": 365, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [2048, 2, 1]], ["c_unroll", "sp", [256, 16]]], "t": "direct"}], "v": 0.1, "r": [[0.0064386485625], 0, 2.906545400619507, 1538806320.712949]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 25088], "float16"], ["TENSOR", [4096, 25088], "float16"], ["TENSOR", [4096], "float16"], null], {}, ["dense", [1, 25088, "float16"], [4096, 25088, "float16"], [4096, "float16"], 0], {"i": 911, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [2048, 2, 1]], ["c_unroll", "sp", [448, 56]]], "t": "direct"}], "v": 0.1, "r": [[0.046200917], 0, 3.784418821334839, 1538807329.712466]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float16"], ["TENSOR", [64, 3, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 224, 224, "float16"], [64, 3, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 12273933, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [224, 1, 1]], ["tile_ow", "sp", [2, 28, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.007613095785714286], 0, 1.4565367698669434, 1538822787.719077], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 512], "float16"], ["TENSOR", [1000, 512], "float16"], ["TENSOR", [1000], "float16"], null], {}, ["dense", [1, 512, "float16"], [1000, 512, "float16"], [1000, "float16"], 0], {"i": 403, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [200, 5, 1]], ["c_unroll", "sp", [32, 16]]], "t": "direct"}], "v": 0.1, "r": [[0.0002798621299734748], 0, 6.161244869232178, 1538823142.338813]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [512, 256, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [512, 256, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 22687, "c": null, "e": [["tile_co", "sp", [64, 1, 8]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 1, 7]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007593721278195489], 0, 4.604989767074585, 1538824093.0268483], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [512, 256, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [512, 256, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {"i": 22799, "c": null, "e": [["tile_co", "sp", [16, 4, 8]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0041767452], 0, 1.1835463047027588, 1538824713.4414806], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [256, 128, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [256, 128, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 406247, "c": null, "e": [["tile_co", "sp", [4, 4, 16]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0007015416013986015], 0, 2.981767177581787, 1538825366.3654752], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [256, 128, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [256, 128, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {"i": 172962, "c": null, "e": [["tile_co", "sp", [4, 8, 8]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0033754749], 0, 7.035157203674316, 1538826524.7637076], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [128, 64, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [128, 64, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 554069, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.000781552140625], 0, 4.545984268188477, 1538827656.1871963], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [128, 64, 3, 3], "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [128, 64, 3, 3, "float16"], [2, 2], [1, 1], [1, 1], "NCHW", "float16"], {"i": 543048, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.00282742925], 0, 4.216508388519287, 1538828567.5136054], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_dense", [["TENSOR", [1, 2048], "float16"], ["TENSOR", [1000, 2048], "float16"], ["TENSOR", [1000], "float16"], null], {}, ["dense", [1, 2048, "float16"], [1000, 2048, "float16"], [1000, "float16"], 0], {"i": 505, "c": null, "e": [["tile_y", "sp", [1, 1, 1]], ["tile_x", "sp", [100, 10, 1]], ["c_unroll", "sp", [64, 32]]], "t": "direct"}], "v": 0.1, "r": [[0.0008482203644067796], 0, 4.211421012878418, 1538829333.01879]}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float16"], ["TENSOR", [512, 2048, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 2048, 7, 7, "float16"], [512, 2048, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 41224, "c": null, "e": [["tile_co", "sp", [16, 4, 8]], ["tile_oh", "sp", [1, 1, 7]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004025511440000001], 0, 3.007218837738037, 1538830538.451566], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float16"], ["TENSOR", [2048, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 7, 7, "float16"], [2048, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 56204, "c": null, "e": [["tile_co", "sp", [32, 4, 16]], ["tile_oh", "sp", [7, 1, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.004389022130434783], 0, 1.7869715690612793, 1538831207.4736938], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float16"], ["TENSOR", [2048, 1024, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1024, 14, 14, "float16"], [2048, 1024, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 57918, "c": null, "e": [["tile_co", "sp", [128, 1, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [1, 7, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.008714922], 0, 5.736236572265625, 1538831797.1927795], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float16"], ["TENSOR", [512, 1024, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1024, 14, 14, "float16"], [512, 1024, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 19890, "c": null, "e": [["tile_co", "sp", [16, 2, 16]], ["tile_oh", "sp", [1, 7, 1]], ["tile_ow", "sp", [7, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0026814866842105266], 0, 2.5446197986602783, 1538833676.7023625], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float16"], ["TENSOR", [256, 1024, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 1024, 14, 14, "float16"], [256, 1024, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 399535, "c": null, "e": [["tile_co", "sp", [16, 2, 8]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [1, 7, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0029870688529411763], 0, 1.9580085277557373, 1538834820.7910173], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [1024, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [1024, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 433924, "c": null, "e": [["tile_co", "sp", [16, 4, 16]], ["tile_oh", "sp", [7, 1, 2]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0027764130810810812], 0, 2.772752523422241, 1538836099.165978], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float16"], ["TENSOR", [1024, 512, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 28, 28, "float16"], [1024, 512, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 590541, "c": null, "e": [["tile_co", "sp", [32, 2, 16]], ["tile_oh", "sp", [7, 2, 1]], ["tile_ow", "sp", [7, 1, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0064862606875], 0, 5.323763370513916, 1538837300.9190485], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float16"], ["TENSOR", [256, 512, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 28, 28, "float16"], [256, 512, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 153528, "c": null, "e": [["tile_co", "sp", [2, 8, 16]], ["tile_oh", "sp", [14, 1, 1]], ["tile_ow", "sp", [7, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0025278042250000002], 0, 5.475460529327393, 1538838533.2757318], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float16"], ["TENSOR", [128, 512, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 28, 28, "float16"], [128, 512, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1219560, "c": null, "e": [["tile_co", "sp", [2, 8, 8]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.003034867411764706], 0, 1.0496666431427002, 1538839990.3638096], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [512, 128, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [512, 128, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1916009, "c": null, "e": [["tile_co", "sp", [16, 4, 8]], ["tile_oh", "sp", [14, 1, 2]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.002731956027027027], 0, 1.7346396446228027, 1538840872.9403327], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float16"], ["TENSOR", [512, 256, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 56, 56, "float16"], [512, 256, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1880048, "c": null, "e": [["tile_co", "sp", [2, 16, 16]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [1, 14, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0065488609375], 0, 3.5417277812957764, 1538841831.5018895], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float16"], ["TENSOR", [128, 256, 1, 1], "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 56, 56, "float16"], [128, 256, 1, 1, "float16"], [2, 2], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1004429, "c": null, "e": [["tile_co", "sp", [1, 8, 16]], ["tile_oh", "sp", [28, 1, 1]], ["tile_ow", "sp", [7, 4, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "none", "vec"]]], "t": "direct"}], "r": [[0.0027073713157894736], 0, 1.6657164096832275, 1538844252.107011], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float16"], ["TENSOR", [64, 256, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 56, 56, "float16"], [64, 256, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1145781, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [56, 1, 1]], ["tile_ow", "sp", [14, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035728516071428574], 0, 4.678251266479492, 1538845049.1921308], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [256, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [256, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 4313320, "c": null, "e": [["tile_co", "sp", [16, 2, 8]], ["tile_oh", "sp", [28, 2, 1]], ["tile_ow", "sp", [7, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.0035525226206896553], 0, 2.87395977973938, 1538846593.8810399], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [64, 64, 1, 1], "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [64, 64, 1, 1, "float16"], [1, 1], [0, 0], [1, 1], "NCHW", "float16"], {"i": 1097928, "c": null, "e": [["tile_co", "sp", [2, 4, 8]], ["tile_oh", "sp", [28, 2, 1]], ["tile_ow", "sp", [1, 14, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "t": "direct"}], "r": [[0.001203579380952381], 0, 1.0861639976501465, 1538848076.6331637], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float16"], ["TENSOR", [64, 3, 7, 7], "float16"], [2, 2], [3, 3], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 3, 224, 224, "float16"], [64, 3, 7, 7, "float16"], [2, 2], [3, 3], [1, 1], "NCHW", "float16"], {"i": 6317661, "c": null, "e": [["tile_co", "sp", [1, 8, 8]], ["tile_oh", "sp", [112, 1, 1]], ["tile_ow", "sp", [14, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "t": "direct"}], "r": [[0.006038337117647058], 0, 6.7816314697265625, 1538850043.8729377], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 7, 7], "float16"], ["TENSOR", [32, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 7, 7, "float16"], [32, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 11662, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 4]], ["tile_t2", "sp", [2, 16]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0002704029732620321], 0, 4.193325042724609, 1539256989.0511532], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 14, 14], "float16"], ["TENSOR", [32, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 14, 14, "float16"], [32, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 14261, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 4]], ["tile_t2", "sp", [1, 32]], ["c_unroll", "sp", [16, 8]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.000789238546875], 0, 5.3365867137908936, 1539257583.3258145], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float16"], ["TENSOR", [32, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 56, 56, "float16"], [32, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 18661, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [32, 4]], ["tile_t2", "sp", [4, 8]], ["c_unroll", "sp", [16, 8]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.006905089200000001], 0, 3.8691892623901367, 1539259405.5481424], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float16"], ["TENSOR", [384, 448, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 448, 8, 8, "float16"], [384, 448, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 74936, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 56]], ["tile_t2", "sp", [6, 64]], ["c_unroll", "sp", [112, 4]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.003061332363636364], 0, 9.108967781066895, 1539260723.5946913], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float16"], ["TENSOR", [96, 96, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 96, 35, 35, "float16"], [96, 96, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 50292, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [6, 16]], ["tile_t2", "sp", [1, 96]], ["c_unroll", "sp", [48, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0038921106153846155], 0, 1.1923186779022217, 1539261950.717845], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float16"], ["TENSOR", [96, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 35, 35, "float16"], [96, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 32461, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 8]], ["tile_t2", "sp", [12, 8]], ["c_unroll", "sp", [8, 8]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.00313774934375], 0, 3.9298360347747803, 1539262576.3087442], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 13, 13], "float16"], ["TENSOR", [256, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 13, 13, "float16"], [256, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 13067, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [4, 16]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0012899777820512821], 0, 3.5680699348449707, 1539263425.348902], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 27, 27], "float16"], ["TENSOR", [256, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 27, 27, "float16"], [256, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 13042, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [8, 8]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.003569257482758621], 0, 4.436291456222534, 1539264025.297396], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float16"], ["TENSOR", [512, 512, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 14, 14, "float16"], [512, 512, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 15917, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [32, 16]], ["tile_t2", "sp", [4, 128]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.009784086818181818], 0, 2.8494021892547607, 1539264636.0501614], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float16"], ["TENSOR", [512, 512, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 28, 28, "float16"], [512, 512, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 18542, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [16, 32]], ["tile_t2", "sp", [32, 16]], ["c_unroll", "sp", [64, 8]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.02645109575], 0, 21.205586910247803, 1539265485.0608034], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float16"], ["TENSOR", [512, 256, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 28, 28, "float16"], [512, 256, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 19192, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [2, 128]], ["tile_t2", "sp", [4, 128]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.014060732125], 0, 4.281729221343994, 1539267047.7894878], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float16"], ["TENSOR", [256, 256, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 56, 56, "float16"], [256, 256, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 15917, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [16, 16]], ["tile_t2", "sp", [2, 128]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.03138662075], 0, 14.635979652404785, 1539268581.9926639], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float16"], ["TENSOR", [256, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 56, 56, "float16"], [256, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 14942, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [4, 32]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.017603293333333332], 0, 14.38584041595459, 1539270116.7375104], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float16"], ["TENSOR", [128, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 112, 112, "float16"], [128, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 15567, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [2, 64]], ["tile_t2", "sp", [4, 32]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.043861566], 0, 18.059239625930786, 1539271561.31132], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float16"], ["TENSOR", [128, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 112, 112, "float16"], [128, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 13642, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [1, 64]], ["tile_t2", "sp", [4, 32]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.02590457475], 0, 14.528096675872803, 1539272775.5760543], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float16"], ["TENSOR", [64, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 224, 224, "float16"], [64, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 16342, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [16, 4]], ["tile_t2", "sp", [16, 4]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.074903163], 0, 9.019776105880737, 1539274112.9994435], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float16"], ["TENSOR", [512, 512, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 512, 7, 7, "float16"], [512, 512, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 15967, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [8, 64]], ["tile_t2", "sp", [4, 128]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.0034370355333333334], 0, 8.091655731201172, 1539275339.858657], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float16"], ["TENSOR", [256, 256, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 256, 14, 14, "float16"], [256, 256, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 15692, "c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 8], ["tile_t1", "sp", [32, 8]], ["tile_t2", "sp", [4, 64]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.003097792484848485], 0, 2.525359630584717, 1539276221.117567], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float16"], ["TENSOR", [128, 128, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 128, 28, 28, "float16"], [128, 128, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 18161, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [2, 64]], ["tile_t2", "sp", [32, 4]], ["c_unroll", "sp", [16, 8]], ["yt", "ot", 4]], "t": "winograd"}], "r": [[0.003984467692307692], 0, 19.46451187133789, 1539277299.2439451], "v": 0.1}
+{"i": ["opencl -model=rk3399 -device=mali", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float16"], ["TENSOR", [64, 64, 3, 3], "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {}, ["conv2d", [1, 64, 56, 56, "float16"], [64, 64, 3, 3, "float16"], [1, 1], [1, 1], [1, 1], "NCHW", "float16"], {"i": 18786, "c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 4]], ["tile_t2", "sp", [16, 4]], ["c_unroll", "sp", [8, 8]], ["yt", "ot", 8]], "t": "winograd"}], "r": [[0.0050456273], 0, 3.70674729347229, 1539278982.2172227], "v": 0.1}
+{"v": 0.1, "r": [[0.0067990563], 0, 22.31303906440735, 1559268431.1042297], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 16, 16], "float32"], ["TENSOR", [24, 1024, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 16, 16, "float32"], [24, 1024, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 64]], ["tile_t2", "sp", [1, 24]], ["c_unroll", "sp", [256, 4]], ["yt", "ot", 4]], "i": 14049, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.017026335], 0, 4.165199518203735, 1559270460.5974245], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 16, 16], "float32"], ["TENSOR", [126, 1024, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 16, 16, "float32"], [126, 1024, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [256, 4]], ["tile_t2", "sp", [14, 9]], ["c_unroll", "sp", [128, 8]], ["yt", "ot", 2]], "i": 7144, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0005459967], 0, 3.179990530014038, 1559272070.366818], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 2, 2], "float32"], ["TENSOR", [126, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 2, 2, "float32"], [126, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [8, 32]], ["tile_t2", "sp", [6, 21]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 2]], "i": 7410, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0004307925], 0, 6.326959848403931, 1559273078.769135], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 2, 2], "float32"], ["TENSOR", [24, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 2, 2, "float32"], [24, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [4, 64]], ["tile_t2", "sp", [4, 6]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 4]], "i": 14840, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0003103615], 0, 4.570157766342163, 1559273725.7148163], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 1, 1], "float32"], ["TENSOR", [16, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 1, 1, "float32"], [16, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [32, 8]], ["tile_t2", "sp", [16, 1]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 8]], "i": 15076, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0005705555], 0, 1.499587059020996, 1559274415.4181104], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 1, 1], "float32"], ["TENSOR", [84, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 1, 1, "float32"], [84, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [128, 2]], ["tile_t2", "sp", [2, 42]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 16]], "i": 28575, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0030230367], 0, 13.507185220718384, 1559277613.6141922], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 8, 8], "float32"], ["TENSOR", [126, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 8, 8, "float32"], [126, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [4, 128]], ["tile_t2", "sp", [1, 126]], ["c_unroll", "sp", [64, 8]], ["yt", "ot", 2]], "i": 7674, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0240283161], 0, 5.942343235015869, 1559279137.098004], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [84, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 32, 32, "float32"], [84, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 64]], ["tile_t2", "sp", [3, 28]], ["c_unroll", "sp", [64, 8]], ["yt", "ot", 4]], "i": 17016, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0012389921], 0, 8.82778549194336, 1559280792.2976594], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 8, 8], "float32"], ["TENSOR", [24, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 8, 8, "float32"], [24, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [32, 16]], ["tile_t2", "sp", [3, 8]], ["c_unroll", "sp", [128, 4]], ["yt", "ot", 8]], "i": 18800, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0009596070999999999], 0, 4.888888120651245, 1559282674.247306], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 4, 4], "float32"], ["TENSOR", [24, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 4, 4, "float32"], [24, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [8, 32]], ["tile_t2", "sp", [1, 24]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 4]], "i": 15300, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.001159603], 0, 10.345348596572876, 1559283845.216672], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 4, 4], "float32"], ["TENSOR", [126, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 4, 4, "float32"], [126, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [2, 128]], ["tile_t2", "sp", [1, 126]], ["c_unroll", "sp", [64, 4]], ["yt", "ot", 1]], "i": 2870, "t": "winograd"}]}
+{"v": 0.1, "r": [[0.0229268894], 0, 16.91090679168701, 1559286752.655922], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 16, 16], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 16, 16, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [64, 4, 4]], ["tile_oh", "sp", [8, 2, 1]], ["tile_ow", "sp", [2, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 663389, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0013563563], 0, 2.7268688678741455, 1559287888.8589427], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 4, 4], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 4, 4, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [64, 2, 2]], ["tile_oh", "sp", [1, 2, 1]], ["tile_ow", "sp", [1, 2, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 18820, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0260932133], 0, 16.115710020065308, 1559289448.8561137], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 64, 64], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 64, 64, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [8, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 3898460, "t": "direct"}]}
+{"v": 0.1, "r": [[0.004960769], 0, 12.755321264266968, 1559292139.811835], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 64, 64], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 64, 64, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [32, 8, 1]], ["tile_y", "sp", [8, 1, 8]], ["tile_x", "sp", [4, 4, 4]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "i": 160833, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0005681632], 0, 3.639293909072876, 1559294895.475335], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 1, 1], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 1, 1, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [1, 64, 2]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "i": 5666, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0172897079], 0, 9.107371807098389, 1559296034.005847], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 64, 128, 128], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 128, 128, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [128, 1, 1]], ["tile_ow", "sp", [16, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 2166929, "t": "direct"}]}
+{"v": 0.1, "r": [[0.002727912], 0, 2.8317878246307373, 1559299247.5511405], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 32, 32, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [8, 8, 8]], ["tile_y", "sp", [4, 1, 8]], ["tile_x", "sp", [8, 2, 2]], ["ann_spatial", "an", ["none", "none", "vec"]]], "i": 57450, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0012113232], 0, 5.9466331005096436, 1559299788.4069092], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 8, 8], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 8, 8, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 16, 2]], ["tile_oh", "sp", [1, 1, 4]], ["tile_ow", "sp", [4, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["vec", "unroll", "unroll"]]], "i": 257818, "t": "direct"}]}
+{"v": 0.1, "r": [[0.022410575699999997], 0, 20.323858499526978, 1559305436.2717206], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 32, 256, 256], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 256, 256, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [4, 4, 4]], ["tile_oh", "sp", [256, 1, 1]], ["tile_ow", "sp", [64, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 2629635, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0139412919], 0, 2.466806650161743, 1559306617.3624876], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 64, 64], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 64, 64, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [32, 2, 1]], ["tile_ow", "sp", [1, 16, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 1503245, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0007715711999999999], 0, 4.0566606521606445, 1559308641.2995014], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 2, 2], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 2, 2, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [32, 8, 1]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "i": 5358, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0012850186], 0, 1.2971477508544922, 1559309502.9273202], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 8, 8], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 8, 8, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [1, 8, 1]], ["tile_ow", "sp", [2, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 377045, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0101076691], 0, 1.9567008018493652, 1559313232.1334348], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [16, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 32, 32, "float32"], [16, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [8, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 307766, "t": "direct"}]}
+{"v": 0.1, "r": [[0.011620686600000001], 0, 2.385251760482788, 1559314051.6586921], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 16, 16], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 16, 16, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [32, 8, 4]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [4, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 1583034, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0115034275], 0, 6.077609539031982, 1559318883.0443814], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 256, 256], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 256, 256, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [8, 8, 1]], ["tile_y", "sp", [128, 1, 1]], ["tile_x", "sp", [4, 8, 4]], ["ann_spatial", "an", ["none", "none", "vec"]]], "i": 90723, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0063606962], 0, 13.796538352966309, 1559322352.2132034], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 16, 16], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 16, 16, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [4, 16, 4]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [1, 4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 412446, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0065607544], 0, 1.4856116771697998, 1559324358.867971], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 16, 16], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 16, 16, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 16, 4]], ["tile_oh", "sp", [2, 4, 1]], ["tile_ow", "sp", [1, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 587533, "t": "direct"}]}
+{"v": 0.1, "r": [[0.030098252399999997], 0, 6.67178750038147, 1559325545.6214337], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 128, 128], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 128, 128, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [4, 8, 4]], ["tile_oh", "sp", [128, 1, 1]], ["tile_ow", "sp", [4, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 4968882, "t": "direct"}]}
+{"v": 0.1, "r": [[0.008022639600000001], 0, 1.5782084465026855, 1559327225.313344], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 256, 256], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 256, 256, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [4, 8, 1]], ["tile_y", "sp", [256, 1, 1]], ["tile_x", "sp", [4, 8, 8]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 238143, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0232035072], 0, 9.449883460998535, 1559330753.302027], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 32, 32, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [4, 32, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [1, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 2538714, "t": "direct"}]}
+{"v": 0.1, "r": [[0.000753604], 0, 5.480754852294922, 1559332216.319852], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 1, 1], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 1, 1, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [256, 1, 1]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "vec", "none"]]], "i": 5355, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0018052519], 0, 11.674946546554565, 1559333199.218867], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 32, 32, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [64, 8, 1]], ["tile_y", "sp", [16, 1, 1]], ["tile_x", "sp", [2, 2, 4]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "i": 57753, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0005852842], 0, 8.469638347625732, 1559336462.9289045], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 2, 2], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 2, 2, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [16, 4, 2]], ["tile_oh", "sp", [1, 2, 1]], ["tile_ow", "sp", [2, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "i": 12358, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0013613112], 0, 5.2336509227752686, 1559337920.329327], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 16, 16], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 16, 16, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [128, 8, 1]], ["tile_y", "sp", [8, 2, 1]], ["tile_x", "sp", [1, 2, 8]], ["ann_spatial", "an", ["unroll", "none", "unroll"]]], "i": 146589, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0006678313], 0, 5.503016233444214, 1559339272.7628164], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 4, 4], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 4, 4, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [4, 1, 1]], ["tile_ow", "sp", [4, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 142577, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0105797613], 0, 1.868051290512085, 1559340099.6998003], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 3, 512, 512], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 512, 512, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [1, 8, 4]], ["tile_oh", "sp", [256, 1, 1]], ["tile_ow", "sp", [2, 32, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 4528454, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0121671448], 0, 14.77577829360962, 1559343492.3986123], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 32, 32], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 32, 32, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"c": null, "e": [["tile_co", "sp", [4, 32, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [1, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 2538714, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0085935121], 0, 12.707570552825928, 1559345983.5270083], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 128, 128], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 128, 128, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [4, 8, 4]], ["tile_y", "sp", [1, 1, 128]], ["tile_x", "sp", [1, 32, 4]], ["ann_spatial", "an", ["vec", "none", "unroll"]]], "i": 820350, "t": "direct"}]}
+{"v": 0.1, "r": [[0.003444655], 0, 12.095378875732422, 1559346643.4826336], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 64, 64], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 64, 64, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [2, 8, 16]], ["tile_y", "sp", [2, 1, 16]], ["tile_x", "sp", [1, 8, 4]], ["ann_spatial", "an", ["none", "none", "vec"]]], "i": 53763, "t": "direct"}]}
+{"v": 0.1, "r": [[0.0062151208], 0, 10.776216745376587, 1559348857.8262913], "i": ["opencl -device=mali -model=rk3399", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 128, 128], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 128, 128, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"c": null, "e": [["tile_c", "sp", [8, 8, 2]], ["tile_y", "sp", [32, 2, 1]], ["tile_x", "sp", [4, 4, 4]], ["ann_spatial", "an", ["none", "none", "unroll"]]], "i": 43391, "t": "direct"}]}
+{"r": [[0.015079753900000001], 0, 10.554686546325684, 1559360302.1311853], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 16, 16], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 16, 16, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [8, 64]], ["tile_t2", "sp", [8, 64]], ["c_unroll", "sp", [256, 2]], ["yt", "ot", 4]], "i": 15762, "c": null}], "v": 0.1}
+{"r": [[0.0343192734], 0, 14.90624451637268, 1559361929.758143], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 2048, 16, 16], "float32"], ["TENSOR", [126, 2048, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 16, 16, "float32"], [126, 2048, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 128]], ["tile_t2", "sp", [7, 18]], ["c_unroll", "sp", [256, 8]], ["yt", "ot", 2]], "i": 7354, "c": null}], "v": 0.1}
+{"r": [[0.013624853900000001], 0, 5.254331827163696, 1559365112.0147774], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 2048, 16, 16], "float32"], ["TENSOR", [24, 2048, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 16, 16, "float32"], [24, 2048, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 2], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 128]], ["tile_t2", "sp", [3, 8]], ["c_unroll", "sp", [512, 4]], ["yt", "ot", 4]], "i": 13749, "c": null}], "v": 0.1}
+{"r": [[0.0420852927], 0, 14.921562194824219, 1559365901.5487695], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 64, 128, 128], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 128, 128, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 64]], ["tile_t2", "sp", [2, 32]], ["c_unroll", "sp", [32, 2]], ["yt", "ot", 4]], "i": 12062, "c": null}], "v": 0.1}
+{"r": [[0.0182222698], 0, 29.225221157073975, 1559370858.1307228], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 32, 32], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 32, 32, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [64, 4]], ["tile_t2", "sp", [64, 4]], ["c_unroll", "sp", [128, 2]], ["yt", "ot", 4]], "i": 14862, "c": null}], "v": 0.1}
+{"r": [[0.0005663256999999999], 0, 11.192577123641968, 1559372872.1332726], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 1, 1], "float32"], ["TENSOR", [84, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 1, 1, "float32"], [84, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [2, 64]], ["tile_t2", "sp", [4, 21]], ["c_unroll", "sp", [16, 8]], ["yt", "ot", 4]], "i": 16890, "c": null}], "v": 0.1}
+{"r": [[0.026175567599999998], 0, 13.929192066192627, 1559374151.747619], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 64, 64], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 64, 64, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 4], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [1, 128]], ["tile_t2", "sp", [4, 32]], ["c_unroll", "sp", [64, 2]], ["yt", "ot", 4]], "i": 15587, "c": null}], "v": 0.1}
+{"r": [[0.00039482909999999995], 0, 10.892908573150635, 1559375931.6253314], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 2, 2], "float32"], ["TENSOR", [16, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 2, 2, "float32"], [16, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [16, 16]], ["tile_t2", "sp", [16, 1]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 2]], "i": 7100, "c": null}], "v": 0.1}
+{"r": [[0.0003525069], 0, 7.068357229232788, 1559376853.808379], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 1, 1], "float32"], ["TENSOR", [16, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 1, 1, "float32"], [16, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [16, 8]], ["tile_t2", "sp", [16, 1]], ["c_unroll", "sp", [16, 8]], ["yt", "ot", 4]], "i": 11075, "c": null}], "v": 0.1}
+{"r": [[0.0008209203], 0, 6.32829475402832, 1559377722.370887], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 2, 2], "float32"], ["TENSOR", [84, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 2, 2, "float32"], [84, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 2], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 1], ["tile_t1", "sp", [16, 16]], ["tile_t2", "sp", [21, 4]], ["c_unroll", "sp", [32, 8]], ["yt", "ot", 4]], "i": 16260, "c": null}], "v": 0.1}
+{"r": [[0.0504606098], 0, 5.744432210922241, 1559378619.35616], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 32, 32], "float32"], ["TENSOR", [84, 1024, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 32, 32, "float32"], [84, 1024, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "winograd", "e": [["tile_size", "ot", 4], ["tile_bna", "ot", 1], ["tile_bnb", "ot", 4], ["tile_t1", "sp", [16, 64]], ["tile_t2", "sp", [21, 4]], ["c_unroll", "sp", [128, 8]], ["yt", "ot", 2]], "i": 10536, "c": null}], "v": 0.1}
+{"r": [[0.0469983427], 0, 24.310810565948486, 1559384692.5221972], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 32, 32], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 32, 32, "float32"], [512, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 16, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [4, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 2706188, "c": null}], "v": 0.1}
+{"r": [[0.0235342584], 0, 18.89966607093811, 1559386983.2081497], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 32, 32], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 32, 32, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 32, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [1, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 3250196, "c": null}], "v": 0.1}
+{"r": [[0.0260966263], 0, 23.66214871406555, 1559388589.8834436], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 64, 64], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 64, 64, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [16, 8, 4]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [4, 4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 4636962, "c": null}], "v": 0.1}
+{"r": [[0.0537888465], 0, 3.5150814056396484, 1559390497.2170131], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 64, 64], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 64, 64, "float32"], [256, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 8, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [8, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 2133830, "c": null}], "v": 0.1}
+{"r": [[0.0007365504999999999], 0, 4.358729839324951, 1559392310.4553668], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 2, 2], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 2, 2, "float32"], [128, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [64, 2, 1]], ["tile_oh", "sp", [1, 1, 1]], ["tile_ow", "sp", [1, 1, 1]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "unroll"]]], "i": 1369, "c": null}], "v": 0.1}
+{"r": [[0.0582818468], 0, 23.872197151184082, 1559393441.3053942], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 128, 128, 128], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 128, 128, "float32"], [128, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 4, 4]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [16, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 1226753, "c": null}], "v": 0.1}
+{"r": [[0.0606701301], 0, 7.118222236633301, 1559394945.4275196], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 32, 32], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 32, 32, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [16, 16, 8]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [4, 2, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 831907, "c": null}], "v": 0.1}
+{"r": [[0.022508868600000002], 0, 16.180081844329834, 1559397635.338322], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 16, 16], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 16, 16, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [512, 1, 4]], ["tile_oh", "sp", [8, 2, 1]], ["tile_ow", "sp", [1, 4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 732521, "c": null}], "v": 0.1}
+{"r": [[0.06352168459999999], 0, 23.5455265045166, 1559399702.9598434], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 128, 128], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 128, 128, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [4, 16, 8]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [2, 16, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 4630811, "c": null}], "v": 0.1}
+{"r": [[0.0285724813], 0, 5.5840654373168945, 1559409009.5330079], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 64, 64], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 64, 64, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [2, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 3036116, "c": null}], "v": 0.1}
+{"r": [[0.031217316099999996], 0, 7.627999305725098, 1559410758.7509966], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 64, 128, 128], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 128, 128, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [16, 4, 4]], ["tile_oh", "sp", [64, 2, 1]], ["tile_ow", "sp", [16, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 6266224, "c": null}], "v": 0.1}
+{"r": [[0.0125094221], 0, 7.986847400665283, 1559412127.0351486], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 2048, 16, 16], "float32"], ["TENSOR", [256, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 16, 16, "float32"], [256, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [16, 4, 4]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [4, 1, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 421219, "c": null}], "v": 0.1}
+{"r": [[0.061824211], 0, 22.509861707687378, 1559415622.6345937], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 64, 64], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 64, 64, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 16, 8]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [8, 2, 2]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 1319506, "c": null}], "v": 0.1}
+{"r": [[0.023426749599999998], 0, 6.4067089557647705, 1559417596.6296756], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 2048, 16, 16], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 16, 16, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 16, 4]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [1, 4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 1333223, "c": null}], "v": 0.1}
+{"r": [[0.0508724617], 0, 22.191868543624878, 1559419059.8858], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 32, 32], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 32, 32, "float32"], [512, 512, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [8, 16, 4]], ["tile_oh", "sp", [16, 1, 1]], ["tile_ow", "sp", [1, 4, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 1333223, "c": null}], "v": 0.1}
+{"r": [[0.022218506199999998], 0, 8.271515846252441, 1559420637.9464512], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 32, 32], "float32"], ["TENSOR", [16, 1024, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 32, 32, "float32"], [16, 1024, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [1, 4, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [4, 2, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "unroll"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 314696, "c": null}], "v": 0.1}
+{"r": [[0.0620065658], 0, 17.783520221710205, 1559423083.2628298], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 3, 512, 512], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 512, 512, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [2, 8, 4]], ["tile_oh", "sp", [256, 1, 1]], ["tile_ow", "sp", [8, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 6092116, "c": null}], "v": 0.1}
+{"r": [[0.050658181700000006], 0, 11.255446910858154, 1559426201.1772442], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 512, 64, 64], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 64, 64, "float32"], [256, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [2, 32, 4]], ["tile_oh", "sp", [64, 1, 1]], ["tile_ow", "sp", [2, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 3795142, "c": null}], "v": 0.1}
+{"r": [[0.0352125878], 0, 2.4257562160491943, 1559428961.9145355], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 128, 128], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 128, 128, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [1, 16, 4]], ["tile_oh", "sp", [128, 1, 1]], ["tile_ow", "sp", [4, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 1651121, "c": null}], "v": 0.1}
+{"r": [[0.0577996719], 0, 6.9262919425964355, 1559431611.3340707], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 256, 128, 128], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 128, 128, "float32"], [128, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [1, 32, 4]], ["tile_oh", "sp", [128, 1, 1]], ["tile_ow", "sp", [4, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["unroll", "none"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 5108852, "c": null}], "v": 0.1}
+{"r": [[0.010040239199999999], 0, 8.024097919464111, 1559440110.6810794], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 64, 128, 128], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 128, 128, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [2, 8, 4]], ["tile_oh", "sp", [128, 1, 1]], ["tile_ow", "sp", [1, 32, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "unroll"]], ["ann_spatial", "an", ["unroll", "unroll", "vec"]]], "i": 3902992, "c": null}], "v": 0.1}
+{"r": [[0.0248425953], 0, 4.38378381729126, 1559444221.1852398], "i": ["opencl -device=mali -model=rk3399", "topi_nn_conv2d", [["TENSOR", [1, 1024, 32, 32], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 32, 32, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"t": "direct", "e": [["tile_co", "sp", [4, 16, 4]], ["tile_oh", "sp", [32, 1, 1]], ["tile_ow", "sp", [1, 8, 4]], ["reorder_0", "re", [0, 1, 2, 3, 4, 5, 6, 9, 7, 8]], ["ann_reduce", "an", ["none", "none"]], ["ann_spatial", "an", ["none", "unroll", "vec"]]], "i": 826896, "c": null}], "v": 0.1}
diff --git a/tophub/opencl_v0.03.log b/tophub/opencl_v0.03.log
new file mode 100644
index 0000000..82c06f4
--- /dev/null
+++ b/tophub/opencl_v0.03.log
@@ -0,0 +1,99 @@
+# This is the pre-tuned parameters for opencl backend
+# TVM downloaded this during compilation
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [192, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [192, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 599486, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.342781593590029e-05], 0, 2.1750707626342773, 1535942689.128338], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [448, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [448, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2484179, "c": null, "e": [["tile_f", "sp", [14, 1, 16, 2]], ["tile_y", "sp", [2, 1, 2, 2]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.454847116430903e-05], 0, 2.046814203262329, 1535944879.5233593], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [384, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [384, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1240970, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 2]], ["tile_y", "sp", [2, 2, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 64]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.417353552109628e-05], 0, 1.1570026874542236, 1535946614.386767], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 8, 8], "float32"], ["TENSOR", [320, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 8, 8, "float32"], [320, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 733868, "c": null, "e": [["tile_f", "sp", [20, 2, 8, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [64, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.338510284697508e-05], 0, 3.2822842597961426, 1535947802.53725], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [192, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [192, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5706686, "c": null, "e": [["tile_f", "sp", [12, 1, 16, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [80, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.1632856215748484e-05], 0, 1.9522826671600342, 1535950169.2365394], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 448, 8, 8], "float32"], ["TENSOR", [384, 448, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 448, 8, 8, "float32"], [384, 448, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 359623, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [3, 1, 128, 1]], ["tile_x", "sp", [1, 1, 1, 4]], ["tile_rc", "sp", [56, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[8.294825981205085e-05], 0, 2.0430052280426025, 1535951107.239359], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [448, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [448, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 8529068, "c": null, "e": [["tile_f", "sp", [28, 2, 8, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [32, 40]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.8743621209302324e-05], 0, 2.2016215324401855, 1535952399.1231866], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 3, 1], "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 3, 1, "float32"], [1, 1], [1, 0], [1, 1], "NCHW", "float32"], {"i": 32537930, "c": null, "e": [["tile_f", "sp", [12, 1, 8, 2]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [12, 32]], ["tile_ry", "sp", [3, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.9095067702339854e-05], 0, 1.279364824295044, 1535954070.0828395], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 384, 8, 8], "float32"], ["TENSOR", [384, 384, 1, 3], "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 384, 8, 8, "float32"], [384, 384, 1, 3, "float32"], [1, 1], [0, 1], [1, 1], "NCHW", "float32"], {"i": 32519690, "c": null, "e": [["tile_f", "sp", [12, 1, 8, 2]], ["tile_y", "sp", [1, 1, 8, 1]], ["tile_x", "sp", [2, 1, 4, 1]], ["tile_rc", "sp", [12, 32]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.8831115547024953e-05], 0, 1.1108198165893555, 1535956028.9667456], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [384, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [384, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 15065930, "c": null, "e": [["tile_f", "sp", [12, 1, 8, 2]], ["tile_y", "sp", [2, 1, 4, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [80, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.358940787470089e-05], 0, 1.7977416515350342, 1535957603.3573184], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1280, 8, 8], "float32"], ["TENSOR", [320, 1280, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1280, 8, 8, "float32"], [320, 1280, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 868282, "c": null, "e": [["tile_f", "sp", [10, 2, 16, 1]], ["tile_y", "sp", [4, 1, 2, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [80, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.9327814489357546e-05], 0, 1.9696648120880127, 1535958717.215122], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2614152, "c": null, "e": [["tile_f", "sp", [6, 1, 32, 1]], ["tile_y", "sp", [8, 1, 1, 1]], ["tile_x", "sp", [1, 1, 8, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.083673854059609e-05], 0, 0.9849481582641602, 1535959906.3386817], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [320, 192, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [320, 192, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6319892, "c": null, "e": [["tile_f", "sp", [10, 1, 32, 1]], ["tile_y", "sp", [1, 1, 8, 1]], ["tile_x", "sp", [8, 1, 1, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.1193430016402407e-05], 0, 3.6221978664398193, 1535962514.8839347], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 624492, "c": null, "e": [["tile_f", "sp", [8, 1, 6, 4]], ["tile_y", "sp", [1, 1, 17, 1]], ["tile_x", "sp", [17, 1, 1, 1]], ["tile_rc", "sp", [32, 6]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010311639354395604], 0, 3.2178592681884766, 1535963369.2810538], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 17, 17], "float32"], ["TENSOR", [192, 192, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 17, 17, "float32"], [192, 192, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 180148, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [24, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.991701497902936e-05], 0, 1.1238107681274414, 1535964446.7551522], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 282292, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.453999006458022e-05], 0, 0.9816148281097412, 1535965554.1984963], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 312287, "c": null, "e": [["tile_f", "sp", [16, 2, 5, 1]], ["tile_y", "sp", [1, 1, 17, 1]], ["tile_x", "sp", [17, 1, 1, 1]], ["tile_rc", "sp", [32, 5]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.679889559877176e-05], 0, 1.0611205101013184, 1535966642.8761525], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [192, 160, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [192, 160, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 604851, "c": null, "e": [["tile_f", "sp", [8, 3, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [20, 8]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.634229465648855e-05], 0, 0.9411647319793701, 1535967539.0730731], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 160, 17, 17], "float32"], ["TENSOR", [160, 160, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 160, 17, 17, "float32"], [160, 160, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 191776, "c": null, "e": [["tile_f", "sp", [8, 4, 5, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 10]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.506076505766767e-05], 0, 0.955650806427002, 1535968464.5376928], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [160, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [160, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 216864, "c": null, "e": [["tile_f", "sp", [8, 4, 5, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [64, 12]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.578553904820012e-05], 0, 1.1183710098266602, 1535969541.9601598], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 104884, "c": null, "e": [["tile_f", "sp", [6, 4, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.025110642570281e-05], 0, 0.9510493278503418, 1535970261.585139], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 128902, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [1, 1, 17, 1]], ["tile_x", "sp", [17, 1, 1, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.752582439176544e-05], 0, 0.9907910823822021, 1535971071.4585369], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [192, 128, 7, 1], "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [192, 128, 7, 1, "float32"], [1, 1], [3, 0], [1, 1], "NCHW", "float32"], {"i": 355601, "c": null, "e": [["tile_f", "sp", [12, 2, 4, 2]], ["tile_y", "sp", [1, 1, 17, 1]], ["tile_x", "sp", [17, 1, 1, 1]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [7, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.026941405152225e-05], 0, 1.0132758617401123, 1535971904.5251734], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 17, 17], "float32"], ["TENSOR", [128, 128, 1, 7], "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 17, 17, "float32"], [128, 128, 1, 7, "float32"], [1, 1], [0, 3], [1, 1], "NCHW", "float32"], {"i": 160342, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.329875036075036e-05], 0, 1.020085096359253, 1535972678.511968], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [128, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [128, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 79702, "c": null, "e": [["tile_f", "sp", [8, 2, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [96, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.872497418688694e-05], 0, 0.9671390056610107, 1535973259.084541], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 768, 17, 17], "float32"], ["TENSOR", [192, 768, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 768, 17, 17, "float32"], [192, 768, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 427443, "c": null, "e": [["tile_f", "sp", [8, 3, 8, 1]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [48, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.786999106858055e-05], 0, 1.0456700325012207, 1535974319.6014495], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 575335, "c": null, "e": [["tile_f", "sp", [3, 1, 8, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [16, 6]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.053760632261551e-05], 0, 0.9735841751098633, 1535975347.1402736], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [384, 288, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [384, 288, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1524646, "c": null, "e": [["tile_f", "sp", [8, 2, 12, 2]], ["tile_y", "sp", [17, 1, 1, 1]], ["tile_x", "sp", [1, 1, 17, 1]], ["tile_rc", "sp", [288, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.0002976247188118812], 0, 0.9495353698730469, 1535976878.8081284], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [48, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [48, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 152985, "c": null, "e": [["tile_f", "sp", [2, 1, 6, 4]], ["tile_y", "sp", [7, 1, 5, 1]], ["tile_x", "sp", [7, 1, 5, 1]], ["tile_rc", "sp", [48, 6]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.844753159577084e-05], 0, 0.9524886608123779, 1535978408.9331803], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 288, 35, 35], "float32"], ["TENSOR", [64, 288, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 288, 35, 35, "float32"], [64, 288, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1666600, "c": null, "e": [["tile_f", "sp", [4, 2, 4, 2]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [36, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.3495715065502186e-05], 0, 3.2704806327819824, 1535979950.0025866], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [48, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [48, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1415744, "c": null, "e": [["tile_f", "sp", [3, 2, 4, 2]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.7756828716256043e-05], 0, 3.1801819801330566, 1535981312.0779464], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 35, 35], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 35, 35, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 849448, "c": null, "e": [["tile_f", "sp", [4, 2, 4, 2]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.0726684901531727e-05], 0, 0.9589455127716064, 1535982286.648798], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [32, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [32, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 708193, "c": null, "e": [["tile_f", "sp", [1, 4, 4, 1]], ["tile_y", "sp", [5, 1, 7, 1]], ["tile_x", "sp", [5, 1, 7, 1]], ["tile_rc", "sp", [12, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.086791046073602e-05], 0, 1.0268540382385254, 1535983359.5222244], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 96, 35, 35], "float32"], ["TENSOR", [96, 96, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 96, 35, 35, "float32"], [96, 96, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 4173095, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 4, 12]], ["tile_x", "sp", [3, 2, 54, 1]], ["tile_rc", "sp", [12, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[3.52060051558472e-05], 0, 1.1003777980804443, 1535985608.257474], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 35, 35], "float32"], ["TENSOR", [96, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 35, 35, "float32"], [96, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2448295, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 1, 4, 12]], ["tile_x", "sp", [3, 2, 54, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[2.6800415660511362e-05], 0, 1.1152160167694092, 1535988971.9531376], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 48, 35, 35], "float32"], ["TENSOR", [64, 48, 5, 5], "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 48, 35, 35, "float32"], [64, 48, 5, 5, "float32"], [1, 1], [2, 2], [1, 1], "NCHW", "float32"], {"i": 4135177, "c": null, "e": [["tile_f", "sp", [2, 1, 32, 1]], ["tile_y", "sp", [5, 1, 1, 7]], ["tile_x", "sp", [7, 1, 5, 1]], ["tile_rc", "sp", [16, 3]], ["tile_ry", "sp", [1, 5]], ["tile_rx", "sp", [1, 5]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.869008277404922e-05], 0, 1.0535945892333984, 1535991479.9687867], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [48, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [48, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 627303, "c": null, "e": [["tile_f", "sp", [3, 1, 4, 4]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.3947474349442379e-05], 0, 0.9443972110748291, 1535993295.6901307], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 192, 35, 35], "float32"], ["TENSOR", [64, 192, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 192, 35, 35, "float32"], [64, 192, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 677434, "c": null, "e": [["tile_f", "sp", [4, 1, 4, 4]], ["tile_y", "sp", [35, 1, 1, 1]], ["tile_x", "sp", [1, 1, 35, 1]], ["tile_rc", "sp", [48, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.6291235000542477e-05], 0, 2.0045502185821533, 1535994701.8173513], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 80, 73, 73], "float32"], ["TENSOR", [192, 80, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 80, 73, 73, "float32"], [192, 80, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 271675, "c": null, "e": [["tile_f", "sp", [8, 3, 2, 4]], ["tile_y", "sp", [71, 1, 1, 1]], ["tile_x", "sp", [1, 1, 71, 1]], ["tile_rc", "sp", [80, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[0.00039790470105820105], 0, 1.0133476257324219, 1535995615.459637], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 73, 73], "float32"], ["TENSOR", [80, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 73, 73, "float32"], [80, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 72820, "c": null, "e": [["tile_f", "sp", [5, 4, 2, 1]], ["tile_y", "sp", [73, 1, 1, 1]], ["tile_x", "sp", [1, 1, 73, 1]], ["tile_rc", "sp", [4, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.628059531304652e-05], 0, 2.2699286937713623, 1535996319.9562645], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 147, 147], "float32"], ["TENSOR", [64, 32, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 147, 147, "float32"], [64, 32, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 18619824, "c": null, "e": [["tile_f", "sp", [1, 1, 32, 2]], ["tile_y", "sp", [21, 1, 1, 7]], ["tile_x", "sp", [7, 1, 7, 3]], ["tile_rc", "sp", [32, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.0002058361879286694], 0, 3.723945140838623, 1535997522.3302574], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 149, 149], "float32"], ["TENSOR", [32, 32, 3, 3], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 149, 149, "float32"], [32, 32, 3, 3, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4885515, "c": null, "e": [["tile_f", "sp", [1, 2, 16, 1]], ["tile_y", "sp", [49, 3, 1, 1]], ["tile_x", "sp", [7, 1, 7, 3]], ["tile_rc", "sp", [32, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [3, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.693514331826986e-05], 0, 2.1389546394348145, 1535998840.1841948], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 299, 299], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 299, 299, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5855, "c": null, "e": [["tile_f", "sp", [2, 2, 1, 2]], ["tile_y", "sp", [149, 1, 1, 1]], ["tile_x", "sp", [1, 1, 149, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.8901995718423374e-05], 0, 0.9440114498138428, 1535999867.0320988], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 813594, "c": null, "e": [["tile_f", "sp", [32, 1, 16, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [7, 1, 1, 1]], ["tile_rc", "sp", [256, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00021621475216138328], 0, 0.8917639255523682, 1535866883.280491], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 110920, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [16, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.48512566078221e-05], 0, 5.045914173126221, 1535867796.3817327], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 605914, "c": null, "e": [["tile_f", "sp", [32, 1, 16, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [7, 1, 1, 1]], ["tile_rc", "sp", [128, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00011497667049808429], 0, 3.344743251800537, 1535868716.7805383], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3907436, "c": null, "e": [["tile_f", "sp", [8, 1, 16, 2]], ["tile_y", "sp", [7, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [3, 1]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[8.380603852596314e-05], 0, 0.9502370357513428, 1535869900.1723077], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 147865, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 1]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [16, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[9.69275167396343e-06], 0, 0.9375286102294922, 1535871182.315236], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 3780511, "c": null, "e": [["tile_f", "sp", [8, 2, 16, 1]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 2]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.495314048392922e-05], 0, 2.2556657791137695, 1535872698.2013803], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3153679, "c": null, "e": [["tile_f", "sp", [4, 1, 8, 4]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[7.75432750089272e-06], 0, 1.1662640571594238, 1535876319.7245193], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 15216898, "c": null, "e": [["tile_f", "sp", [4, 1, 16, 2]], ["tile_y", "sp", [7, 2, 2, 1]], ["tile_x", "sp", [7, 1, 4, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.09023181570338e-05], 0, 9.874000072479248, 1535878659.5294101], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12546385, "c": null, "e": [["tile_f", "sp", [2, 1, 8, 4]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.237825215529076e-06], 0, 7.1099395751953125, 1535880001.3221123], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 72845770, "c": null, "e": [["tile_f", "sp", [2, 1, 16, 2]], ["tile_y", "sp", [4, 7, 2, 1]], ["tile_x", "sp", [7, 1, 4, 2]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.223034496124031e-05], 0, 1.1339690685272217, 1535881228.9030168], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 7, 7], "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 7, 7, "float32"], [2, 2], [3, 3], [1, 1], "NCHW", "float32"], {"i": 36658173, "c": null, "e": [["tile_f", "sp", [2, 2, 2, 8]], ["tile_y", "sp", [4, 4, 7, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 7]], ["tile_rx", "sp", [1, 7]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.973905860927153e-05], 0, 1.0524730682373047, 1535883202.333643], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 7, 7, "float32"], [1024, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 218017, "c": null, "e": [["tile_f", "sp", [128, 1, 4, 2]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[7.626181074297188e-05], 0, 0.9035851955413818, 1535895423.4092102], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 1024, 7, 7], "float32"], ["TENSOR", [1024, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 1024, 7, 7, "float32"], [1024, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 12024, "c": null, "e": [["tile_f", "sp", [256, 2, 2, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.651787883478315e-06], 0, 2.234114170074463, 1535895988.8837981], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [1024, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 246005, "c": null, "e": [["tile_f", "sp", [32, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 1, 7]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.992867340963535e-05], 0, 3.2919204235076904, 1535896528.8578045], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2227, "c": null, "e": [["tile_f", "sp", [64, 1, 4, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.621260630833994e-06], 0, 1.0052404403686523, 1535897064.3757832], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 1898195, "c": null, "e": [["tile_f", "sp", [16, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 7, 2]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.714989559623948e-05], 0, 0.9037055969238281, 1535898086.2151685], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 512, 14, 14, "float32"], [512, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 255035, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 2]], ["tile_y", "sp", [1, 2, 7, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[4.62717939793613e-06], 0, 3.6174981594085693, 1535899046.58141], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [512, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 141760, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.980510835750066e-05], 0, 1.9354472160339355, 1535900878.3579087], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 64680, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [1, 1, 14, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.3707222380918375e-06], 0, 1.1024589538574219, 1535901160.5987856], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [256, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 3016464, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 4]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.3563316158156913e-05], 0, 0.9629583358764648, 1535902221.0082664], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [256, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 256, 28, 28, "float32"], [256, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 908160, "c": null, "e": [["tile_f", "sp", [256, 1, 1, 1]], ["tile_y", "sp", [2, 1, 7, 2]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.624383343928799e-06], 0, 2.0657784938812256, 1535903259.8653224], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [256, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [256, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6976464, "c": null, "e": [["tile_f", "sp", [8, 1, 8, 4]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.3044423123697012e-05], 0, 2.186422109603882, 1535904185.8950636], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 659640, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [1, 1, 7, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[8.203848099885502e-06], 0, 1.1230483055114746, 1535905223.5894065], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [128, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 26371343, "c": null, "e": [["tile_f", "sp", [4, 4, 8, 1]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.7854345896769402e-05], 0, 1.1477160453796387, 1535906530.7973206], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [128, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 128, 56, 56, "float32"], [128, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 273840, "c": null, "e": [["tile_f", "sp", [128, 1, 1, 1]], ["tile_y", "sp", [4, 1, 7, 2]], ["tile_x", "sp", [1, 2, 28, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.1798222475263075e-05], 0, 1.9302966594696045, 1535908062.110772], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [128, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [128, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 12566617, "c": null, "e": [["tile_f", "sp", [2, 1, 4, 8]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [1, 1, 56, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.4122842600432777e-05], 0, 3.6323812007904053, 1535909851.2175212], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [64, 1, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 64, 112, 112, "float32"], [64, 1, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "float32"], {"i": 2346456, "c": null, "e": [["tile_f", "sp", [64, 1, 1, 1]], ["tile_y", "sp", [14, 1, 4, 1]], ["tile_x", "sp", [1, 1, 56, 1]], ["auto_unroll_max_step", "ot", 256], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.6541795141168745e-05], 0, 1.0568253993988037, 1535910829.1991868], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [64, 32, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 32, 112, 112, "float32"], [64, 32, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 44888048, "c": null, "e": [["tile_f", "sp", [2, 2, 8, 2]], ["tile_y", "sp", [56, 2, 1, 1]], ["tile_x", "sp", [7, 1, 16, 1]], ["tile_rc", "sp", [4, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[1.7799144348238644e-05], 0, 2.851109266281128, 1535911768.7366812], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_depthwise_conv2d_nchw", [["TENSOR", [1, 32, 112, 112], "float32"], ["TENSOR", [32, 1, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "float32"], {}, ["depthwise_conv2d_nchw", [1, 32, 112, 112, "float32"], [32, 1, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "float32"], {"i": 349720, "c": null, "e": [["tile_f", "sp", [32, 1, 1, 1]], ["tile_y", "sp", [7, 1, 2, 4]], ["tile_x", "sp", [1, 1, 112, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.0858482830911588e-05], 0, 0.9344322681427002, 1535913153.1235378], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [32, 3, 3, 3], "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [32, 3, 3, 3, "float32"], [2, 2], [1, 1], [1, 1], "NCHW", "float32"], {"i": 24494779, "c": null, "e": [["tile_f", "sp", [2, 1, 2, 4]], ["tile_y", "sp", [28, 2, 1, 2]], ["tile_x", "sp", [1, 1, 112, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[1.027193146182814e-05], 0, 2.9933173656463623, 1535915195.945269], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 2048, 7, 7], "float32"], ["TENSOR", [512, 2048, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 2048, 7, 7, "float32"], [512, 2048, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 227068, "c": null, "e": [["tile_f", "sp", [32, 2, 8, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [128, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00010686423257467993], 0, 0.9232778549194336, 1535867303.1553254], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 7, 7], "float32"], ["TENSOR", [2048, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 7, 7, "float32"], [2048, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 79030, "c": null, "e": [["tile_f", "sp", [128, 1, 16, 1]], ["tile_y", "sp", [1, 7, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.712371454138702e-05], 0, 1.3648619651794434, 1535868119.6373053], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [2048, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [2048, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 211177, "c": null, "e": [["tile_f", "sp", [32, 1, 64, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 7, 1, 1]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[0.00014776954527559056], 0, 5.306933641433716, 1535868908.027305], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [512, 1024, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [512, 1024, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 132040, "c": null, "e": [["tile_f", "sp", [16, 1, 32, 1]], ["tile_y", "sp", [7, 1, 1, 1]], ["tile_x", "sp", [1, 1, 7, 1]], ["tile_rc", "sp", [64, 16]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[5.30868634594977e-05], 0, 1.0211074352264404, 1535870556.7032566], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 1024, 14, 14], "float32"], ["TENSOR", [256, 1024, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 1024, 14, 14, "float32"], [256, 1024, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 148566, "c": null, "e": [["tile_f", "sp", [16, 1, 8, 2]], ["tile_y", "sp", [7, 1, 2, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [128, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[4.147794883081156e-05], 0, 0.9938898086547852, 1535871441.9406877], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 14, 14], "float32"], ["TENSOR", [1024, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 14, 14, "float32"], [1024, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 183420, "c": null, "e": [["tile_f", "sp", [32, 2, 8, 2]], ["tile_y", "sp", [7, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.9574581707076678e-05], 0, 0.9183127880096436, 1535872645.4647102], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [1024, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [1024, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 256582, "c": null, "e": [["tile_f", "sp", [16, 4, 16, 1]], ["tile_y", "sp", [7, 2, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[6.666271079520214e-05], 0, 0.9824354648590088, 1535874962.9403217], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [256, 512, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [256, 512, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 570265, "c": null, "e": [["tile_f", "sp", [16, 2, 8, 1]], ["tile_y", "sp", [14, 1, 1, 1]], ["tile_x", "sp", [1, 1, 14, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.215648131539611e-05], 0, 0.99273681640625, 1535878228.9984486], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [128, 512, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [128, 512, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 2546667, "c": null, "e": [["tile_f", "sp", [4, 2, 16, 1]], ["tile_y", "sp", [7, 1, 2, 2]], ["tile_x", "sp", [7, 1, 4, 1]], ["tile_rc", "sp", [64, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[3.262830028267015e-05], 0, 0.9175102710723877, 1535879342.9020834], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [512, 128, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [512, 128, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 6133969, "c": null, "e": [["tile_f", "sp", [16, 1, 4, 8]], ["tile_y", "sp", [14, 2, 1, 1]], ["tile_x", "sp", [1, 1, 28, 1]], ["tile_rc", "sp", [64, 2]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.2630357713768663e-05], 0, 1.0878932476043701, 1535880369.9766262], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [512, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [512, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 4012837, "c": null, "e": [["tile_f", "sp", [4, 8, 16, 1]], ["tile_y", "sp", [28, 1, 1, 1]], ["tile_x", "sp", [1, 2, 14, 1]], ["tile_rc", "sp", [64, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.576615161649944e-05], 0, 2.4738738536834717, 1535882784.7719417], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 28, 28], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 28, 28, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 5088753, "c": null, "e": [["tile_f", "sp", [2, 1, 64, 1]], ["tile_y", "sp", [14, 1, 2, 1]], ["tile_x", "sp", [2, 7, 1, 2]], ["tile_rc", "sp", [32, 4]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.4963060007317964e-05], 0, 0.9677717685699463, 1535885179.2824543], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [128, 256, 1, 1], "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [128, 256, 1, 1, "float32"], [2, 2], [0, 0], [1, 1], "NCHW", "float32"], {"i": 5808751, "c": null, "e": [["tile_f", "sp", [2, 2, 32, 1]], ["tile_y", "sp", [14, 1, 2, 1]], ["tile_x", "sp", [7, 1, 4, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[2.8338558367963735e-05], 0, 2.3057353496551514, 1535885892.4977822], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [64, 256, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [64, 256, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 25955710, "c": null, "e": [["tile_f", "sp", [2, 1, 16, 2]], ["tile_y", "sp", [14, 1, 2, 2]], ["tile_x", "sp", [7, 1, 8, 1]], ["tile_rc", "sp", [32, 8]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "direct"}], "r": [[3.047601584721658e-05], 0, 1.3348917961120605, 1535887402.021358], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 56, 56], "float32"], ["TENSOR", [256, 64, 1, 1], "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 56, 56, "float32"], [256, 64, 1, 1, "float32"], [1, 1], [0, 0], [1, 1], "NCHW", "float32"], {"i": 9860591, "c": null, "e": [["tile_f", "sp", [8, 4, 8, 1]], ["tile_y", "sp", [28, 2, 1, 1]], ["tile_x", "sp", [2, 1, 28, 1]], ["tile_rc", "sp", [16, 4]], ["tile_ry", "sp", [1, 1]], ["tile_rx", "sp", [1, 1]], ["auto_unroll_max_step", "ot", 512], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[2.8265906061746987e-05], 0, 0.9717552661895752, 1535889103.8465264], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 14, 14], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 14, 14, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 70670, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [2, 2, 128, 1]], ["tile_x", "sp", [7, 7, 1, 1]], ["tile_rc", "sp", [128, 4]], ["auto_unroll_max_step", "ot", 0], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0001320726672535211], 0, 1.2616889476776123, 1535916700.0729654], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 512, 28, 28], "float32"], ["TENSOR", [512, 512, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 512, 28, 28, "float32"], [512, 512, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1169650, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [64, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00027391162815884476], 0, 1.2446036338806152, 1535917844.5701568], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 28, 28], "float32"], ["TENSOR", [512, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 28, 28, "float32"], [512, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1059650, "c": null, "e": [["tile_size", "ot", 2], ["tile_b", "sp", [16, 1, 1, 1]], ["tile_y", "sp", [4, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [32, 8]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00015555756476683938], 0, 1.1849892139434814, 1535920310.5176947], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 256, 56, 56], "float32"], ["TENSOR", [256, 256, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 256, 56, 56, "float32"], [256, 256, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 629746, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 1, 32, 4]], ["tile_x", "sp", [7, 7, 4, 1]], ["tile_rc", "sp", [64, 4]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00021653877777777776], 0, 2.4224183559417725, 1535921926.4664695], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 56, 56], "float32"], ["TENSOR", [256, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 56, 56, "float32"], [256, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 698740, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [16, 1, 1, 16]], ["tile_x", "sp", [1, 2, 98, 1]], ["tile_rc", "sp", [32, 4]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00016098087299893274], 0, 2.0246737003326416, 1535923803.6749535], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 128, 112, 112], "float32"], ["TENSOR", [128, 128, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 128, 112, 112, "float32"], [128, 128, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1389683, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 2, 16, 4]], ["tile_x", "sp", [28, 7, 4, 1]], ["tile_rc", "sp", [64, 2]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00030931965030674846], 0, 2.1483683586120605, 1535926949.863795], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 112, 112], "float32"], ["TENSOR", [128, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 112, 112, "float32"], [128, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 1315539, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [1, 1, 16, 8]], ["tile_x", "sp", [49, 2, 4, 2]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.00021050380138888888], 0, 2.089472532272339, 1535929147.0209825], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 64, 224, 224], "float32"], ["TENSOR", [64, 64, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 64, 224, 224, "float32"], [64, 64, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 2196845, "c": null, "e": [["tile_size", "ot", 4], ["tile_b", "sp", [36, 1, 1, 1]], ["tile_y", "sp", [2, 1, 2, 16]], ["tile_x", "sp", [49, 2, 32, 1]], ["tile_rc", "sp", [8, 8]], ["auto_unroll_max_step", "ot", 128], ["unroll_explicit", "ot", 1]], "t": "winograd"}], "r": [[0.0004802369904458598], 0, 6.3500049114227295, 1535933803.1287143], "v": 0.1}
+{"i": ["opencl -model=1080ti", "topi_nn_conv2d", [["TENSOR", [1, 3, 224, 224], "float32"], ["TENSOR", [64, 3, 3, 3], "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {}, ["conv2d", [1, 3, 224, 224, "float32"], [64, 3, 3, 3, "float32"], [1, 1], [1, 1], [1, 1], "NCHW", "float32"], {"i": 93760310, "c": null, "e": [["tile_f", "sp", [8, 2, 4, 1]], ["tile_y", "sp", [56, 4, 1, 1]], ["tile_x", "sp", [7, 1, 32, 1]], ["tile_rc", "sp", [3, 1]], ["tile_ry", "sp", [1, 3]], ["tile_rx", "sp", [1, 3]], ["auto_unroll_max_step", "ot", 1500], ["unroll_explicit", "ot", 0]], "t": "direct"}], "r": [[5.1508467961826863e-05], 0, 0.9860408306121826, 1535934993.9483538], "v": 0.1}