Replace lut yaml with table header

BUILD.bazel

@@ -195,6 +195,7 @@ MICROKERNEL_DEFS = [
     "src/xx-fill/xx-fill.h",
     "src/xx-pad/xx-pad.h",
     "src/xx-transposev/xx-transposev.h",
+    "src/x8-lut/x8-lut.h",
     "src/x8-packq/x8-packq.h",
     "src/x8-packw/x8-packw.h",
     "src/x8-transposec/x8-transposec.h",

scripts/generate-tests.sh

@@ -192,9 +192,6 @@ tools/generate-rdsum-test.py --spec test/f32-rdsum.yaml --output test/f32-rdsum.
 tools/generate-rdsum-test.py --spec test/qs8-rdsum-minmax-fp32.yaml --output test/qs8-rdsum-minmax-fp32.cc &
 tools/generate-rdsum-test.py --spec test/qu8-rdsum.yaml --output test/qu8-rdsum.cc &
 
-### Tests for LUT micro-kernels
-tools/generate-lut-test.py --spec test/x8-lut.yaml --output test/x8-lut.cc &
-
 ### Tests for Conv HWC layout micro-kernels
 tools/generate-conv-hwc-test.py --spec test/f32-conv-hwc.yaml --output test/f32-conv-hwc.cc &
 

src/x8-lut/x8-lut.h

@@ -0,0 +1,80 @@
+// Copyright 2023 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#ifndef XNN_UKERNEL_WITH_PARAMS
+#define XNN_UKERNEL_WITH_PARAMS(arch_flags, ukernel, batch_size, datatype, params_type, init_params)                   \
+  XNN_UKERNEL(arch_flags, ukernel, batch_size, datatype)
+#define XNN_DEFINED_UKERNEL_WITH_PARAMS
+#endif
+
+#ifndef XNN_UKERNEL
+#define XNN_UKERNEL(arch_flags, ukernel, batch_size, datatype)                                                         \
+  XNN_UKERNEL_WITH_PARAMS(arch_flags, ukernel, batch_size, datatype, void, /*init_params=*/nullptr)
+#define XNN_DEFINED_UKERNEL
+#endif
+
+XNN_UKERNEL_WITH_PARAMS(0, xnn_x8_lut_ukernel__scalar_u1, 1, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(0, xnn_x8_lut_ukernel__scalar_u2, 2, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(0, xnn_x8_lut_ukernel__scalar_u4, 4, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(0, xnn_x8_lut_ukernel__scalar_u8, 8, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(0, xnn_x8_lut_ukernel__scalar_u16, 16, uint8_t, void, nullptr)
+
+#if XNN_ARCH_ARM64
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u16, 16, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u48, 48, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u64, 64, uint8_t, void, nullptr)
+#endif  // XNN_ARCH_ARM64
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__ssse3_u16, 16, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__ssse3_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx_u16, 16, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx_u48, 48, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx_u64, 64, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx2_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx2_u64, 64, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx2_u96, 96, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_ssse3, xnn_x8_lut_ukernel__avx2_u128, 128, uint8_t, void, nullptr)
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
+
+#if XNN_ENABLE_AVX512VBMI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512vbmi, xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u64, 64, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512vbmi, xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u128, 128, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512vbmi, xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u192, 192, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512vbmi, xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u256, 256, uint8_t, void, nullptr)
+#endif  // XNN_ENABLE_AVX512VBMI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+
+#if XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512skx, xnn_x8_lut_ukernel__avx512skx_vpshufb_u64, 64, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512skx, xnn_x8_lut_ukernel__avx512skx_vpshufb_u128, 128, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512skx, xnn_x8_lut_ukernel__avx512skx_vpshufb_u192, 192, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_x86_avx512skx, xnn_x8_lut_ukernel__avx512skx_vpshufb_u256, 256, uint8_t, void, nullptr)
+#endif  // XNN_ENABLE_AVX512SKX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
+
+#if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmsimd_u16, 16, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmsimd_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmsimd_u48, 48, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmsimd_u64, 64, uint8_t, void, nullptr)
+#endif  // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
+
+#if XNN_ARCH_WASMRELAXEDSIMD
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmpshufb_u16, 16, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmpshufb_u32, 32, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmpshufb_u48, 48, uint8_t, void, nullptr)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_wasmsimd, xnn_x8_lut_ukernel__wasmpshufb_u64, 64, uint8_t, void, nullptr)
+#endif  // XNN_ARCH_WASMRELAXEDSIMD
+
+#ifdef XNN_DEFINED_UKERNEL_WITH_PARAMS
+#undef XNN_DEFINED_UKERNEL_WITH_PARAMS
+#undef XNN_UKERNEL_WITH_PARAMS
+#endif
+
+#ifdef XNN_DEFINED_UKERNEL
+#undef XNN_DEFINED_UKERNEL
+#undef XNN_UKERNEL
+#endif

src/xnnpack/lut.h

@@ -18,56 +18,14 @@ extern "C" {
 #endif
 
 
-#define DECLARE_X8_LUT_UKERNEL_FUNCTION(fn_name) \
+#define XNN_UKERNEL(arch_flags, fn_name, batch_tile, datatype) \
   XNN_INTERNAL void fn_name(                     \
       size_t n,                                  \
       const uint8_t* x,                          \
       uint8_t* y,                                \
       const uint8_t table[XNN_RESTRICT XNN_MIN_ELEMENTS(256)]);
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__scalar_u1)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__scalar_u2)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__scalar_u4)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__scalar_u8)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__scalar_u16)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u16)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u32)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u48)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_u64)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__ssse3_u16)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__ssse3_u32)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx_u16)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx_u32)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx_u48)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx_u64)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx2_u32)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx2_u64)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx2_u96)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx2_u128)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512skx_vpshufb_u64)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512skx_vpshufb_u128)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512skx_vpshufb_u192)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512skx_vpshufb_u256)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u64)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u128)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u192)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__avx512vbmi_vpermx2b_u256)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmpshufb_u16)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmpshufb_u32)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmpshufb_u48)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmpshufb_u64)
-
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmsimd_u16)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmsimd_u32)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmsimd_u48)
-DECLARE_X8_LUT_UKERNEL_FUNCTION(xnn_x8_lut_ukernel__wasmsimd_u64)
+#include "x8-lut/x8-lut.h"
+#undef XNN_UKERNEL
 
 
 #define DECLARE_U8_LUT32NORM_UKERNEL_FUNCTION(fn_name) \

test/BUILD.bazel

@@ -1059,7 +1059,6 @@ xnnpack_unit_test(
 xnnpack_unit_test(
     name = "x8_lut_test",
     srcs = [
-        "lut-microkernel-tester.h",
         "x8-lut.cc",
     ],
     deps = MICROKERNEL_TEST_DEPS,