graph: backend: compiler: ops: support transpose with no permutation …

…axis

src/graph/backend/graph_compiler/core/src/compiler/ir/graph/transform/tensor_view_transform.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
- * Copyright 2021-2023 Intel Corporation
+ * Copyright 2021-2024 Intel Corporation
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -215,9 +215,10 @@ bool should_transform_reorder(const sc_op_ptr &node) {
 
 static bool should_transform_transpose(const sc_op_ptr &node) {
     assert(node->isa<transpose_op_t>());
-    if (node->info_.inputs_[0]->details_.get_format()
-            != node->info_.outputs_[0]->details_.get_format()) {
-        // means that we have permuted the data format
+    // all transpose op can be converted to tensor_view as long as layout
+    // propagation is in effect
+    if (node->attrs_.get_or_else("layout_transformed", false)) {
+        // means that we have transformed the data format
         return true;
     }
     return false;

src/graph/backend/graph_compiler/core/src/ops/fusible/memory_movement.cpp

@@ -506,16 +506,9 @@ void transpose_op_t::query_format(context_ptr ctx,
     auto in_format = info_.inputs_[0]->details_.get_format();
     auto in_strides = info_.inputs_[0]->details_.get_strides();
     auto in_format_code = in_format.format_code_;
-    int batch_dims = info_.inputs_[0]->details_.get_plain_dims().size()
-            - in_format_code.norig_dims();
     std::unordered_map<int, int> order_map;
     for (size_t i = 0; i < order_.size(); ++i) {
-        COMPILE_ASSERT((order_[i] >= batch_dims)
-                        == (static_cast<int>(i) >= batch_dims),
-                "Permutation on batch dims is not supported.")
-        if (order_[i] >= batch_dims && static_cast<int>(i) >= batch_dims) {
-            order_map[order_[i] - batch_dims] = i - batch_dims;
-        }
+        order_map[order_[i]] = i;
     }
 
     std::vector<int> storage_args;
@@ -525,6 +518,8 @@ void transpose_op_t::query_format(context_ptr ctx,
     }
     auto out_format = sc_data_format_t(storage_args, in_format.blocks_);
 
+    attrs_["layout_transformed"] = true;
+
     supported_ins.resize(1);
     supported_outs.resize(1);
     supported_ins[0].emplace_back(std::make_pair(in_format, in_strides));

tests/gtests/graph/unit/backend/graph_compiler/core/test_transpose.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
- * Copyright 2020-2023 Intel Corporation
+ * Copyright 2020-2024 Intel Corporation
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -37,10 +37,11 @@ using namespace dnnl::impl::graph::gc;
 
 template <class T>
 static std::vector<T> ref_transpose(const std::vector<T> &data,
-        const sc_dims &input_dims, const std::vector<int> &axis) {
-    assert(axis.size() == 2);
+        const sc_dims &input_dims, std::vector<int> axis) {
+    assert(axis.size() == 2 || axis.empty());
     const int num_of_loops = input_dims.size();
     sc_dims output_dims = input_dims;
+    if (axis.empty()) { axis = {0, 0}; }
     std::swap(output_dims[axis[0]], output_dims[axis[1]]);
     std::vector<int> lp_vars(num_of_loops);
     std::vector<T> ret(data.size());
@@ -155,4 +156,5 @@ TEST(GCCore_CPU_transpose_test, TestQueryFormat) {
 TEST(GCCore_CPU_transpose_test, TestSingleTranspose) {
     transpose_test({4, 8, 16, 32}, {0, 1, 3, 2}, {2, 3});
     transpose_test({4, 8, 16, 32}, {0, 2, 1, 3}, {1, 2});
+    transpose_test({4, 8, 16, 32}, {0, 1, 2, 3}, {});
 }