forked from openvinotoolkit/openvino
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[CPU] MergeTransposeReorder extending
- Loading branch information
Showing
7 changed files
with
635 additions
and
206 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Large diffs are not rendered by default.
Oops, something went wrong.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
176 changes: 176 additions & 0 deletions
176
src/plugins/intel_cpu/tests/functional/subgraph_tests/src/merge_transpose_reorder.cpp
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,176 @@ | ||
| // Copyright (C) 2018-2023 Intel Corporation | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
| // | ||
|
|
||
| #include <memory> | ||
| #include <string> | ||
| #include <tuple> | ||
| #include <vector> | ||
|
|
||
| #include "common_test_utils/common_utils.hpp" | ||
| #include "openvino/opsets/opset10.hpp" | ||
| #include "ov_models/builders.hpp" | ||
| #include "shared_test_classes/base/ov_subgraph.hpp" | ||
| #include "test_utils/cpu_test_utils.hpp" | ||
| #include "transformations/utils/utils.hpp" | ||
|
|
||
| using namespace CPUTestUtils; | ||
| using namespace ov::test; | ||
|
|
||
| namespace CPUSubgraphTestsDefinitions { | ||
| template <typename NodeType, typename... Args> | ||
| static std::shared_ptr<ov::Node> make_layer_with_bias(Args&&... args) { | ||
| const auto node = std::make_shared<NodeType>(std::forward<Args>(args)...); | ||
| const auto& precision = node->get_output_element_type(0); | ||
| const auto bias_const = ngraph::builder::makeConstant(precision, ov::Shape{}, std::vector<float>{}, true); | ||
| const auto bias = std::make_shared<ov::opset10::Add>(node, bias_const); | ||
| return bias; | ||
| } | ||
|
|
||
| /* | ||
| Parameter(4D) | ||
| | | ||
| Reshape(3D) | ||
| | | ||
| Transpose(0, 2, 1) | ||
| | | ||
| MatMul | ||
| | | ||
| Transpose(0, 2, 1) | ||
| | | ||
| Reshape(4D) | ||
| | | ||
| GroupConvolution | ||
| | | ||
| Reshape(3D) | ||
| | | ||
| Transpose(0, 2, 1) | ||
| | | ||
| MatMul | ||
| */ | ||
| using MergeTransposeReorderTestParams = std::tuple<InputShape, ElementType, size_t>; | ||
| class MergeTransposeReorderCPUTest : public testing::WithParamInterface<MergeTransposeReorderTestParams>, virtual public SubgraphBaseTest, public CPUTestsBase { | ||
| public: | ||
| static std::string getTestCaseName(const testing::TestParamInfo<MergeTransposeReorderTestParams> &obj) { | ||
| InputShape input_shape; | ||
| ElementType precision; | ||
| size_t optimized_reorders_count; | ||
| std::tie(input_shape, precision, optimized_reorders_count) = obj.param; | ||
|
|
||
| std::ostringstream results; | ||
| results << "IS=(" << ov::test::utils::partialShape2str({input_shape.first}) << "_"; | ||
| results << ")_TS=("; | ||
| for (const auto& static_shape : input_shape.second) { | ||
| results << ov::test::utils::vec2str(static_shape) << "_"; | ||
| } | ||
| results << ")_precision=" << precision; | ||
| results << ")_optimized_reorders_count=" << optimized_reorders_count; | ||
| return results.str(); | ||
| } | ||
|
|
||
| protected: | ||
| void SetUp() override { | ||
| targetDevice = ov::test::utils::DEVICE_CPU; | ||
| InputShape input_shape; | ||
| ElementType precision; | ||
| std::tie(input_shape, precision, m_optimized_reorders_count) = this->GetParam(); | ||
| init_input_shapes({input_shape}); | ||
|
|
||
| const auto shapeof_subgraph_prc = ov::element::i32; | ||
| OPENVINO_ASSERT(inputDynamicShapes[0].rank().is_static() && inputDynamicShapes[0].size() == 4, "initSubgraph: only 4D shapes are supported"); | ||
| OPENVINO_ASSERT(inputDynamicShapes[0][1].is_static(), "initSubgraph: only static channels dim is supported"); | ||
|
|
||
| const auto param = std::make_shared<ov::opset10::Parameter>(precision, inputDynamicShapes[0]); | ||
| const auto reshape_const_1 = ov::opset10::Constant::create(shapeof_subgraph_prc, {3}, {0, 0, -1}); | ||
| const auto reshape_1 = std::make_shared<ov::opset10::Reshape>(param, reshape_const_1, true); | ||
|
|
||
| const auto transpose_const_1 = ov::opset10::Constant::create(shapeof_subgraph_prc, {3}, {0, 2, 1}); | ||
| const auto transpose_1 = std::make_shared<ov::opset10::Transpose>(reshape_1, transpose_const_1); | ||
|
|
||
| const size_t channels = inputDynamicShapes[0][1].get_length(); | ||
| const size_t fc_out_channels = 512; | ||
| const auto fc_weights_1 = ngraph::builder::makeConstant(precision, ov::Shape{fc_out_channels, channels}, std::vector<float>{}, true); | ||
| const auto fc_1 = make_layer_with_bias<ov::opset10::MatMul>(transpose_1, fc_weights_1, false, true); | ||
|
|
||
| const auto transpose_const_2 = ov::opset10::Constant::create(shapeof_subgraph_prc, {3}, {0, 2, 1}); | ||
| const auto transpose_2 = std::make_shared<ov::opset10::Transpose>(fc_1, transpose_const_2); | ||
| const auto spatial_dims = ov::op::util::node_to_get_shape_value_of_indices_from_shape_source(param, {2, 3}, {}, shapeof_subgraph_prc); | ||
| const auto unchangable_dims = ov::opset10::Constant::create(shapeof_subgraph_prc, {2}, {0, 0}); | ||
| const auto reshape_const_2 = ov::op::util::make_try_fold<ov::opset10::Concat>(ov::OutputVector{unchangable_dims, spatial_dims}, 0); | ||
| const auto reshape_2 = std::make_shared<ov::opset10::Reshape>(transpose_2, reshape_const_2, true); | ||
|
|
||
| const auto conv_weights = ngraph::builder::makeConstant(precision, ov::Shape{fc_out_channels, 1, 1, 3, 3}, std::vector<float>{}, true); | ||
| const auto conv_with_bias = make_layer_with_bias<ov::opset10::GroupConvolution>(reshape_2, | ||
| conv_weights, | ||
| ov::Strides{1, 1}, | ||
| ov::CoordinateDiff{1, 1}, | ||
| ov::CoordinateDiff{1, 1}, | ||
| ov::Strides{1, 1}); | ||
| // It's necessary to force acdb layout to be sure that the reorder, which changes dims order, will be inserted | ||
| // (by default acdb layout is chosen only on >= AVX512 platforms) | ||
| const auto conv = conv_with_bias->get_input_node_shared_ptr(0); | ||
| const auto acdb_format = CPUTestUtils::cpu_memory_format_t::acdb; | ||
| conv->get_rt_info() = makeCPUInfo({acdb_format}, {acdb_format}, {}); | ||
|
|
||
| const auto dim_h = ov::op::util::node_to_get_shape_value_of_indices_from_shape_source(param, {2}, {}, shapeof_subgraph_prc); | ||
| const auto dim_w = ov::op::util::node_to_get_shape_value_of_indices_from_shape_source(param, {3}, {}, shapeof_subgraph_prc); | ||
| const auto fused_spatial_dims = ov::op::util::make_try_fold<ov::opset10::Multiply>(dim_h, dim_w); | ||
| const auto reshape_const_3 = ov::op::util::make_try_fold<ov::opset10::Concat>(ov::OutputVector{unchangable_dims, fused_spatial_dims}, 0); | ||
| const auto reshape_3 = std::make_shared<ov::opset10::Reshape>(conv_with_bias, reshape_const_3, true); | ||
| const auto transpose_const_3 = ov::opset10::Constant::create(shapeof_subgraph_prc, {3}, {0, 2, 1}); | ||
| const auto transpose_3 = std::make_shared<ov::opset10::Transpose>(reshape_3, transpose_const_3); | ||
|
|
||
| const auto fc_weights_2 = ngraph::builder::makeConstant(precision, ov::Shape{channels, fc_out_channels}, std::vector<float>{}, true); | ||
| const auto fc_2 = make_layer_with_bias<ov::opset10::MatMul>(transpose_3, fc_weights_2, false, true); | ||
| function = std::make_shared<ov::Model>(fc_2, ov::ParameterVector{param}, "MergeTransposeReorderModel"); | ||
| } | ||
|
|
||
| void validate_exec_graph() { | ||
| const size_t original_reshape_transpose_count = 3; | ||
| const size_t non_optimized_reshape_transpose_count = original_reshape_transpose_count - m_optimized_reorders_count; | ||
| CheckNumberOfNodesWithType(compiledModel, "Transpose", non_optimized_reshape_transpose_count); | ||
| CheckNumberOfNodesWithType(compiledModel, "Reshape", non_optimized_reshape_transpose_count); | ||
| size_t fake_reorder_count = 0; | ||
| for (const auto& node : compiledModel.get_runtime_model()->get_ops()) { | ||
| const auto& rtInfo = node->get_rt_info(); | ||
| auto it = rtInfo.find(ExecGraphInfoSerialization::LAYER_TYPE); | ||
| IE_ASSERT(rtInfo.end() != it); | ||
| if (it->second.as<std::string>() == "Reorder" && node->get_friendly_name().find("_fake") != std::string::npos) { | ||
| fake_reorder_count++; | ||
| } | ||
| } | ||
| ASSERT_EQ(fake_reorder_count, m_optimized_reorders_count); | ||
| } | ||
|
|
||
| private: | ||
| size_t m_optimized_reorders_count = 0; | ||
| }; | ||
|
|
||
| TEST_P(MergeTransposeReorderCPUTest, CompareWithRefs) { | ||
| run(); | ||
| validate_exec_graph(); | ||
| } | ||
|
|
||
| namespace { | ||
| std::vector<InputShape> static_shapes = { | ||
| InputShape{{}, {{1, 32, 16, 16}}}, | ||
| }; | ||
|
|
||
| INSTANTIATE_TEST_SUITE_P(smoke_MergeTransposeReorder_static, MergeTransposeReorderCPUTest, | ||
| ::testing::Combine(::testing::ValuesIn(static_shapes), | ||
| ::testing::Values(ElementType::f32), | ||
| ::testing::Values(2)), | ||
| MergeTransposeReorderCPUTest::getTestCaseName); | ||
|
|
||
| std::vector<InputShape> dynamic_shapes = { | ||
| InputShape{{-1, 32, -1, -1}, {{1, 32, 16, 16}}}, | ||
| InputShape{{-1, 32, 16, 16}, {{1, 32, 16, 16}}}, | ||
| }; | ||
|
|
||
| INSTANTIATE_TEST_SUITE_P(smoke_MergeTransposeReorder_dynamic, MergeTransposeReorderCPUTest, | ||
| ::testing::Combine(::testing::ValuesIn(dynamic_shapes), | ||
| ::testing::Values(ElementType::f32), | ||
| ::testing::Values(0)), | ||
| MergeTransposeReorderCPUTest::getTestCaseName); | ||
| } // namespace | ||
| } // namespace CPUSubgraphTestsDefinitions |
Oops, something went wrong.