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.
[GNA] Remove transposes around MatMul
- Loading branch information
1 parent
7580bd3
commit b02adb0
Showing
16 changed files
with
365 additions
and
109 deletions.
There are no files selected for viewing
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
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
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
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
125 changes: 125 additions & 0 deletions
125
inference-engine/src/gna_plugin/transformations/handle_transposes_around_matmul.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 +1,125 @@ | ||
| // Copyright (C) 2021 Intel Corporation | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
| // | ||
|
|
||
| #include "transformations/handle_transposes_around_matmul.hpp" | ||
|
|
||
| #include <numeric> | ||
|
|
||
| #include <ngraph/opsets/opset7.hpp> | ||
| #include <ngraph/pattern/op/wrap_type.hpp> | ||
| #include <ngraph/pattern/op/or.hpp> | ||
| #include <ngraph/rt_info.hpp> | ||
|
|
||
| #include "gna_plugin_log.hpp" | ||
|
|
||
| using namespace GNAPluginNS; | ||
|
|
||
| NGRAPH_RTTI_DEFINITION(HandleTransposesAroundMatMul, "HandleTransposesAroundMatMul", 0); | ||
| NGRAPH_RTTI_DEFINITION(HandleTransposeBeforeMatMul, "HandleTransposeBeforeMatMul", 0); | ||
| NGRAPH_RTTI_DEFINITION(HandleTransposeAfterMatMul, "HandleTransposeAfterMatMul", 0); | ||
|
|
||
| static void ReplaceTransposeWithReshape(std::shared_ptr<ngraph::Node> transpose_node) { | ||
| auto shape = transpose_node->get_output_shape(0); | ||
| auto reshape_const = std::make_shared<ngraph::opset7::Constant>(ngraph::element::Type_t::i64, | ||
| ngraph::Shape{shape.size()}, shape); | ||
| auto reshape_node = std::make_shared<ngraph::opset7::Reshape>(transpose_node->input_value(0), reshape_const, false); | ||
| reshape_node->set_friendly_name(transpose_node->get_friendly_name() + "/reshape"); | ||
| ngraph::copy_runtime_info(transpose_node, reshape_node); | ||
| transpose_node->output(0).replace(reshape_node->output(0)); | ||
| } | ||
|
|
||
| static void InsertTranspose(std::shared_ptr<ngraph::Node> prev_node, const std::string& base_name) { | ||
| auto consumers = prev_node->output(0).get_target_inputs(); | ||
| const auto orig_shape = prev_node->get_output_shape(0); | ||
| std::vector<size_t> transpose_ids; | ||
| for (size_t i = 0; i < orig_shape.size(); ++i) { | ||
| if (orig_shape[i] > 1) { | ||
| transpose_ids.push_back(i); | ||
| } | ||
| } | ||
| IE_ASSERT(transpose_ids.size() == 2); | ||
| std::vector<size_t> permute_order(orig_shape.size()); | ||
| std::iota(std::begin(permute_order), std::end(permute_order), 0); | ||
| std::swap(permute_order[transpose_ids[0]], permute_order[transpose_ids[1]]); | ||
|
|
||
| auto transpose_order = ngraph::opset7::Constant::create(ngraph::element::i64, ngraph::Shape{permute_order.size()}, permute_order); | ||
| auto transpose = std::make_shared<ngraph::opset7::Transpose>(prev_node, transpose_order); | ||
| transpose->set_friendly_name(base_name + "/in_transpose"); | ||
|
|
||
| auto reshapeConstAfter = std::make_shared<ngraph::opset7::Constant>(ngraph::element::Type_t::i64, | ||
| ngraph::Shape{orig_shape.size()}, orig_shape); | ||
| auto reshapeAfter = std::make_shared<ngraph::opset7::Reshape>(transpose, reshapeConstAfter, false); | ||
| reshapeAfter->set_friendly_name(base_name + "/reshape_after_transpose"); | ||
| ngraph::copy_runtime_info(prev_node, ngraph::NodeVector{transpose, reshapeAfter}); | ||
|
|
||
| for (auto input : consumers) { | ||
| input.replace_source_output(reshapeAfter); | ||
| } | ||
| } | ||
|
|
||
| static bool IsTransposeSupported(const ngraph::Shape& shape) { | ||
| auto shape_no_1 = shape; | ||
| shape_no_1.erase(std::remove(shape_no_1.begin(), shape_no_1.end(), 1), shape_no_1.end()); | ||
| if (shape_no_1.size() != 2) return false; | ||
| size_t min, max; | ||
| std::tie(min, max) = std::minmax(shape_no_1[0], shape_no_1[1]); | ||
| return min <= 8 && max % 8 == 0; | ||
| } | ||
|
|
||
| HandleTransposeBeforeMatMul::HandleTransposeBeforeMatMul() { | ||
| auto reshape = ngraph::pattern::wrap_type<ngraph::opset7::Reshape>({ngraph::pattern::any_input(), | ||
| ngraph::pattern::any_input()}, VerifyReshape()); | ||
| auto transpose = ngraph::pattern::wrap_type<ngraph::opset7::Transpose>({reshape, | ||
| ngraph::pattern::any_input()}); | ||
| auto matmul_input = std::make_shared<ngraph::pattern::op::Or>(ngraph::OutputVector{reshape, transpose}); | ||
| auto matmul1 = ngraph::pattern::wrap_type<ngraph::opset7::MatMul>({matmul_input, ngraph::pattern::any_input()}); | ||
| auto matmul2 = ngraph::pattern::wrap_type<ngraph::opset7::MatMul>({ngraph::pattern::any_input(), matmul_input}); | ||
| auto matmul = std::make_shared<ngraph::pattern::op::Or>(ngraph::OutputVector{matmul1, matmul2}); | ||
|
|
||
| ngraph::matcher_pass_callback callback = [=](ngraph::pattern::Matcher &m) { | ||
| const auto& pattern_map = m.get_pattern_value_map(); | ||
| auto transpose_it = pattern_map.find(transpose); | ||
| if (transpose_it != std::end(pattern_map)) { | ||
| ReplaceTransposeWithReshape(transpose_it->second.get_node_shared_ptr()); | ||
| } else { | ||
| auto reshape_node = pattern_map.at(reshape).get_node_shared_ptr(); | ||
| if (!IsTransposeSupported(reshape_node->get_output_shape(0))) return false; | ||
| auto matmul_it = pattern_map.find(matmul1); | ||
| auto matmul_out = matmul_it != std::end(pattern_map) ? matmul_it->second : pattern_map.at(matmul2); | ||
| InsertTranspose(reshape_node, matmul_out.get_node_shared_ptr()->get_friendly_name()); | ||
| } | ||
| return true; | ||
| }; | ||
|
|
||
| auto m = std::make_shared<ngraph::pattern::Matcher>(matmul, "HandleTransposeBeforeMatMul"); | ||
| this->register_matcher(m, callback); | ||
| } | ||
|
|
||
| HandleTransposeAfterMatMul::HandleTransposeAfterMatMul() { | ||
| auto matmul = ngraph::pattern::wrap_type<ngraph::opset7::MatMul>(); | ||
| auto fq = ngraph::pattern::wrap_type<ngraph::opset7::FakeQuantize>({matmul, ngraph::pattern::any_input(), | ||
| ngraph::pattern::any_input(), ngraph::pattern::any_input(), ngraph::pattern::any_input()}); | ||
| auto transpose_input = std::make_shared<ngraph::pattern::op::Or>(ngraph::OutputVector{matmul, fq}); | ||
| auto transpose = ngraph::pattern::wrap_type<ngraph::opset7::Transpose>({transpose_input, ngraph::pattern::any_input()}); | ||
| auto reshape_input = std::make_shared<ngraph::pattern::op::Or>(ngraph::OutputVector{transpose_input, transpose}); | ||
| auto reshape = ngraph::pattern::wrap_type<ngraph::opset7::Reshape>({reshape_input, | ||
| ngraph::pattern::any_input()}, VerifyReshape()); | ||
|
|
||
| ngraph::matcher_pass_callback callback = [=](ngraph::pattern::Matcher &m) { | ||
| const auto& pattern_map = m.get_pattern_value_map(); | ||
| auto transpose_it = pattern_map.find(transpose); | ||
| if (transpose_it != std::end(pattern_map)) { | ||
| ReplaceTransposeWithReshape(transpose_it->second.get_node_shared_ptr()); | ||
| } else { | ||
| auto reshape_node = pattern_map.at(reshape).get_node_shared_ptr(); | ||
| if (!IsTransposeSupported(reshape_node->get_input_shape(0))) return false; | ||
| auto matmul_node = pattern_map.at(matmul).get_node_shared_ptr(); | ||
| InsertTranspose(matmul_node, matmul_node->get_friendly_name()); | ||
| } | ||
| return true; | ||
| }; | ||
|
|
||
| auto m = std::make_shared<ngraph::pattern::Matcher>(reshape, "HandleTransposeAfterMatMul"); | ||
| this->register_matcher(m, callback); | ||
| } |
74 changes: 74 additions & 0 deletions
74
inference-engine/src/gna_plugin/transformations/handle_transposes_around_matmul.hpp
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 +1,74 @@ | ||
| // Copyright (C) 2021 Intel Corporation | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
| // | ||
|
|
||
| #pragma once | ||
|
|
||
| #include <ngraph/pass/graph_rewrite.hpp> | ||
|
|
||
| namespace GNAPluginNS { | ||
|
|
||
| struct VerifyReshape { | ||
| bool operator()(const ngraph::Output<ngraph::Node>& reshape_out) const { | ||
| auto in_shape = reshape_out.get_node_shared_ptr()->get_input_shape(0); | ||
| auto out_shape = reshape_out.get_node_shared_ptr()->get_output_shape(0); | ||
|
|
||
| // Check if Reshape changes the final 2d shape of Affine primitive | ||
| in_shape.erase(std::remove(in_shape.begin(), in_shape.end(), 1), in_shape.end()); | ||
| out_shape.erase(std::remove(out_shape.begin(), out_shape.end(), 1), out_shape.end()); | ||
| return in_shape != out_shape; | ||
| } | ||
| }; | ||
|
|
||
| /** | ||
| * @brief Inserts Transpose before MatMul or removes it (if it exists) if there is Reshape | ||
| * before MatMul which changes the batch size: | ||
| * [1, A*B] [1, A*B] | ||
| * | | | ||
| * Reshape Reshape | ||
| * | | | ||
| * [1, A, 1, B] [1, A, 1, B] | ||
| * | | | ||
| * | Transpose | ||
| * | -> | | ||
| * | <- [1, B, 1, A] | ||
| * | | | ||
| * MatMul MatMul | ||
| */ | ||
| class HandleTransposeBeforeMatMul : public ngraph::pass::MatcherPass { | ||
| public: | ||
| NGRAPH_RTTI_DECLARATION; | ||
| HandleTransposeBeforeMatMul(); | ||
| }; | ||
|
|
||
| /** | ||
| * @brief Inserts Transpose after MatMul or removes it (if it exists) if there is Reshape | ||
| * after MatMul which changes the batch size: | ||
| * MatMul MatMul | ||
| * | | | ||
| * [1, A, 1, B] [1, A, 1, B] | ||
| * | | | ||
| * | Transpose | ||
| * | -> | | ||
| * | <- [1, B, 1, A] | ||
| * | | | ||
| * Reshape Reshape | ||
| * | | | ||
| * [1, A*B] [1, A*B] | ||
| */ | ||
| class HandleTransposeAfterMatMul : public ngraph::pass::MatcherPass { | ||
| public: | ||
| NGRAPH_RTTI_DECLARATION; | ||
| HandleTransposeAfterMatMul(); | ||
| }; | ||
|
|
||
| class HandleTransposesAroundMatMul: public ngraph::pass::GraphRewrite { | ||
| public: | ||
| NGRAPH_RTTI_DECLARATION; | ||
| HandleTransposesAroundMatMul() { | ||
| add_matcher<HandleTransposeBeforeMatMul>(); | ||
| add_matcher<HandleTransposeAfterMatMul>(); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace GNAPluginNS |
67 changes: 0 additions & 67 deletions
67
inference-engine/src/gna_plugin/transformations/insert_transpose_before_matmul.cpp
This file was deleted.
Oops, something went wrong.
Oops, something went wrong.