From 7fa9bbf6fc5a94e312ffc0ef4748175fc8dbddd1 Mon Sep 17 00:00:00 2001 From: Vladislav Golubev Date: Mon, 27 Sep 2021 11:50:17 +0300 Subject: [PATCH] [LPT] Avoid using std::shared_ptr when creating a node (#7357) * [LPT] Avoid using std::shared_ptr when creating a node * [LPT] removed unused files * [LPT] D2STransformation: transform & isPrecisionPreserved methods are moved to base class * [LPT] Revert redundant changes --- .../include/low_precision/concat.hpp | 2 - .../include/low_precision/depth_to_space.hpp | 2 - .../transparent_base_transformation.hpp | 1 + .../low_precision_transformations/src/add.cpp | 4 +- .../operation_precision_restriction.cpp | 19 --- .../src/concat.cpp | 34 +---- .../src/convert.cpp | 2 +- .../src/convolution_backprop_data.cpp | 2 +- .../create_precisions_dependent_attribute.cpp | 22 ---- .../src/depth_to_space.cpp | 15 --- .../src/fake_quantize.cpp | 18 +-- .../src/fold_convert.cpp | 2 +- .../src/fuse_convert.cpp | 18 +-- .../src/fuse_fake_quantize.cpp | 15 ++- .../src/fuse_multiply_to_fake_quantize.cpp | 12 +- .../src/fuse_subtract_to_fake_quantize.cpp | 12 +- .../src/mat_mul.cpp | 2 +- .../src/multiply.cpp | 10 +- .../low_precision_transformations/src/mvn.cpp | 2 +- .../src/network_helper.cpp | 124 ++++++++---------- .../pull_reshape_through_dequantization.cpp | 16 +-- .../pull_transpose_through_dequantization.cpp | 14 +- .../src/reshape.cpp | 2 +- .../rt_info/intervals_alignment_attribute.cpp | 6 +- .../src/rt_info/shared_value_attribute.cpp | 16 --- .../src/squeeze.cpp | 2 +- .../src/strided_slice.cpp | 6 +- .../src/subtract.cpp | 14 +- .../src/transparent_base_transformation.cpp | 29 ++-- .../src/unsqueeze.cpp | 2 +- 30 files changed, 158 insertions(+), 267 deletions(-) delete mode 100644 inference-engine/src/low_precision_transformations/src/common/operation_precision_restriction.cpp delete mode 100644 inference-engine/src/low_precision_transformations/src/create_precisions_dependent_attribute.cpp delete mode 100644 inference-engine/src/low_precision_transformations/src/rt_info/shared_value_attribute.cpp diff --git a/inference-engine/src/low_precision_transformations/include/low_precision/concat.hpp b/inference-engine/src/low_precision_transformations/include/low_precision/concat.hpp index 65cb9694eb8..c1f752972ad 100644 --- a/inference-engine/src/low_precision_transformations/include/low_precision/concat.hpp +++ b/inference-engine/src/low_precision_transformations/include/low_precision/concat.hpp @@ -39,8 +39,6 @@ protected: NodeVector& convertNodes, NodeVector& subtractNodes, NodeVector& multiplyNodes) const; - - std::shared_ptr concatenateDeqNodes(NodeVector& nodes) const; }; } // namespace low_precision diff --git a/inference-engine/src/low_precision_transformations/include/low_precision/depth_to_space.hpp b/inference-engine/src/low_precision_transformations/include/low_precision/depth_to_space.hpp index b02ead7321b..5a199454eb6 100644 --- a/inference-engine/src/low_precision_transformations/include/low_precision/depth_to_space.hpp +++ b/inference-engine/src/low_precision_transformations/include/low_precision/depth_to_space.hpp @@ -14,8 +14,6 @@ class LP_TRANSFORMATIONS_API DepthToSpaceTransformation : public TransparentBase public: NGRAPH_RTTI_DECLARATION; DepthToSpaceTransformation(const Params& params = Params()); - bool transform(TransformationContext &context, ngraph::pattern::Matcher &m) override; - bool isPrecisionPreserved(std::shared_ptr layer) const noexcept override; bool canBeTransformed(const TransformationContext& context, std::shared_ptr layer) const override; }; diff --git a/inference-engine/src/low_precision_transformations/include/low_precision/transparent_base_transformation.hpp b/inference-engine/src/low_precision_transformations/include/low_precision/transparent_base_transformation.hpp index 05b0dbebc01..d1f87f92f86 100644 --- a/inference-engine/src/low_precision_transformations/include/low_precision/transparent_base_transformation.hpp +++ b/inference-engine/src/low_precision_transformations/include/low_precision/transparent_base_transformation.hpp @@ -18,6 +18,7 @@ public: ~TransparentBaseTransformation() override {}; bool transform(TransformationContext& context, ngraph::pattern::Matcher &m) override; bool canBeTransformed(const TransformationContext& context, std::shared_ptr layer) const override; + bool isPrecisionPreserved(std::shared_ptr layer) const noexcept override; }; } // namespace low_precision diff --git a/inference-engine/src/low_precision_transformations/src/add.cpp b/inference-engine/src/low_precision_transformations/src/add.cpp index da720928ff7..e4068c74a51 100644 --- a/inference-engine/src/low_precision_transformations/src/add.cpp +++ b/inference-engine/src/low_precision_transformations/src/add.cpp @@ -176,13 +176,13 @@ bool AddTransformation::transform(TransformationContext& context, ngraph::patter // after : Y = SC2 * ( SC1' * (X1 - SH1') + X2 ) , where : // SC1' = SC1 / SC2 // SH1' = SH1 + SC2 * SH2 / SC1 - std::shared_ptr newSubtractFullPathValues = fold( + auto newSubtractFullPathValues = fold( subtractFullPathValues, fold( fold(subtractEmptyPathValues, multiplyEmptyPathValues), multiplyFullPathValues)); - std::shared_ptr newMultiplyFullPathValues = fold(multiplyFullPathValues, multiplyEmptyPathValues); + auto newMultiplyFullPathValues = fold(multiplyFullPathValues, multiplyEmptyPathValues); if (NetworkHelper::isZeroConst(newSubtractFullPathValues)) { newSubtractFullPathValues = nullptr; diff --git a/inference-engine/src/low_precision_transformations/src/common/operation_precision_restriction.cpp b/inference-engine/src/low_precision_transformations/src/common/operation_precision_restriction.cpp deleted file mode 100644 index 0ec085d7245..00000000000 --- a/inference-engine/src/low_precision_transformations/src/common/operation_precision_restriction.cpp +++ /dev/null @@ -1,19 +0,0 @@ -// Copyright (C) 2021 Intel Corporation -// SPDX-License-Identifier: Apache-2.0 -// - -#include "low_precision/common/operation_precision_restriction.hpp" - -#include -#include -#include -#include - -#include -#include -#include -#include "low_precision/network_helper.hpp" -#include "low_precision/rt_info/precisions_attribute.hpp" - -using namespace ngraph; - diff --git a/inference-engine/src/low_precision_transformations/src/concat.cpp b/inference-engine/src/low_precision_transformations/src/concat.cpp index 8df69b6fb21..19d4f8daa37 100644 --- a/inference-engine/src/low_precision_transformations/src/concat.cpp +++ b/inference-engine/src/low_precision_transformations/src/concat.cpp @@ -70,20 +70,11 @@ bool ConcatTransformation::transform(TransformationContext& context, ngraph::pat } } - auto broadcastElementWiseConst = []( - // FakeQuantize constant shape must be broadcastable to the shape on data. - std::shared_ptr operation, - const ngraph::Shape targetShape) -> std::shared_ptr { - auto targetShapeConst = std::make_shared( - element::i64, ngraph::Shape{ targetShape.size() }, - targetShape); - - auto broadcast = ngraph::pass::low_precision::fold( - operation, - targetShapeConst, - ngraph::op::AutoBroadcastType::NUMPY); - - return broadcast; + // FakeQuantize constant shape must be broadcastable to the shape on data. + auto broadcastElementWiseConst = [](std::shared_ptr operation, const Shape targetShape) { + auto targetShapeConst = std::make_shared(element::i64, Shape{ targetShape.size() }, targetShape); + auto broadcast = fold(operation, targetShapeConst); + return broadcast; }; bool someDqInLowPrecision = std::any_of( @@ -247,15 +238,8 @@ void ConcatTransformation::fillDequantizationNodes( // FakeQuantize constant shape must be broadcastable to the shape on data. std::shared_ptr operation, const ngraph::Shape targetShape) -> std::shared_ptr { - auto targetShapeConst = std::make_shared( - element::i64, ngraph::Shape{ targetShape.size() }, - targetShape); - - auto broadcast = ngraph::pass::low_precision::fold( - operation, - targetShapeConst, - ngraph::op::AutoBroadcastType::NUMPY); - + auto targetShapeConst = opset1::Constant::create(element::i64, ngraph::Shape{ targetShape.size() }, targetShape); + auto broadcast = fold(operation, targetShapeConst); return broadcast; }; @@ -308,10 +292,6 @@ void ConcatTransformation::fillDequantizationNodes( } } -std::shared_ptr ConcatTransformation::concatenateDeqNodes(NodeVector& nodes) const { - return nodes.size() == 1ul ? nodes[0] : fold(nodes, 1); -} - bool ConcatTransformation::isHandled(const TransformationContext& context, const std::vector>& quantizationOperations) { for (const std::shared_ptr& quantizationLayer : quantizationOperations) { if (context.quantizedFakeQuantizeNames.find(quantizationLayer->get_friendly_name()) != context.quantizedFakeQuantizeNames.end()) { diff --git a/inference-engine/src/low_precision_transformations/src/convert.cpp b/inference-engine/src/low_precision_transformations/src/convert.cpp index fbc64e0db62..82c6982b59a 100644 --- a/inference-engine/src/low_precision_transformations/src/convert.cpp +++ b/inference-engine/src/low_precision_transformations/src/convert.cpp @@ -49,7 +49,7 @@ bool ConvertTransformation::transform(TransformationContext& context, ngraph::pa const ngraph::element::Type precisionBefore = convert->get_input_element_type(0); std::shared_ptr subtract = std::make_shared>( - convert->get_input_node_shared_ptr(0), + convert->input_value(0), std::make_shared(precisionBefore, Shape{}, std::vector({ 0 }))); NetworkHelper::setOutDataPrecision(subtract, convert->get_output_element_type(0)); diff --git a/inference-engine/src/low_precision_transformations/src/convolution_backprop_data.cpp b/inference-engine/src/low_precision_transformations/src/convolution_backprop_data.cpp index 7802a498a0e..cec1b73b9a0 100644 --- a/inference-engine/src/low_precision_transformations/src/convolution_backprop_data.cpp +++ b/inference-engine/src/low_precision_transformations/src/convolution_backprop_data.cpp @@ -181,7 +181,7 @@ bool ConvolutionBackpropDataTransformation::transform(TransformationContext &con zeroPointShape[1] = static_cast(weightsPShape[1].get_length()); auto zeroPointConstant = fold( - subtractFromWeights->get_input_node_shared_ptr(1), + subtractFromWeights->input_value(1), std::make_shared(element::i32, Shape{zeroPointShape.size()}, zeroPointShape)); replace_node(subtractFromWeights->get_input_node_shared_ptr(1), zeroPointConstant); } diff --git a/inference-engine/src/low_precision_transformations/src/create_precisions_dependent_attribute.cpp b/inference-engine/src/low_precision_transformations/src/create_precisions_dependent_attribute.cpp deleted file mode 100644 index 7ddd060b06d..00000000000 --- a/inference-engine/src/low_precision_transformations/src/create_precisions_dependent_attribute.cpp +++ /dev/null @@ -1,22 +0,0 @@ -// Copyright (C) 2021 Intel Corporation -// SPDX-License-Identifier: Apache-2.0 -// - -#include "low_precision/create_precisions_dependent_attribute.hpp" - -#include -#include -#include -#include -#include -#include - -#include -#include -#include -#include "low_precision/rt_info/precisions_attribute.hpp" -#include "low_precision/rt_info/precision_preserved_attribute.hpp" -#include "low_precision/network_helper.hpp" - -using namespace ngraph; -using namespace ngraph::pass::low_precision; diff --git a/inference-engine/src/low_precision_transformations/src/depth_to_space.cpp b/inference-engine/src/low_precision_transformations/src/depth_to_space.cpp index b5de8dd661c..4339ad33503 100644 --- a/inference-engine/src/low_precision_transformations/src/depth_to_space.cpp +++ b/inference-engine/src/low_precision_transformations/src/depth_to_space.cpp @@ -29,21 +29,6 @@ DepthToSpaceTransformation::DepthToSpaceTransformation(const Params& params) : T this->register_matcher(m, callback); } -bool DepthToSpaceTransformation::transform(TransformationContext &context, ngraph::pattern::Matcher &m) { - std::shared_ptr depthToSpace = m.get_match_root(); - if (!canBeTransformed(context, depthToSpace)) { - return false; - } - - depthToSpace = NetworkHelper::separateInStandaloneBranch(depthToSpace); - moveDequantizationAfter(context, depthToSpace, NetworkHelper::getDequantization(depthToSpace), true); - return true; -} - -bool DepthToSpaceTransformation::isPrecisionPreserved(std::shared_ptr layer) const noexcept { - return true; -} - bool DepthToSpaceTransformation::canBeTransformed(const TransformationContext& context, std::shared_ptr layer) const { if (!LayerTransformation::canBeTransformed(context, layer)) { return false; diff --git a/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp index b731ed22a63..deb7ec1ea28 100644 --- a/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp +++ b/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp @@ -67,7 +67,7 @@ static std::shared_ptr updateShape(std::shared_ptr constantOp, const return constantOp; } -static std::shared_ptr getData(const std::shared_ptr& eltwise) { +static std::shared_ptr getDataNode(const std::shared_ptr& eltwise) { if (!ov::is_type(eltwise->get_input_node_shared_ptr(0))) { return eltwise->get_input_node_shared_ptr(0); } @@ -123,7 +123,7 @@ bool FakeQuantizeTransformation::checkElementwise(const std::shared_ptr& e } } - return fq::getData(eltwise) != nullptr; + return fq::getDataNode(eltwise) != nullptr; } std::shared_ptr FakeQuantizeTransformation::fuseElementwise( @@ -132,8 +132,8 @@ std::shared_ptr FakeQuantizeTransformation::fuseElementwis const std::shared_ptr& fakeQuantize) const { const std::shared_ptr eltwise = fakeQuantize->get_input_node_shared_ptr(0); - std::shared_ptr inputLowConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(1), deqPrecision); - std::shared_ptr inputHighConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(2), deqPrecision); + std::shared_ptr inputLowConst_f32 = foldConvert(fakeQuantize->input_value(1), deqPrecision); + std::shared_ptr inputHighConst_f32 = foldConvert(fakeQuantize->input_value(2), deqPrecision); std::shared_ptr constant = fq::getConstant(eltwise); if (ov::is_type(eltwise) && checkElementwise(eltwise)) { @@ -166,10 +166,10 @@ std::shared_ptr FakeQuantizeTransformation::fuseElementwis inputLowConst_f32 = fq::updateShape(fold(inputLowConst_f32, value), fakeQuantize->get_output_partial_shape(0)); inputHighConst_f32 = fq::updateShape(fold(inputHighConst_f32, value), fakeQuantize->get_output_partial_shape(0)); } else if (ov::is_type(eltwise) && checkElementwise(eltwise)) { - if (ov::is_type(fq::getData(eltwise)) || - ov::is_type(fq::getData(eltwise)) || - ov::is_type(fq::getData(eltwise)) || - ov::is_type(fq::getData(eltwise))) { + if (ov::is_type(fq::getDataNode(eltwise)) || + ov::is_type(fq::getDataNode(eltwise)) || + ov::is_type(fq::getDataNode(eltwise)) || + ov::is_type(fq::getDataNode(eltwise))) { return nullptr; } @@ -189,7 +189,7 @@ std::shared_ptr FakeQuantizeTransformation::fuseElementwis return nullptr; } - const auto data = fq::getData(eltwise); + const auto data = fq::getDataNode(eltwise); const size_t outputIdx = NetworkHelper::getParentOutputIndex(data, eltwise); const auto newFakeQuantize = ov::as_type_ptr(fakeQuantize->clone_with_new_inputs({ diff --git a/inference-engine/src/low_precision_transformations/src/fold_convert.cpp b/inference-engine/src/low_precision_transformations/src/fold_convert.cpp index f7a3255df49..d17ac6e1a82 100644 --- a/inference-engine/src/low_precision_transformations/src/fold_convert.cpp +++ b/inference-engine/src/low_precision_transformations/src/fold_convert.cpp @@ -42,7 +42,7 @@ bool FoldConvertTransformation::transform(TransformationContext& context, ngraph return; } - const auto resultConstant = ngraph::pass::low_precision::foldConvert(convert->get_input_node_shared_ptr(0), convert->output(0).get_element_type()); + const auto resultConstant = ngraph::pass::low_precision::foldConvert(convert->input_value(0), convert->get_output_element_type(0)); assert(ov::is_type(resultConstant)); replace_node(convert, resultConstant); diff --git a/inference-engine/src/low_precision_transformations/src/fuse_convert.cpp b/inference-engine/src/low_precision_transformations/src/fuse_convert.cpp index 3533bb66213..07f56f8b572 100644 --- a/inference-engine/src/low_precision_transformations/src/fuse_convert.cpp +++ b/inference-engine/src/low_precision_transformations/src/fuse_convert.cpp @@ -47,8 +47,8 @@ std::shared_ptr removeConvertIfPossibleForSubtract( if (NetworkHelper::checkConstantValuePrecision(precisionBeforeConvert, subtract->get_input_node_shared_ptr(1))) { newSubtract = std::make_shared>( std::vector{ element::f32, element::f32 }, std::vector{}, - ngraph::op::TemporaryReplaceOutputType(convert->get_input_source_output(0), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(subtract->get_input_node_shared_ptr(1), element::f32).get()); + ngraph::op::TemporaryReplaceOutputType(convert->input_value(0), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(subtract->input_value(1), element::f32).get()); NetworkHelper::setOutDataPrecisionForTypeRelaxed(newSubtract, subtract->get_output_element_type(0)); replace_node(subtract, newSubtract); } @@ -63,11 +63,11 @@ bool FuseConvertTransformation::transform(TransformationContext& context, ngraph } const auto convert = ov::as_type_ptr(op->get_input_node_shared_ptr(0)); - std::shared_ptr parent = convert->get_input_node_shared_ptr(0); + auto parent = convert->input_value(0); - if (ov::is_type(parent)) { + if (ov::is_type(parent.get_node_shared_ptr())) { auto convertedConstant = foldConvert(parent, convert->get_convert_element_type()); - NetworkHelper::copyInfo(parent, convertedConstant); + NetworkHelper::copyInfo(parent.get_node_shared_ptr(), convertedConstant); replace_node(convert, convertedConstant); } else { std::shared_ptr newOp; @@ -77,15 +77,15 @@ bool FuseConvertTransformation::transform(TransformationContext& context, ngraph } else if (ov::is_type(op)) { newOp = std::make_shared>( std::vector{ element::f32, element::f32 }, std::vector{}, - ngraph::op::TemporaryReplaceOutputType(convert->get_input_source_output(0), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(op->get_input_node_shared_ptr(1), element::f32).get()); + ngraph::op::TemporaryReplaceOutputType(convert->input_value(0), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(op->input_value(1), element::f32).get()); NetworkHelper::setOutDataPrecisionForTypeRelaxed(newOp, op->get_output_element_type(0)); replace_node(op, newOp); } else if (ov::is_type(op)) { newOp = std::make_shared>( std::vector{ element::f32, element::f32 }, std::vector{}, - ngraph::op::TemporaryReplaceOutputType(convert->get_input_source_output(0), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(op->get_input_node_shared_ptr(1), element::f32).get()); + ngraph::op::TemporaryReplaceOutputType(convert->input_value(0), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(op->input_value(1), element::f32).get()); NetworkHelper::setOutDataPrecisionForTypeRelaxed(newOp, op->get_output_element_type(0)); replace_node(op, newOp); } diff --git a/inference-engine/src/low_precision_transformations/src/fuse_fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/fuse_fake_quantize.cpp index 0c897f468a5..bb59172ff3a 100644 --- a/inference-engine/src/low_precision_transformations/src/fuse_fake_quantize.cpp +++ b/inference-engine/src/low_precision_transformations/src/fuse_fake_quantize.cpp @@ -54,7 +54,7 @@ std::shared_ptr updateShape(std::shared_ptr op, const PartialShape& return op; } -std::shared_ptr getData(const std::shared_ptr& eltwise) { +std::shared_ptr getDataNode(const std::shared_ptr& eltwise) { if (!ov::is_type(eltwise->get_input_node_shared_ptr(0))) { return eltwise->get_input_node_shared_ptr(0); } @@ -108,7 +108,7 @@ bool eltwiseWithConstant(const std::shared_ptr& eltwise) { } } - return getData(eltwise) != nullptr; + return getDataNode(eltwise) != nullptr; } } // namespace fuse_fq @@ -144,8 +144,8 @@ std::shared_ptr FuseFakeQuantizeTransformation::handle( inputLowConst = fuse_fq::updateShape(fold(inputLowConst, value), fakeQuantize->get_output_partial_shape(0)); inputHightConst = fuse_fq::updateShape(fold(inputHightConst, value), fakeQuantize->get_output_partial_shape(0)); } else if (ov::is_type(eltwise) && fuse_fq::eltwiseWithConstant(eltwise)) { - if (ov::is_type(fuse_fq::getData(eltwise)) || - ov::is_type(fuse_fq::getData(eltwise))) { + if (ov::is_type(fuse_fq::getDataNode(eltwise)) || + ov::is_type(fuse_fq::getDataNode(eltwise))) { return nullptr; } @@ -157,15 +157,18 @@ std::shared_ptr FuseFakeQuantizeTransformation::handle( inputHightConst = fuse_fq::updateShape(fold(inputHightConst, value), fakeQuantize->get_output_partial_shape(0)); } else if (ov::is_type(eltwise)) { // issue #40611 - if ((eltwise->input(0).get_element_type() == element::i32) && (eltwise->output(0).get_element_type() == element::f32)) { + if ((eltwise->get_input_element_type(0) == element::i32) && (eltwise->get_output_element_type(0) == element::f32)) { return nullptr; } } else { return nullptr; } + const auto data = fuse_fq::getDataNode(eltwise); + const size_t outputIdx = NetworkHelper::getParentOutputIndex(data, eltwise); + std::shared_ptr newFakeQuantize = ov::as_type_ptr(fakeQuantize->clone_with_new_inputs({ - fuse_fq::getData(eltwise), + data->output(outputIdx), inputLowConst, inputHightConst, fakeQuantize->input_value(3), diff --git a/inference-engine/src/low_precision_transformations/src/fuse_multiply_to_fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/fuse_multiply_to_fake_quantize.cpp index 3cab73ba3e9..9d48b5a23f3 100644 --- a/inference-engine/src/low_precision_transformations/src/fuse_multiply_to_fake_quantize.cpp +++ b/inference-engine/src/low_precision_transformations/src/fuse_multiply_to_fake_quantize.cpp @@ -46,9 +46,12 @@ bool FuseMultiplyToFakeQuantizeTransformation::transform(TransformationContext& } const auto multiplyConstant = multiply->get_input_node_shared_ptr(1); + if (!ov::is_type(multiplyConstant)) { + return false; + } - auto outputLowConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(3), deqPrecision); - auto outputHighConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(4), deqPrecision); + auto outputLowConst_f32 = foldConvert(fakeQuantize->input_value(3), deqPrecision); + auto outputHighConst_f32 = foldConvert(fakeQuantize->input_value(4), deqPrecision); const auto value = multiplyConstant->get_output_element_type(0) == element::f32 ? multiplyConstant : @@ -57,9 +60,6 @@ bool FuseMultiplyToFakeQuantizeTransformation::transform(TransformationContext& outputLowConst_f32 = fold(outputLowConst_f32, value); outputHighConst_f32 = fold(outputHighConst_f32, value); - const auto fakeQuantizeParent = fakeQuantize->get_input_node_shared_ptr(0); - const size_t parentIndex = NetworkHelper::getParentOutputIndex(fakeQuantizeParent, fakeQuantize); - const auto inputLow = foldConvert(fakeQuantize->input_value(1), deqPrecision); const auto inputHigh = foldConvert(fakeQuantize->input_value(2), deqPrecision); NetworkHelper::copyInfo(fakeQuantize->get_input_node_shared_ptr(1), inputLow); @@ -69,7 +69,7 @@ bool FuseMultiplyToFakeQuantizeTransformation::transform(TransformationContext& auto newFakeQuantize = std::make_shared>( opset1::FakeQuantize( - fakeQuantizeParent->output(parentIndex), + fakeQuantize->input_value(0), inputLow, inputHigh, outputLowConst_f32, diff --git a/inference-engine/src/low_precision_transformations/src/fuse_subtract_to_fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/fuse_subtract_to_fake_quantize.cpp index edd8ee35cb4..d4e1aafee4c 100644 --- a/inference-engine/src/low_precision_transformations/src/fuse_subtract_to_fake_quantize.cpp +++ b/inference-engine/src/low_precision_transformations/src/fuse_subtract_to_fake_quantize.cpp @@ -45,9 +45,12 @@ bool FuseSubtractToFakeQuantizeTransformation::transform(TransformationContext& } const auto subtractConstant = subtract->get_input_node_shared_ptr(1); + if (!ov::is_type(subtractConstant)) { + return false; + } - auto outputLowConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(3), deqPrecision); - auto outputHighConst_f32 = foldConvert(fakeQuantize->get_input_node_shared_ptr(4), deqPrecision); + auto outputLowConst_f32 = foldConvert(fakeQuantize->input_value(3), deqPrecision); + auto outputHighConst_f32 = foldConvert(fakeQuantize->input_value(4), deqPrecision); const auto value = subtractConstant->get_output_element_type(0) == element::f32 ? subtractConstant : @@ -56,9 +59,6 @@ bool FuseSubtractToFakeQuantizeTransformation::transform(TransformationContext& outputLowConst_f32 = fold(outputLowConst_f32, value); outputHighConst_f32 = fold(outputHighConst_f32, value); - const auto fakeQuantizeParent = fakeQuantize->get_input_node_shared_ptr(0); - const size_t parentIndex = NetworkHelper::getParentOutputIndex(fakeQuantizeParent, fakeQuantize); - const auto inputLow = foldConvert(fakeQuantize->input_value(1), deqPrecision); const auto inputHigh = foldConvert(fakeQuantize->input_value(2), deqPrecision); NetworkHelper::copyInfo(fakeQuantize->get_input_node_shared_ptr(1), inputLow); @@ -68,7 +68,7 @@ bool FuseSubtractToFakeQuantizeTransformation::transform(TransformationContext& auto newFakeQuantize = std::make_shared>( opset1::FakeQuantize( - fakeQuantizeParent->output(parentIndex), + fakeQuantize->input_value(0), inputLow, inputHigh, outputLowConst_f32, diff --git a/inference-engine/src/low_precision_transformations/src/mat_mul.cpp b/inference-engine/src/low_precision_transformations/src/mat_mul.cpp index 6b0a1cdfeca..9d9fd8ecce0 100644 --- a/inference-engine/src/low_precision_transformations/src/mat_mul.cpp +++ b/inference-engine/src/low_precision_transformations/src/mat_mul.cpp @@ -109,7 +109,7 @@ bool MatMulTransformation::transform(TransformationContext &context, ngraph::pat // multiply by weights: [1, ..., 1, Y] x [Y, Z] => [1, ..., 1, Z] const auto newSubConst = NetworkHelper::toScalarIfPossible(fold( broadcastedConst, - foldConvert(newMatMul->get_input_node_shared_ptr(1), newMatMul->get_element_type()), + foldConvert(newMatMul->input_value(1), newMatMul->get_element_type()), newMatMul->get_transpose_a(), newMatMul->get_transpose_b())); diff --git a/inference-engine/src/low_precision_transformations/src/multiply.cpp b/inference-engine/src/low_precision_transformations/src/multiply.cpp index ea713ee6b27..68792d2d3c7 100644 --- a/inference-engine/src/low_precision_transformations/src/multiply.cpp +++ b/inference-engine/src/low_precision_transformations/src/multiply.cpp @@ -77,10 +77,10 @@ bool MultiplyTransformation::transform(TransformationContext& context, ngraph::p return false; } - auto multiplyParent = multiply->get_input_source_output(multiplyBranch.first); - auto constParent = multiply->get_input_source_output(multiplyBranch.first == 0 ? 1 : 0); - auto multiplyParentParent = multiplyParent.get_node_shared_ptr()->get_input_source_output(multiplyBranch.second); - auto multiplyParentConst = multiplyParent.get_node_shared_ptr()->get_input_source_output(multiplyBranch.second == 0 ? 1 : 0); + auto multiplyParent = multiply->input_value(multiplyBranch.first); + auto constParent = multiply->input_value(multiplyBranch.first == 0 ? 1 : 0); + auto multiplyParentParent = multiplyParent.get_node_shared_ptr()->input_value(multiplyBranch.second); + auto multiplyParentConst = multiplyParent.get_node_shared_ptr()->input_value(multiplyBranch.second == 0 ? 1 : 0); newMultiply = std::make_shared>( std::vector{ element::f32, element::f32 }, @@ -127,7 +127,7 @@ bool MultiplyTransformation::transform(TransformationContext& context, ngraph::p // before: Y = (SC1 * (X1 - SH1)) * (SC2 * X2) // after : Y = (SC1' * (X1 - SH1)) * (X2) , where : // SC1' = SC1 * SC2 - std::shared_ptr newMultiplyValuesFullPath = fold(multiplyValuesEmptyPath, multiplyValuesFullPath); + auto newMultiplyValuesFullPath = fold(multiplyValuesEmptyPath, multiplyValuesFullPath); OutputVector inputs{ {}, {} }; inputs[emptyPathIndex] = dequantizationEmptyPath.data; inputs[fullPathIndex] = std::make_shared( diff --git a/inference-engine/src/low_precision_transformations/src/mvn.cpp b/inference-engine/src/low_precision_transformations/src/mvn.cpp index 34f43655604..72ca086ced5 100644 --- a/inference-engine/src/low_precision_transformations/src/mvn.cpp +++ b/inference-engine/src/low_precision_transformations/src/mvn.cpp @@ -149,7 +149,7 @@ bool MVNTransformation::transform(TransformationContext &context, ngraph::patter if (ov::is_type(mvn)) { newMVN = mvn->copy_with_new_inputs({dequantization.data}); } else { - newMVN = mvn->copy_with_new_inputs({dequantization.data, mvn->get_input_node_shared_ptr(1)}); + newMVN = mvn->copy_with_new_inputs({dequantization.data, mvn->input_value(1)}); } NetworkHelper::setOutDataPrecisionForTypeRelaxed(newMVN, deqPrecision); NetworkHelper::copyInfo(mvn, newMVN); diff --git a/inference-engine/src/low_precision_transformations/src/network_helper.cpp b/inference-engine/src/low_precision_transformations/src/network_helper.cpp index 6d38a1ba5f8..e026fe015c3 100644 --- a/inference-engine/src/low_precision_transformations/src/network_helper.cpp +++ b/inference-engine/src/low_precision_transformations/src/network_helper.cpp @@ -233,10 +233,10 @@ std::shared_ptr NetworkHelper::swapMultiplyAndAdd(std::shared_ptrget_input_source_output(multiplyInputBranch); - auto a = multiply->get_input_node_shared_ptr(multiplyInputBranch == 0 ? 1 : 0); - auto b = addAfterMultiply->get_input_node_shared_ptr(multiplyBranch == 0 ? 1 : 0); - std::shared_ptr bDivA; + const auto x = multiply->input_value(multiplyInputBranch); + auto a = as_type_ptr(multiply->get_input_node_shared_ptr(multiplyInputBranch == 0 ? 1 : 0)); + auto b = as_type_ptr(addAfterMultiply->get_input_node_shared_ptr(multiplyBranch == 0 ? 1 : 0)); + std::shared_ptr bDivA; const auto aPShape = a->get_output_partial_shape(0); assert(aPShape.is_static()); @@ -248,8 +248,8 @@ std::shared_ptr NetworkHelper::swapMultiplyAndAdd(std::shared_ptr bValues = ov::as_type_ptr(b)->cast_vector(); - const std::vector aValues = ov::as_type_ptr(a)->cast_vector(); + const std::vector bValues = b->cast_vector(); + const std::vector aValues = a->cast_vector(); const bool aBroadcasted = bValues.size() > aValues.size(); const bool bBroadcasted = bValues.size() < aValues.size(); std::vector bDivAValues(aBroadcasted ? bValues.size() : aValues.size()); @@ -271,16 +271,16 @@ std::shared_ptr NetworkHelper::swapMultiplyAndAdd(std::shared_ptr(b, a); + b = as_type_ptr(foldConvert(b->output(0), element::f32)); + a = as_type_ptr(foldConvert(a->output(0), element::f32)); + bDivA = as_type_ptr(fold(b->output(0), a->output(0))); // TODO: issue #49868 - bDivA = foldConvert(bDivA, a->get_output_element_type(0)); + bDivA = as_type_ptr(foldConvert(bDivA->output(0), a->get_element_type())); } OutputVector inputs{ {}, {} }; inputs[0] = x; - inputs[1] = bDivA; + inputs[1] = bDivA->output(0); std::shared_ptr newAdd = std::make_shared>( std::vector{element::f32, element::f32}, @@ -292,8 +292,8 @@ std::shared_ptr NetworkHelper::swapMultiplyAndAdd(std::shared_ptr>( std::vector{element::f32, element::f32}, std::vector{ multiply->get_output_element_type(0) }, - ngraph::op::TemporaryReplaceOutputType(newAdd, element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(a, element::f32).get()); + ngraph::op::TemporaryReplaceOutputType(newAdd->output(0), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(a->output(0), element::f32).get()); copyInfo({ multiply, newMultiply }, newMultiply); replace_node(addAfterMultiply, newMultiply); @@ -460,7 +460,7 @@ std::shared_ptr NetworkHelper::optimizeMultipliesAfter } auto newInput = multiply->input_value(1 - constant1->output(0).get_target_inputs().begin()->get_index()); - auto multiplyResult = fold(constant1, constant2); + auto multiplyResult = fold(constant1->output(0), constant2->output(0)); { // optimize constant shape: used in rfcn-resnet101-coco const auto multiplyResultConstant = ov::as_type_ptr(multiplyResult); @@ -526,13 +526,13 @@ FakeQuantizeDequantization NetworkHelper::foldDequantization(const std::shared_p } if (dequantization.subtract != nullptr) { - if (dequantization.subtract->input(0).get_element_type() != dequantization.subtract->input(1).get_element_type()) { + if (dequantization.subtract->get_input_element_type(0) != dequantization.subtract->get_input_element_type(1)) { return dequantization; } if (dequantization.subtractConvert != nullptr) { const auto convertionResult = foldConvert( - dequantization.subtractConstant, + dequantization.subtractConstant->output(0), dequantization.subtractConvert->get_element_type()); if (ov::is_type(convertionResult)) { replace_node(dequantization.subtractConvert, convertionResult); @@ -541,8 +541,8 @@ FakeQuantizeDequantization NetworkHelper::foldDequantization(const std::shared_p } const std::shared_ptr result = fold( - dequantization.subtract->get_input_node_shared_ptr(0), - dequantization.subtract->get_input_node_shared_ptr(1)); + dequantization.subtract->input_value(0), + dequantization.subtract->input_value(1)); if (ov::is_type(result)) { if (inPlace) { copyInfo(dequantization.subtract, result); @@ -555,18 +555,18 @@ FakeQuantizeDequantization NetworkHelper::foldDequantization(const std::shared_p } if (dequantization.multiply != nullptr) { - if (dequantization.multiply->input(0).get_element_type() != dequantization.multiply->input(1).get_element_type()) { + if (dequantization.multiply->get_input_element_type(0) != dequantization.multiply->get_input_element_type(1)) { return dequantization; } std::shared_ptr result = fold( - dequantization.multiply->get_input_node_shared_ptr(0), - dequantization.multiply->get_input_node_shared_ptr(1)); + dequantization.multiply->input_value(0), + dequantization.multiply->input_value(1)); if (!ov::is_type(result)) { return dequantization; } if (dequantization.multiply->get_output_element_type(0) != result->get_element_type()) { - result = foldConvert(result, dequantization.multiply->get_output_element_type(0)); + result = foldConvert(result->output(0), dequantization.multiply->get_output_element_type(0)); } if (inPlace) { copyInfo(dequantization.multiply, result); @@ -599,7 +599,7 @@ std::shared_ptr NetworkHelper::separateInStandaloneBranch(std::sha outputs.push_back(input.get_source_output()); } - auto subtract = dequantization.subtract->clone_with_new_inputs({parent, parentOnWeights->clone_with_new_inputs(outputs) }); + auto subtract = dequantization.subtract->clone_with_new_inputs({parent, parentOnWeights->clone_with_new_inputs(outputs)->output(0) }); subtract->set_friendly_name(""); copy_runtime_info(parent.get_node_shared_ptr(), subtract); parent = subtract->output(0); @@ -608,7 +608,7 @@ std::shared_ptr NetworkHelper::separateInStandaloneBranch(std::sha if (dequantization.multiply != nullptr) { auto multiply = dequantization.multiply->clone_with_new_inputs({ parent, - dequantization.multiply->get_input_node_shared_ptr(1)->clone_with_new_inputs({}) }); + dequantization.multiply->get_input_node_shared_ptr(1)->clone_with_new_inputs({})->output(0) }); multiply->set_friendly_name(""); copy_runtime_info(parent.get_node_shared_ptr(), multiply); parent = multiply->output(0); @@ -650,11 +650,11 @@ std::shared_ptr NetworkHelper::fuseConvert(const std::shar std::shared_ptr newFakeQuantize = std::make_shared>( std::vector{ element::f32, element::f32, element::f32, element::f32, element::f32 }, std::vector{}, - ngraph::op::TemporaryReplaceOutputType(fakeQuantize->get_input_node_shared_ptr(0), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(fakeQuantize->get_input_node_shared_ptr(1), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(fakeQuantize->get_input_node_shared_ptr(2), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(fakeQuantize->get_input_node_shared_ptr(3), element::f32).get(), - ngraph::op::TemporaryReplaceOutputType(fakeQuantize->get_input_node_shared_ptr(4), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(fakeQuantize->input_value(0), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(fakeQuantize->input_value(1), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(fakeQuantize->input_value(2), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(fakeQuantize->input_value(3), element::f32).get(), + ngraph::op::TemporaryReplaceOutputType(fakeQuantize->input_value(4), element::f32).get(), fakeQuantize->get_levels()); NetworkHelper::setOutDataPrecisionForTypeRelaxed(newFakeQuantize, node->get_output_element_type(0)); replace_node(node->shared_from_this(), newFakeQuantize); @@ -889,14 +889,14 @@ std::shared_ptr NetworkHelper::composeFakeQuantize(const s if (dequantization.subtract != nullptr) { const auto subtractValue = (dequantization.subtractConvert == nullptr) ? dequantization.subtractConstant : - foldConvert(dequantization.subtractConstant, dequantization.subtractConvert->output(0).get_element_type()); + foldConvert(dequantization.subtractConstant->output(0), dequantization.subtractConvert->get_destination_type()); const std::shared_ptr replacement = std::make_shared>( newFakeQuantize->input_value(0), newFakeQuantize->input_value(1), newFakeQuantize->input_value(2), - fold(newFakeQuantize->get_input_node_shared_ptr(3), subtractValue), - fold(newFakeQuantize->get_input_node_shared_ptr(4), subtractValue), + fold(newFakeQuantize->input_value(3), subtractValue), + fold(newFakeQuantize->input_value(4), subtractValue), newFakeQuantize->get_levels(), newFakeQuantize->get_auto_broadcast()); replace_node(dequantization.subtract, replacement); @@ -907,11 +907,9 @@ std::shared_ptr NetworkHelper::composeFakeQuantize(const s if (dequantization.multiply != nullptr) { // multiply different precision constants (value1 & value2) and convert result to first argument precision (value1) - auto multiply = []( - const std::shared_ptr& value1, - const std::shared_ptr& value2) -> std::shared_ptr { - const ngraph::element::Type precision1 = value1->output(0).get_element_type(); - const ngraph::element::Type precision2 = value2->output(0).get_element_type(); + auto multiply = [](const Output& value1, const Output& value2) { + const ngraph::element::Type precision1 = value1.get_element_type(); + const ngraph::element::Type precision2 = value2.get_element_type(); // 1) precision1 & precision2 are not equal but similar // 2) precision2 >= precision1 assert((precision2.is_real() == precision1.is_real()) && (precision2.bitwidth() >= precision1.bitwidth())); @@ -921,7 +919,7 @@ std::shared_ptr NetworkHelper::composeFakeQuantize(const s value2); if (output->output(0).get_element_type() != precision1) { - output = foldConvert(output, precision1); + output = foldConvert(output->output(0), precision1); } return output; @@ -931,8 +929,8 @@ std::shared_ptr NetworkHelper::composeFakeQuantize(const s newFakeQuantize->input_value(0ul), newFakeQuantize->input_value(1ul), newFakeQuantize->input_value(2ul), - multiply(newFakeQuantize->get_input_node_shared_ptr(3ul), dequantization.multiplyConstant), - multiply(newFakeQuantize->get_input_node_shared_ptr(4ul), dequantization.multiplyConstant), + multiply(newFakeQuantize->input_value(3ul), dequantization.multiplyConstant), + multiply(newFakeQuantize->input_value(4ul), dequantization.multiplyConstant), newFakeQuantize->get_levels(), newFakeQuantize->get_auto_broadcast()); @@ -956,8 +954,6 @@ std::tuple, std::shared_ptr> NetworkHelper::decompos const bool updatePrecision, const element::Type deqPrecision, const size_t outChannelsShapeIndex) { - using std::make_shared; - const auto outputLow = fq->input_value(3); const auto outputHigh = fq->input_value(4); @@ -1015,8 +1011,8 @@ std::tuple, std::shared_ptr> NetworkHelper::decompos nullptr; std::shared_ptr scale = std::make_shared(element::f32, outputLow.get_shape(), scales); - auto newMin = make_shared(outputLow.get_element_type(), outputLow.get_shape(), minValues); - auto newMax = make_shared(outputLow.get_element_type(), outputLow.get_shape(), maxValues); + auto newMin = std::make_shared(outputLow.get_element_type(), outputLow.get_shape(), minValues); + auto newMax = std::make_shared(outputLow.get_element_type(), outputLow.get_shape(), maxValues); if (isScalarLike(newMin)) { newMin = toScalar(newMin); @@ -1072,7 +1068,7 @@ std::tuple, std::shared_ptr> NetworkHelper::decompos std::shared_ptr newFqConstant = ov::as_type_ptr(newFQ); if (ov::is_type(newFQ)) { - convert = foldConvert(newFQ, precision); + convert = foldConvert(newFQ->output(0), precision); } else if (ov::is_type(newFQ)) { newFQ = setOutDataPrecision(ov::as_type_ptr(newFQ), precision); convert = newFQ; @@ -1192,11 +1188,9 @@ FakeQuantizeDequantization NetworkHelper::createDequantizationFromFakeQuantize( const bool hasZeroPoint, const bool updatePrecision, const element::Type deqPrecision) { - using std::make_shared; - const ngraph::element::Type_t fqPrecision = fq->get_output_element_type(0); - auto newMin = make_shared(fqPrecision, Shape{}, min); - auto newMax = make_shared(fqPrecision, Shape{}, max); + auto newMin = std::make_shared(fqPrecision, Shape{}, min); + auto newMax = std::make_shared(fqPrecision, Shape{}, max); auto outputLow = fq->input_value(3); auto outputHigh = fq->input_value(4); @@ -1205,12 +1199,12 @@ FakeQuantizeDequantization NetworkHelper::createDequantizationFromFakeQuantize( const std::shared_ptr scale = ov::as_type_ptr(foldConvert(fold( fold(outputHigh, outputLow), - fold(newMax, newMin)), deqPrecision)); + fold(newMax->output(0), newMin->output(0))), deqPrecision)); assert(scale != nullptr); std::shared_ptr shift = hasZeroPoint ? ov::as_type_ptr(foldConvert(fold( - fold(fold(newMin, outputHigh), fold(newMax, outputLow)), + fold(fold(newMin->output(0), outputHigh), fold(newMax->output(0), outputLow)), fold(outputHigh, outputLow)), deqPrecision)) : nullptr; assert((!hasZeroPoint) || (hasZeroPoint && shift != nullptr)); @@ -1240,7 +1234,7 @@ FakeQuantizeDequantization NetworkHelper::createDequantizationFromFakeQuantize( std::shared_ptr subtract; if (shift != nullptr) { - subtract = make_shared>(parent, shift); + subtract = std::make_shared>(parent, shift); subtract->set_output_type(0, deqPrecision, subtract->get_output_partial_shape(0)); parent = subtract; } else { @@ -1416,16 +1410,16 @@ FakeQuantizeDequantization NetworkHelper::normalizeDequantization(FakeQuantizeDe return dequantization; } if (dequantization.multiply != nullptr && ov::as_type_ptr(dequantization.multiply->get_input_node_shared_ptr(0))) { - std::shared_ptr leftParent = dequantization.multiply->get_input_node_shared_ptr(0); - std::shared_ptr rightParent = dequantization.multiply->get_input_node_shared_ptr(1); + const auto leftParent = dequantization.multiply->input_value(0); + const auto rightParent = dequantization.multiply->input_value(1); std::shared_ptr normalized_multiply = ov::as_type_ptr( dequantization.multiply->clone_with_new_inputs({rightParent, leftParent})); replace_node(dequantization.multiply, normalized_multiply); dequantization.multiply = normalized_multiply; } if (dequantization.subtract != nullptr && ov::as_type_ptr(dequantization.subtract->get_input_node_shared_ptr(0))) { - std::shared_ptr leftParent = dequantization.subtract->get_input_node_shared_ptr(0); - std::shared_ptr rightParent = dequantization.subtract->get_input_node_shared_ptr(1); + const auto leftParent = dequantization.subtract->input_value(0); + const auto rightParent = dequantization.subtract->input_value(1); std::shared_ptr normalized_subtract = ov::as_type_ptr( dequantization.subtract->clone_with_new_inputs({rightParent, leftParent})); replace_node(dequantization.subtract, normalized_subtract); @@ -1452,7 +1446,7 @@ std::shared_ptr NetworkHelper::normalizeDequantizationShape(co std::iota(unsqueezeConstantShape.begin(), unsqueezeConstantShape.end(), 0ul); const auto newConstant = fold( - constant, + constant->output(0), op::Constant::create(element::i32, Shape{ unsqueezeConstantShape.size() }, unsqueezeConstantShape)); return ov::as_type_ptr(newConstant); @@ -1471,13 +1465,13 @@ std::shared_ptr NetworkHelper::normalizeDequantizationShape(co FakeQuantizeDequantizationValues NetworkHelper::createEmptyValues(const FakeQuantizeDequantization& dequantization, const element::Type precision) { const std::shared_ptr multiplyConstant = dequantization.multiply ? dequantization.multiplyConstant->get_element_type() != precision ? - foldConvert(dequantization.multiplyConstant, precision) : + foldConvert(dequantization.multiplyConstant->output(0), precision) : dequantization.multiplyConstant : std::make_shared(precision, Shape({}), std::vector({ 1.f })); const std::shared_ptr subtractConstant = dequantization.subtract ? dequantization.subtractConstant->get_element_type() != precision ? - foldConvert(dequantization.subtractConstant, precision) : + foldConvert(dequantization.subtractConstant->output(0), precision) : dequantization.subtractConstant : std::make_shared(precision, Shape({}), std::vector({ 0.f })); @@ -1538,7 +1532,7 @@ std::shared_ptr NetworkHelper::optimizeSubtract(std::shared_ptr>(data, roundedShift); + replacement = std::make_shared>(data, roundedShift->output(0)); NetworkHelper::copyInfo(subtract, replacement); NetworkHelper::setOutDataPrecisionForTypeRelaxed(replacement, convertOutputType); replace_node(subtract, replacement); @@ -1546,7 +1540,7 @@ std::shared_ptr NetworkHelper::optimizeSubtract(std::shared_ptr(subtractParent) && ov::is_type(subtractParent->get_input_node_shared_ptr(0))) { - auto replacement = std::make_shared>(data, subtractParent->get_input_node_shared_ptr(0)); + auto replacement = std::make_shared>(data, subtractParent->input_value(0)); NetworkHelper::copyInfo(subtract, replacement); NetworkHelper::setOutDataPrecisionForTypeRelaxed(replacement, convertOutputType); replace_node(subtract, replacement); @@ -1569,11 +1563,9 @@ NetworkHelper::InsertDequantizationResult NetworkHelper::moveDequantizationAfter (NetworkHelper::getDequantization(operation).multiplyConstant == nullptr) || (NetworkHelper::getDequantization(operation).multiplyConstant.get() == dequantization.multiplyConstant.get())); - std::vector> inputs(operation->get_input_size()); - for (size_t i = 0; i < operation->get_input_size(); ++i) { - inputs[i] = operation->get_input_node_shared_ptr(i); - } + assert(operation->get_output_size() == 1); + OutputVector inputs = operation->input_values(); const size_t dequantizationIndex = getChildInputIndex(dequantization.multiply, operation); inputs[dequantizationIndex] = moveSubtract ? dequantization.data : @@ -1623,7 +1615,7 @@ NetworkHelper::InsertDequantizationResult NetworkHelper::moveDequantizationAfter ngraph::op::TemporaryReplaceOutputType( dequantization.subtractConstant->output(0).get_element_type() == parentPrecision ? dequantization.subtractConstant : - foldConvert(dequantization.subtractConstant, parentPrecision), element::f32).get()); + foldConvert(dequantization.subtractConstant->output(0), parentPrecision), element::f32).get()); ngraph::copy_runtime_info({ newOperation, parent }, parent); } else { parent = std::make_shared(parent, dequantization.subtractConvert); diff --git a/inference-engine/src/low_precision_transformations/src/pull_reshape_through_dequantization.cpp b/inference-engine/src/low_precision_transformations/src/pull_reshape_through_dequantization.cpp index 68617278844..42afb73e21b 100644 --- a/inference-engine/src/low_precision_transformations/src/pull_reshape_through_dequantization.cpp +++ b/inference-engine/src/low_precision_transformations/src/pull_reshape_through_dequantization.cpp @@ -30,15 +30,15 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& reshap assert(ov::is_type(elementwiseValues)); const std::shared_ptr newReshape = ov::as_type_ptr(reshape->clone_with_new_inputs({ - elementwise->get_input_node_shared_ptr(0ul), + elementwise->input_value(0), reshapeValues })); std::shared_ptr newElementwiseValues; - const Shape elementwiseValuesShape = elementwiseValues->output(0).get_shape(); + const Shape elementwiseValuesShape = elementwiseValues->get_output_shape(0); if (!elementwiseValuesShape.empty() && (elementwiseValuesShape.size() != 1ul)) { // update shape constant value to avoid eltwise constan value broadcasting - const Shape elementwiseShape = elementwise->output(0).get_shape(); + const Shape elementwiseShape = elementwise->get_output_shape(0); const std::vector reshapeValuesVector = ov::as_type_ptr(reshapeValues)->cast_vector(); const std::vector newReshapeValuesVector = ngraph::pass::low_precision::NetworkHelper::updateReshapeValues( @@ -47,13 +47,13 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& reshap reshapeValuesVector); const auto newReshapeValues = std::make_shared( - reshapeValues->output(0).get_element_type(), + reshapeValues->get_output_element_type(0), Shape{ newReshapeValuesVector.size() }, newReshapeValuesVector); newElementwiseValues = ngraph::pass::low_precision::fold_reshape( - elementwiseValues->output(0), - newReshapeValues->output(0), + elementwiseValues, + newReshapeValues, ov::as_type_ptr(reshape)->get_special_zero()); assert(ov::is_type(newElementwiseValues)); } else { @@ -71,7 +71,7 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& reshap } std::shared_ptr moveThroughConvert(const std::shared_ptr& reshape, const std::shared_ptr& convert) { - const auto newReshape = reshape->clone_with_new_inputs({ convert->get_input_node_shared_ptr(0), reshape->get_input_node_shared_ptr(1) }); + const auto newReshape = reshape->clone_with_new_inputs({ convert->input_value(0), reshape->input_value(1) }); const auto newConvert = convert->clone_with_new_inputs({ newReshape }); replace_node(reshape, newConvert); copy_runtime_info({ convert, reshape }, { newReshape, newConvert }); @@ -81,7 +81,7 @@ std::shared_ptr moveThroughConvert(const std::shared_ptr& reshape, c void fuseConstant(const std::shared_ptr& reshape, const std::shared_ptr& constant) { ngraph::OutputVector result(1); - reshape->constant_fold(result, { constant->output(0), reshape->get_input_node_ptr(1)->output(0) }); + reshape->constant_fold(result, { constant, reshape->input_value(1) }); const auto newConstant = result[0].get_node_shared_ptr(); replace_node(reshape, newConstant); copy_runtime_info({ constant, reshape }, newConstant); diff --git a/inference-engine/src/low_precision_transformations/src/pull_transpose_through_dequantization.cpp b/inference-engine/src/low_precision_transformations/src/pull_transpose_through_dequantization.cpp index 3ee344884dc..2af61df6734 100644 --- a/inference-engine/src/low_precision_transformations/src/pull_transpose_through_dequantization.cpp +++ b/inference-engine/src/low_precision_transformations/src/pull_transpose_through_dequantization.cpp @@ -30,8 +30,8 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& transp elementwiseValuesConvert->get_input_node_shared_ptr(0ul); assert(ov::is_type(elementwiseValues)); - const auto transposeValuesShape = transposeValues->output(0).get_shape(); - const auto elementwiseValuesShape = elementwiseValues->output(0).get_shape(); + const auto transposeValuesShape = transposeValues->get_output_shape(0); + const auto elementwiseValuesShape = elementwiseValues->get_output_shape(0); if (elementwiseValuesShape.size() != shape_size(transposeValuesShape)) { if (shape_size(elementwiseValuesShape) != 1ul) { return nullptr; @@ -51,8 +51,8 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& transp transposeValues })); const auto newElementwiseValues = ngraph::pass::low_precision::fold( - elementwiseValues->output(0), - transposeValues->output(0)); + elementwiseValues, + transposeValues); assert(ov::is_type(newElementwiseValues)); const auto newElementwise = elementwise->clone_with_new_inputs({ @@ -68,7 +68,7 @@ std::shared_ptr moveThroughElementwise(const std::shared_ptr& transp } std::shared_ptr moveThroughConvert(const std::shared_ptr& transpose, const std::shared_ptr& convert) { - const auto newTranspose = transpose->clone_with_new_inputs({convert->get_input_node_shared_ptr(0), transpose->get_input_node_ptr(1)->output(0) }); + const auto newTranspose = transpose->clone_with_new_inputs({convert->input_value(0), transpose->input_value(1) }); const auto newConvert = convert->clone_with_new_inputs({ newTranspose }); replace_node(transpose, newConvert); copy_runtime_info({ convert, transpose }, { newTranspose, newConvert }); @@ -78,8 +78,8 @@ std::shared_ptr moveThroughConvert(const std::shared_ptr& transpose, void fuseConstant(const std::shared_ptr& transpose, const std::shared_ptr& constant) { const auto newConstant = ngraph::pass::low_precision::fold( - constant->output(0), - transpose->get_input_node_ptr(1)->output(0)); + constant, + transpose->input_value(1)); replace_node(transpose, newConstant); copy_runtime_info({ constant, transpose }, newConstant); diff --git a/inference-engine/src/low_precision_transformations/src/reshape.cpp b/inference-engine/src/low_precision_transformations/src/reshape.cpp index ee8e02e1045..aeba26b7e49 100644 --- a/inference-engine/src/low_precision_transformations/src/reshape.cpp +++ b/inference-engine/src/low_precision_transformations/src/reshape.cpp @@ -63,7 +63,7 @@ void reshapeDequantizationConstant(const std::shared_ptr& resha } } - const auto reshapeOutputPShape = reshape->output(0).get_partial_shape(); + const auto reshapeOutputPShape = reshape->get_output_partial_shape(0); const auto reshapeOutputRank = reshapeOutputPShape.rank(); assert(reshapeOutputRank.is_static()); assert(reshapeOutputRank.get_length() >= 2); diff --git a/inference-engine/src/low_precision_transformations/src/rt_info/intervals_alignment_attribute.cpp b/inference-engine/src/low_precision_transformations/src/rt_info/intervals_alignment_attribute.cpp index 95a6168db9c..cb8b650bac8 100644 --- a/inference-engine/src/low_precision_transformations/src/rt_info/intervals_alignment_attribute.cpp +++ b/inference-engine/src/low_precision_transformations/src/rt_info/intervals_alignment_attribute.cpp @@ -52,7 +52,7 @@ std::shared_ptr>> Va FakeQuantizeDequantization dequantization; { - const auto targetInputs = node->output(0).get_target_inputs(); + const auto targetInputs = node->get_output_target_inputs(0); if (targetInputs.size() == 1ul) { dequantization = NetworkHelper::getDequantizationBelow(node, true); } @@ -75,7 +75,7 @@ std::shared_ptr>> Va auto multiplyResult = dequantization.multiplyConstant == nullptr ? node->get_input_node_ptr(3)->shared_from_this() : fold( - foldConvert(node->get_input_node_ptr(3)->shared_from_this(), params.deqPrecision), + foldConvert(node->input_value(3), params.deqPrecision), dequantization.multiplyConstant); auto multiplyResultConstant = ov::as_type_ptr(multiplyResult); @@ -87,7 +87,7 @@ std::shared_ptr>> Va auto multiplyResult = dequantization.multiplyConstant == nullptr ? node->get_input_node_ptr(4)->shared_from_this() : fold( - foldConvert(node->get_input_node_ptr(4)->shared_from_this(), params.deqPrecision), + foldConvert(node->input_value(4), params.deqPrecision), dequantization.multiplyConstant); auto multiplyResultConstant = ov::as_type_ptr(multiplyResult); diff --git a/inference-engine/src/low_precision_transformations/src/rt_info/shared_value_attribute.cpp b/inference-engine/src/low_precision_transformations/src/rt_info/shared_value_attribute.cpp deleted file mode 100644 index 95cc5fa72ea..00000000000 --- a/inference-engine/src/low_precision_transformations/src/rt_info/shared_value_attribute.cpp +++ /dev/null @@ -1,16 +0,0 @@ -// Copyright (C) 2021 Intel Corporation -// SPDX-License-Identifier: Apache-2.0 -// - -#include "low_precision/rt_info/shared_value_attribute.hpp" - -#include -#include -#include -#include -#include - -#include -#include "low_precision/network_helper.hpp" - -using namespace ngraph; diff --git a/inference-engine/src/low_precision_transformations/src/squeeze.cpp b/inference-engine/src/low_precision_transformations/src/squeeze.cpp index 42d8e7e5932..b6290e10793 100644 --- a/inference-engine/src/low_precision_transformations/src/squeeze.cpp +++ b/inference-engine/src/low_precision_transformations/src/squeeze.cpp @@ -47,7 +47,7 @@ bool SqueezeTransformation::transform(TransformationContext& context, ngraph::pa return NetworkHelper::toScalar(dequantizationOpConstant); } if (constantShape.size() == inputRankValue) { - return ov::as_type_ptr(fold(dequantizationOpConstant, squeeze->get_input_node_shared_ptr(1))); + return ov::as_type_ptr(fold(dequantizationOpConstant, squeeze->input_value(1))); } return dequantizationOpConstant; diff --git a/inference-engine/src/low_precision_transformations/src/strided_slice.cpp b/inference-engine/src/low_precision_transformations/src/strided_slice.cpp index 12744e330ce..a8c709cd31e 100644 --- a/inference-engine/src/low_precision_transformations/src/strided_slice.cpp +++ b/inference-engine/src/low_precision_transformations/src/strided_slice.cpp @@ -62,9 +62,9 @@ std::shared_ptr stridedSliceDeqConstant( const auto result = fold( constant, - stridedSlice->get_input_node_shared_ptr(1), - stridedSlice->get_input_node_shared_ptr(2), - stridedSlice->get_input_node_shared_ptr(3), + stridedSlice->input_value(1), + stridedSlice->input_value(2), + stridedSlice->input_value(3), beginMask, endMask, stridedSlice->get_new_axis_mask(), diff --git a/inference-engine/src/low_precision_transformations/src/subtract.cpp b/inference-engine/src/low_precision_transformations/src/subtract.cpp index 46e2245bcb3..2b1a8a777b1 100644 --- a/inference-engine/src/low_precision_transformations/src/subtract.cpp +++ b/inference-engine/src/low_precision_transformations/src/subtract.cpp @@ -55,10 +55,10 @@ bool SubtractTransformation::transform(TransformationContext& context, ngraph::p // X * SC - SH = X * SC - SH' * SC // SH' = SH / SC std::shared_ptr newSubtract = ov::as_type_ptr(subtract->copy_with_new_inputs({ - dequantization.multiply->get_input_node_shared_ptr(0), + dequantization.multiply->input_value(0), ngraph::pass::low_precision::fold( - subtract->get_input_node_shared_ptr(1), - dequantization.multiply->get_input_node_shared_ptr(1)) + subtract->input_value(1), + dequantization.multiply->input_value(1)) })); std::shared_ptr newMultiply = dequantization.multiply->copy_with_new_inputs({ @@ -72,8 +72,8 @@ bool SubtractTransformation::transform(TransformationContext& context, ngraph::p if (dequantization.subtract != nullptr) { std::shared_ptr newSubtract = ov::as_type_ptr(subtract->copy_with_new_inputs({ - dequantization.subtract->get_input_node_shared_ptr(0), - fold(subtract->get_input_node_shared_ptr(1), dequantization.subtractConstant) + dequantization.subtract->input_value(0), + fold(subtract->input_value(1), dequantization.subtractConstant) })); replace_node(subtract, newSubtract); @@ -86,8 +86,8 @@ bool SubtractTransformation::transform(TransformationContext& context, ngraph::p subtract->set_output_type(0, originalPrecision, subtract->get_output_partial_shape(0)); replace_node(subtract, std::make_shared>( - subtract->get_input_node_shared_ptr(0), - subtract->get_input_node_shared_ptr(1))); + subtract->input_value(0), + subtract->input_value(1))); } return true; } diff --git a/inference-engine/src/low_precision_transformations/src/transparent_base_transformation.cpp b/inference-engine/src/low_precision_transformations/src/transparent_base_transformation.cpp index c89ca0e9144..23d9922c454 100644 --- a/inference-engine/src/low_precision_transformations/src/transparent_base_transformation.cpp +++ b/inference-engine/src/low_precision_transformations/src/transparent_base_transformation.cpp @@ -4,9 +4,7 @@ #include "low_precision/transparent_base_transformation.hpp" -#include #include -#include #include #include "low_precision/network_helper.hpp" @@ -16,27 +14,20 @@ using namespace ngraph::pass; using namespace ngraph::pass::low_precision; bool TransparentBaseTransformation::transform(TransformationContext& context, ngraph::pattern::Matcher &m) { - auto operation = m.get_match_root(); - const std::shared_ptr dequantization = operation->input_value(0).get_node_shared_ptr(); - // const std::shared_ptr dequantizationParent = dequantization->input_value(0).get_node_shared_ptr(); + std::shared_ptr op = m.get_match_root(); + if (!canBeTransformed(context, op)) { + return false; + } - // auto newOperation = operation->copy_with_new_inputs({ dequantizationParent }); - // const auto newDequantization = dequantization->copy_with_new_inputs({ - // newOperation, - // dequantization->input_value(1), - // dequantization->input_value(2) }); - - // const std::string friendlyName = operation->get_friendly_name(); - //// TODO: new operation name has to be unique - // newOperation->set_friendly_name(friendlyName + "_original"); - // newDequantization->set_friendly_name(friendlyName); - - // replace_node(operation, newDequantization); - - // NetworkHelper::moveDequantization(operation, dequantization); + op = NetworkHelper::separateInStandaloneBranch(op); + moveDequantizationAfter(context, op, NetworkHelper::getDequantization(op), true); return true; } bool TransparentBaseTransformation::canBeTransformed(const TransformationContext& context, std::shared_ptr layer) const { return true; } + +bool TransparentBaseTransformation::isPrecisionPreserved(std::shared_ptr layer) const noexcept { + return true; +} diff --git a/inference-engine/src/low_precision_transformations/src/unsqueeze.cpp b/inference-engine/src/low_precision_transformations/src/unsqueeze.cpp index 011bb4d46f0..a515085f017 100644 --- a/inference-engine/src/low_precision_transformations/src/unsqueeze.cpp +++ b/inference-engine/src/low_precision_transformations/src/unsqueeze.cpp @@ -48,7 +48,7 @@ bool UnsqueezeTransformation::transform(TransformationContext& context, ngraph:: } if (constantShape.size() == inputRankValue) { - return ov::as_type_ptr(fold(dequantizationOpConstant, unsqueeze->get_input_node_shared_ptr(1))); + return ov::as_type_ptr(fold(dequantizationOpConstant, unsqueeze->input_value(1))); } return dequantizationOpConstant;