diff options
Diffstat (limited to 'src/backends')
19 files changed, 269 insertions, 349 deletions
diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp index 766bf2d2cc..c9d6c71f18 100644 --- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp +++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -188,8 +188,6 @@ LayerType* FuseConvolution2dLayer(OptimizationViews& optimizationViews, LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement); - replacementLayer->m_Weight = std::move(baseLayer->m_Weight); - replacementLayer->m_Bias = std::move(baseLayer->m_Bias); FuseLayer(optimizationViews, baseLayer, @@ -212,8 +210,6 @@ LayerType* FuseDepthwiseConvolution2dLayer(OptimizationViews& optimizationViews, LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement); - replacementLayer->m_Weight = std::move(baseLayer->m_Weight); - replacementLayer->m_Bias = std::move(baseLayer->m_Bias); FuseLayer(optimizationViews, baseLayer, @@ -242,8 +238,6 @@ LayerType* FuseFullyConnectedLayer(OptimizationViews& optimizationViews, activationLayer, activationDesc); - replacementLayer->m_Weight = std::move(baseLayer->m_Weight); - replacementLayer->m_Bias = std::move(baseLayer->m_Bias); return replacementLayer; } diff --git a/src/backends/backendsCommon/test/DynamicBackendTests.hpp b/src/backends/backendsCommon/test/DynamicBackendTests.hpp index 0d98804954..6a673c6e93 100644 --- a/src/backends/backendsCommon/test/DynamicBackendTests.hpp +++ b/src/backends/backendsCommon/test/DynamicBackendTests.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -1469,8 +1469,6 @@ void CreateReferenceDynamicBackendTestImpl() { outputInfo } }; convolution2dQueueDescriptor.m_Inputs.push_back(nullptr); - auto weights = std::make_unique<ScopedTensorHandle>(weightInfo); - convolution2dQueueDescriptor.m_Weight = weights.get(); // Create a convolution workload with the dummy settings auto workload = referenceWorkloadFactory->CreateWorkload(LayerType::Convolution2d, diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp index 18f11a542e..fb7a0271d4 100644 --- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp +++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -242,10 +242,6 @@ struct DummyConvolutionLayer desc.m_StrideX = 1; desc.m_StrideY = 1; m_Layer = dummyGraph.AddLayer<ConvolutionLayerType>(desc, ""); - m_Layer->m_Weight = std::make_unique<armnn::ScopedTensorHandle>( - armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); - m_Layer->m_Bias = std::make_unique<armnn::ScopedTensorHandle>( - armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); } ~DummyConvolutionLayer() @@ -268,10 +264,28 @@ struct DummyLayer<armnn::DepthwiseConvolution2dLayer> { }; -template<> +// Note: When m_Weight and m_Bias are removed from TransposeConvolution, Transpose can use DummyConvolutionLayer +template <> struct DummyLayer<armnn::TransposeConvolution2dLayer> - : public DummyConvolutionLayer<armnn::TransposeConvolution2dLayer> { + DummyLayer() + { + typename armnn::TransposeConvolution2dLayer::DescriptorType desc; + desc.m_StrideX = 1; + desc.m_StrideY = 1; + m_Layer = dummyGraph.AddLayer<armnn::TransposeConvolution2dLayer>(desc, ""); + m_Layer->m_Weight = std::make_unique<armnn::ScopedTensorHandle>( + armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); + m_Layer->m_Bias = std::make_unique<armnn::ScopedTensorHandle>( + armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); + } + + ~DummyLayer() + { + dummyGraph.EraseLayer(m_Layer); + } + + armnn::TransposeConvolution2dLayer* m_Layer; }; template<> @@ -518,8 +532,6 @@ struct DummyLayer<armnn::FullyConnectedLayer> { armnn::FullyConnectedLayer::DescriptorType desc; m_Layer = dummyGraph.AddLayer<armnn::FullyConnectedLayer>(desc, ""); - m_Layer->m_Weight = std::make_unique<armnn::ScopedTensorHandle>( - armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); } ~DummyLayer() @@ -767,12 +779,6 @@ unsigned int GetNumOutputs(const armnn::Layer& layer) return layer.GetNumOutputSlots(); } -template<> -unsigned int GetNumInputs<armnn::LayerType::Concat>(const armnn::Layer& layer) -{ - IgnoreUnused(layer); - return 2; -} // Tests that the IsLayerSupported() function returns the correct value. // We determined the correct value by *trying* to create the relevant workload and seeing if it matches what we expect. diff --git a/src/backends/backendsCommon/test/LayerReleaseConstantDataTest.cpp b/src/backends/backendsCommon/test/LayerReleaseConstantDataTest.cpp index abfb621c93..b7b514573c 100644 --- a/src/backends/backendsCommon/test/LayerReleaseConstantDataTest.cpp +++ b/src/backends/backendsCommon/test/LayerReleaseConstantDataTest.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -12,8 +12,6 @@ #include <doctest/doctest.h> -#include <utility> - using namespace armnn; using namespace std; @@ -84,29 +82,15 @@ TEST_CASE("ReleaseConvolution2dLayerConstantDataTest") layerDesc.m_StrideY = 4; layerDesc.m_BiasEnabled = true; - Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer"); - - layer->m_Weight = std::make_unique<ScopedTensorHandle>(TensorInfo({ 2, 3, 5, 3 }, - armnn::DataType::Float32)); - layer->m_Bias = std::make_unique<ScopedTensorHandle> - (TensorInfo({ 2 }, GetBiasDataType(armnn::DataType::Float32))); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); - - ConstantLayer* weightsLayer = graph.AddLayer<ConstantLayer>("Weights"); - ConstantLayer* biasLayer = graph.AddLayer<ConstantLayer>("Bias"); + auto* const convolutionLayer = graph.AddLayer<Convolution2dLayer>(layerDesc, "convolution"); + auto* const weightsLayer = graph.AddLayer<ConstantLayer>("weights"); + auto* const biasLayer = graph.AddLayer<ConstantLayer>("bias"); - weightsLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(TensorInfo({ 2, 3, 5, 3 }, - armnn::DataType::Float32)); + TensorInfo weightsInfo = TensorInfo({ 2, 3, 5, 3 }, armnn::DataType::Float32, 1.0, 0.0, true); + TensorInfo biasInfo = TensorInfo({ 2 }, GetBiasDataType(armnn::DataType::Float32), 1.0, 0.0, true); - biasLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>( - TensorInfo({2}, GetBiasDataType(armnn::DataType::Float32))); - - TensorInfo weightsInfo = weightsLayer->m_LayerOutput->GetTensorInfo(); - weightsInfo.SetConstant(); - TensorInfo biasInfo = biasLayer->m_LayerOutput->GetTensorInfo(); - biasInfo.SetConstant(); + weightsLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(biasInfo); weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); @@ -116,21 +100,21 @@ TEST_CASE("ReleaseConvolution2dLayerConstantDataTest") Layer* const output = graph.AddLayer<OutputLayer>(0, "output"); // connect up - Connect(input, layer, TensorInfo({ 2, 3, 8, 16 }, armnn::DataType::Float32)); - weightsLayer->GetOutputSlot().Connect(layer->GetInputSlot(1)); - biasLayer->GetOutputSlot().Connect(layer->GetInputSlot(2)); - Connect(layer, output, TensorInfo({ 2, 2, 2, 10 }, armnn::DataType::Float32)); + Connect(input, convolutionLayer, TensorInfo({ 2, 3, 8, 16 }, armnn::DataType::Float32)); + weightsLayer->GetOutputSlot().Connect(convolutionLayer->GetInputSlot(1)); + biasLayer->GetOutputSlot().Connect(convolutionLayer->GetInputSlot(2)); + Connect(convolutionLayer, output, TensorInfo({ 2, 2, 2, 10 }, armnn::DataType::Float32)); // check the constants that they are not NULL CHECK(weightsLayer->m_LayerOutput != nullptr); CHECK(biasLayer->m_LayerOutput != nullptr); - // free up the constants.. - layer->ReleaseConstantData(); + // free up the constants. + convolutionLayer->ReleaseConstantData(); - // check the constants that they are NULL now - CHECK(weightsLayer->m_LayerOutput == nullptr); - CHECK(biasLayer->m_LayerOutput == nullptr); + // check the constants that they are still not NULL + CHECK(weightsLayer->m_LayerOutput != nullptr); + CHECK(biasLayer->m_LayerOutput != nullptr); } TEST_CASE("ReleaseDepthwiseConvolution2dLayerConstantDataTest") @@ -147,33 +131,39 @@ TEST_CASE("ReleaseDepthwiseConvolution2dLayerConstantDataTest") layerDesc.m_StrideY = 4; layerDesc.m_BiasEnabled = true; - DepthwiseConvolution2dLayer* const layer = graph.AddLayer<DepthwiseConvolution2dLayer>(layerDesc, "layer"); + auto* const convolutionLayer = graph.AddLayer<DepthwiseConvolution2dLayer>(layerDesc, "convolution"); + auto* const weightsLayer = graph.AddLayer<ConstantLayer>("weights"); + auto* const biasLayer = graph.AddLayer<ConstantLayer>("bias"); + + TensorInfo weightsInfo = TensorInfo({ 3, 3, 5, 3 }, armnn::DataType::Float32, 1.0, 0.0, true); + TensorInfo biasInfo = TensorInfo({ 9 }, GetBiasDataType(armnn::DataType::Float32), 1.0, 0.0, true); + + weightsLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(biasInfo); - layer->m_Weight = std::make_unique<ScopedTensorHandle>( - TensorInfo({3, 3, 5, 3}, DataType::Float32)); - layer->m_Bias = std::make_unique<ScopedTensorHandle>( - TensorInfo({9}, DataType::Float32)); - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); + weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); + biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); // create extra layers Layer* const input = graph.AddLayer<InputLayer>(0, "input"); Layer* const output = graph.AddLayer<OutputLayer>(0, "output"); // connect up - Connect(input, layer, TensorInfo({2, 3, 8, 16}, armnn::DataType::Float32)); - Connect(layer, output, TensorInfo({2, 9, 2, 10}, armnn::DataType::Float32)); + Connect(input, convolutionLayer, TensorInfo({2, 3, 8, 16}, armnn::DataType::Float32)); + weightsLayer->GetOutputSlot().Connect(convolutionLayer->GetInputSlot(1)); + biasLayer->GetOutputSlot().Connect(convolutionLayer->GetInputSlot(2)); + Connect(convolutionLayer, output, TensorInfo({2, 9, 2, 10}, armnn::DataType::Float32)); // check the constants that they are not NULL - CHECK(layer->m_Weight != nullptr); - CHECK(layer->m_Bias != nullptr); + CHECK(weightsLayer->m_LayerOutput != nullptr); + CHECK(biasLayer->m_LayerOutput != nullptr); - // free up the constants.. - layer->ReleaseConstantData(); + // free up the constants. + convolutionLayer->ReleaseConstantData(); - // check the constants that they are NULL now - CHECK(layer->m_Weight == nullptr); - CHECK(layer->m_Bias == nullptr); + // check the constants that they are still not NULL + CHECK(weightsLayer->m_LayerOutput != nullptr); + CHECK(biasLayer->m_LayerOutput != nullptr); } TEST_CASE("ReleaseFullyConnectedLayerConstantDataTest") @@ -185,36 +175,42 @@ TEST_CASE("ReleaseFullyConnectedLayerConstantDataTest") layerDesc.m_BiasEnabled = true; layerDesc.m_TransposeWeightMatrix = true; - FullyConnectedLayer* const layer = graph.AddLayer<FullyConnectedLayer>(layerDesc, "layer"); + auto* const fullyConnectedLayer = graph.AddLayer<FullyConnectedLayer>(layerDesc, "layer"); + auto* const weightsLayer = graph.AddLayer<ConstantLayer>("weights"); + auto* const biasLayer = graph.AddLayer<ConstantLayer>("bias"); float inputsQScale = 1.0f; float outputQScale = 2.0f; - layer->m_Weight = std::make_unique<ScopedTensorHandle>( - TensorInfo({7, 20}, DataType::QAsymmU8, inputsQScale, 0)); - layer->m_Bias = std::make_unique<ScopedTensorHandle>( - TensorInfo({7}, GetBiasDataType(DataType::QAsymmU8), inputsQScale)); - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); + TensorInfo weightsInfo = TensorInfo({ 7, 20 }, DataType::QAsymmU8, inputsQScale, 0.0, true); + TensorInfo biasInfo = TensorInfo({ 7 }, GetBiasDataType(DataType::QAsymmU8), inputsQScale, 0.0, true); + + weightsLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared<ScopedTensorHandle>(biasInfo); + + weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); + biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); // create extra layers Layer* const input = graph.AddLayer<InputLayer>(0, "input"); Layer* const output = graph.AddLayer<OutputLayer>(0, "output"); // connect up - Connect(input, layer, TensorInfo({3, 1, 4, 5}, DataType::QAsymmU8, inputsQScale)); - Connect(layer, output, TensorInfo({3, 7}, DataType::QAsymmU8, outputQScale)); + Connect(input, fullyConnectedLayer, TensorInfo({ 3, 1, 4, 5 }, DataType::QAsymmU8, inputsQScale)); + weightsLayer->GetOutputSlot().Connect(fullyConnectedLayer->GetInputSlot(1)); + biasLayer->GetOutputSlot().Connect(fullyConnectedLayer->GetInputSlot(2)); + Connect(fullyConnectedLayer, output, TensorInfo({ 3, 7 }, DataType::QAsymmU8, outputQScale)); // check the constants that they are not NULL - CHECK(layer->m_Weight != nullptr); - CHECK(layer->m_Bias != nullptr); + CHECK(weightsLayer->m_LayerOutput != nullptr); + CHECK(biasLayer->m_LayerOutput != nullptr); - // free up the constants.. - layer->ReleaseConstantData(); + // free up the constants. + fullyConnectedLayer->ReleaseConstantData(); - // check the constants that they are NULL now - CHECK(layer->m_Weight == nullptr); - CHECK(layer->m_Bias == nullptr); + // check the constants that they are still not NULL + CHECK(weightsLayer->m_LayerOutput != nullptr); + CHECK(biasLayer->m_LayerOutput != nullptr); } } diff --git a/src/backends/backendsCommon/test/OptimizeSubgraphViewTests.cpp b/src/backends/backendsCommon/test/OptimizeSubgraphViewTests.cpp index 67354696b0..997fe9850d 100644 --- a/src/backends/backendsCommon/test/OptimizeSubgraphViewTests.cpp +++ b/src/backends/backendsCommon/test/OptimizeSubgraphViewTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -94,13 +94,10 @@ Convolution2dLayer* AddConvolutionLayer(Graph& graph, LayerNameToLayerMap& layersInGraph, const Convolution2dDescriptor& convolutionDescriptor, const std::string& layerName, - const TensorInfo& weightInfo, - const TensorInfo& biasInfo, const TensorInfo& outputInfo) { Convolution2dLayer* const convLayer = graph.AddLayer<Convolution2dLayer>(convolutionDescriptor, layerName.c_str()); CHECK(convLayer); - SetWeightAndBias(convLayer, weightInfo, biasInfo); convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); layersInGraph.insert(std::make_pair(convLayer->GetName(), convLayer)); return convLayer; @@ -335,11 +332,11 @@ SubgraphView::SubgraphViewPtr BuildFullyOptimizableSubgraph1(Graph& graph, Layer // Construct the graph Layer* const inputLayer = AddInputLayer(graph, "input layer", inputInfo); Convolution2dLayer* const convLayer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv layer", weightInfo, biasInfo, outputInfo); + "conv layer", outputInfo); ConstantLayer* const weightsLayer = - AddConstantLayer(graph, layersInGraph, "Weights Layer", constWeightsTensor, outputInfo); - ConstantLayer* const biasLayer = AddConstantLayer(graph, layersInGraph, "Bias Layer", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer", constWeightsTensor, weightInfo); + ConstantLayer* const biasLayer = AddConstantLayer(graph, layersInGraph, "Bias Layer", constBiasTensor, biasInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); @@ -373,7 +370,6 @@ SubgraphView::SubgraphViewPtr BuildFullyOptimizableSubgraph2(Graph& graph, Layer std::vector<float> biasVector(16); ConstTensor constBiasTensor(biasInfo, biasVector); - Convolution2dDescriptor convolutionDescriptor; convolutionDescriptor.m_StrideX = 1; convolutionDescriptor.m_StrideY = 1; @@ -383,40 +379,40 @@ SubgraphView::SubgraphViewPtr BuildFullyOptimizableSubgraph2(Graph& graph, Layer // Construct the graph Layer* const inputLayer = AddInputLayer(graph, "input layer", inputInfo); Convolution2dLayer* const conv1Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv1 layer", weightInfo, biasInfo, outputInfo); + "conv1 layer", outputInfo); ConstantLayer* const weightsLayer1 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer1 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, biasInfo); Convolution2dLayer* const conv2Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv2 layer", weightInfo, biasInfo, outputInfo); + "conv2 layer", outputInfo); ConstantLayer* const weightsLayer2 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 2", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 2", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer2 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 2", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 2", constBiasTensor, biasInfo); Convolution2dLayer* const conv3Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv3 layer", weightInfo, biasInfo, outputInfo); + "conv3 layer", outputInfo); ConstantLayer* const weightsLayer3 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer3 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, biasInfo); Convolution2dLayer* const conv4Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv4 layer", weightInfo, biasInfo, outputInfo); + "conv4 layer", outputInfo); ConstantLayer* const weightsLayer4 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 4", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 4", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer4 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 4", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 4", constBiasTensor, biasInfo); Convolution2dLayer* const conv5Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv5 layer", weightInfo, biasInfo, outputInfo); + "conv5 layer", outputInfo); ConstantLayer* const weightsLayer5 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 5", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 5", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer5 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 5", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 5", constBiasTensor, biasInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); @@ -504,26 +500,26 @@ SubgraphView::SubgraphViewPtr BuildPartiallySupportedSubgraph(Graph& graph, Laye // Construct the graph Layer* const inputLayer = AddInputLayer(graph, "input layer", inputInfo); ConstantLayer* const weightsLayer1 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer1 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, biasInfo); Convolution2dLayer* const conv1Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv1 layer", weightInfo, biasInfo, outputInfo); + "conv1 layer", outputInfo); Pooling2dLayer* const pooling1Layer = AddPoolingLayer(graph, layersInGraph, poolingDescriptor, "pooling1 layer", outputInfo); Pooling2dLayer* const pooling2Layer = AddPoolingLayer(graph, layersInGraph, poolingDescriptor, "pooling2 layer", outputInfo); ConstantLayer* const weightsLayer2 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 2", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 2", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer2 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 2", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 2", constBiasTensor, biasInfo); Convolution2dLayer* const conv2Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv2 layer", weightInfo, biasInfo, outputInfo); + "conv2 layer", outputInfo); Pooling2dLayer* const pooling3Layer = AddPoolingLayer(graph, layersInGraph, poolingDescriptor, "pooling3 layer", outputInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); @@ -581,14 +577,13 @@ SubgraphView::SubgraphViewPtr BuildFullyUnoptimizableSubgraph1(Graph& graph, Lay Layer* const inputLayer = AddInputLayer(graph, "input layer", inputInfo); ConstantLayer* const weightsLayer = - AddConstantLayer(graph, layersInGraph, "Weights Layer unoptimizable", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer unoptimizable", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer = - AddConstantLayer(graph, layersInGraph, "Bias Layer unoptimizable", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer unoptimizable", constBiasTensor, biasInfo); Convolution2dLayer* const convLayer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv layer unoptimizable", weightInfo, biasInfo, - outputInfo); + "conv layer unoptimizable", outputInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); // Connect the network @@ -631,46 +626,36 @@ SubgraphView::SubgraphViewPtr BuildPartiallyOptimizableSubgraph1(Graph& graph, L Layer* const inputLayer = AddInputLayer(graph, "input layer", inputInfo); ConstantLayer* const weightsLayer1 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer1 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer2 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 2 unoptimizable", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 2 unoptimizable", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer2 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 2 unoptimizable", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 2 unoptimizable", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer3 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer3 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer4 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 4 unoptimizable", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 4 unoptimizable", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer4 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 4 unoptimizable", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 4 unoptimizable", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer5 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 5", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 5", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer5 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 5", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 5", constBiasTensor, biasInfo); Convolution2dLayer* const conv1Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv1 layer", weightInfo, biasInfo, outputInfo); - Convolution2dLayer* const conv2Layer = AddConvolutionLayer(graph, - layersInGraph, - convolutionDescriptor, - "conv2 layer unoptimizable", - weightInfo, - biasInfo, - outputInfo); + "conv1 layer", outputInfo); + Convolution2dLayer* const conv2Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, + "conv2 layer unoptimizable", outputInfo); Convolution2dLayer* const conv3Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv3 layer", weightInfo, biasInfo, outputInfo); - Convolution2dLayer* const conv4Layer = AddConvolutionLayer(graph, - layersInGraph, - convolutionDescriptor, - "conv4 layer unoptimizable", - weightInfo, - biasInfo, - outputInfo); + "conv3 layer", outputInfo); + Convolution2dLayer* const conv4Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, + "conv4 layer unoptimizable", outputInfo); Convolution2dLayer* const conv5Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv5 layer", weightInfo, biasInfo, outputInfo); + "conv5 layer", outputInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); @@ -747,25 +732,24 @@ SubgraphView::SubgraphViewPtr BuildPartiallyOptimizableSubgraph2(Graph& graph, L Layer* const input2Layer = AddInputLayer(graph, "input2 layer", inputInfo, 1); ConstantLayer* const weightsLayer1 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 1", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer1 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 1", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer2 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 2 unoptimizable", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 2 unoptimizable", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer2 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 2 unoptimizable", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 2 unoptimizable", constBiasTensor, biasInfo); ConstantLayer* const weightsLayer3 = - AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Weights Layer 3", constWeightsTensor, weightInfo); ConstantLayer* const biasLayer3 = - AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, outputInfo); + AddConstantLayer(graph, layersInGraph, "Bias Layer 3", constBiasTensor, biasInfo); Convolution2dLayer* const conv1Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv1 layer", weightInfo, biasInfo, outputInfo); + "conv1 layer", outputInfo); Convolution2dLayer* const conv2Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv2 layer unoptimizable", weightInfo, biasInfo, - outputInfo); + "conv2 layer unoptimizable", outputInfo); Convolution2dLayer* const conv3Layer = AddConvolutionLayer(graph, layersInGraph, convolutionDescriptor, - "conv3 layer", weightInfo, biasInfo, outputInfo); + "conv3 layer", outputInfo); AdditionLayer* const addLayer = AddAdditionaLayer(graph, layersInGraph, "add layer", outputInfo); Layer* const outputLayer = AddOutputLayer(graph, "output layer"); diff --git a/src/backends/backendsCommon/test/WorkloadDataValidation.cpp b/src/backends/backendsCommon/test/WorkloadDataValidation.cpp index fed21eb911..2a09f6508f 100644 --- a/src/backends/backendsCommon/test/WorkloadDataValidation.cpp +++ b/src/backends/backendsCommon/test/WorkloadDataValidation.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -158,13 +158,8 @@ TEST_CASE("FullyConnectedQueueDescriptor_Validate_RequiredDataMissing") FullyConnectedQueueDescriptor invalidData; WorkloadInfo invalidInfo; - ScopedTensorHandle weightTensor(weightsDesc); - ScopedTensorHandle biasTensor(biasesDesc); - AddInputToWorkload(invalidData, invalidInfo, inputTensorInfo, nullptr); AddOutputToWorkload(invalidData, invalidInfo, outputTensorInfo, nullptr); - invalidData.m_Weight = &weightTensor; - invalidData.m_Bias = &biasTensor; invalidData.m_Parameters.m_BiasEnabled = true; invalidData.m_Parameters.m_TransposeWeightMatrix = false; @@ -678,16 +673,10 @@ TEST_CASE("BiasPerAxisQuantization_ValidateCorrectValues") AddInputToWorkload(queueDescriptor, workloadInfo, weightInfo, nullptr); AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, nullptr); - ScopedTensorHandle weightTensor(weightInfo); - queueDescriptor.m_Weight = &weightTensor; - // Test 1: correct per-axis quantization values const std::vector<float> biasPerAxisScales1 = { 3.75f, 5.25f }; const TensorInfo biasInfo1(biasShape, biasType, biasPerAxisScales1, 0); - ScopedTensorHandle biasHandle1(biasInfo1); - queueDescriptor.m_Bias = &biasHandle1; - AddInputToWorkload(queueDescriptor, workloadInfo, biasInfo1, nullptr); CHECK_NOTHROW(queueDescriptor.Validate(workloadInfo)); @@ -729,16 +718,10 @@ TEST_CASE("BiasPerAxisQuantization_ValidateIncorrectValues") AddInputToWorkload(queueDescriptor, workloadInfo, weightInfo, nullptr); AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, nullptr); - ScopedTensorHandle weightTensor(weightInfo); - queueDescriptor.m_Weight = &weightTensor; - - // Test 2: wrong per-axis quantization values + // Test 2: wrong per-axis quantization values const std::vector<float> biasPerAxisScales2 = { 4.00f, 5.00f }; const TensorInfo biasInfo2(biasShape, biasType, biasPerAxisScales2, 0); - ScopedTensorHandle biasHandle2(biasInfo2); - queueDescriptor.m_Bias = &biasHandle2; - AddInputToWorkload(queueDescriptor, workloadInfo, biasInfo2, nullptr); CHECK_NOTHROW(queueDescriptor.Validate(workloadInfo)); @@ -781,16 +764,10 @@ TEST_CASE("BiasPerAxisQuantization_ValidateInvalidArgumentException") AddInputToWorkload(queueDescriptor, workloadInfo, weightInfo, nullptr); AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, nullptr); - ScopedTensorHandle weightTensor(weightInfo); - queueDescriptor.m_Weight = &weightTensor; - // Test 3: mismatched number of quantization scales const std::vector<float> biasPerAxisScales3 = { 3.75f, 5.25f, 5.25f }; const TensorInfo biasInfo3(biasShape, biasType, biasPerAxisScales3, 0); - ScopedTensorHandle biasHandle3(biasInfo3); - queueDescriptor.m_Bias = &biasHandle3; - AddInputToWorkload(queueDescriptor, workloadInfo, biasInfo3, nullptr); CHECK_THROWS_AS(queueDescriptor.Validate(workloadInfo), InvalidArgumentException); diff --git a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp index 1e0adc169a..8a2d8c8696 100644 --- a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp +++ b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -313,8 +313,6 @@ LayerTestResult<T, 4> SimpleConvolution2dTestImpl( armnn::Convolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); // Permute the kernel if necessary std::vector<T> kernel = originalKernel; @@ -322,12 +320,6 @@ LayerTestResult<T, 4> SimpleConvolution2dTestImpl( { armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data(), sizeof(T)); } - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernel.data()); - - if(biasEnabled) - { - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - } AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); @@ -339,8 +331,6 @@ LayerTestResult<T, 4> SimpleConvolution2dTestImpl( biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padLeft; @@ -442,15 +432,13 @@ LayerTestResult<O, 4> SimpleConvolution2dNhwcTestImpl( std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernel.data()); +// armnn::ScopedTensorHandle weightsTensor(kernelDesc); +// AllocateAndCopyDataToITensorHandle(&weightsTensor, kernel.data()); - armnn::ScopedTensorHandle biasTensor(biasDesc); +// armnn::ScopedTensorHandle biasTensor(biasDesc); armnn::Convolution2dQueueDescriptor data; - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padLeft; @@ -590,18 +578,16 @@ LayerTestResult<T,4> Convolution1dTestImpl( armnn::Convolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelInfo); - armnn::ScopedTensorHandle biasTensor(biasInfo); - - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); +// armnn::ScopedTensorHandle weightsTensor(kernelInfo); +// armnn::ScopedTensorHandle biasTensor(biasInfo); +// +// AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); +// AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); AddInputToWorkload(data, info, inputInfo, inputHandle.get()); AddInputToWorkload(data, info, kernelInfo, weightsHandle.get()); AddOutputToWorkload(data, info, outputInfo, outputHandle.get()); - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; data.m_Parameters.m_StrideX = 1; data.m_Parameters.m_StrideY = stride; data.m_Parameters.m_PadLeft = 0; @@ -1421,25 +1407,14 @@ LayerTestResult<T,4> CompareConvolution2dTestImpl( armnn::Convolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); AddInputToWorkload(data, info, biasDesc, biasHandle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernel.data()); - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernel.data()); AllocateAndCopyDataToITensorHandle(biasHandle.get(), bias.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padX; @@ -1827,15 +1802,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl( armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // Needed in Neon and Cl Workload when permuting. Backend TensorHandle in (2) below will not work. - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. - // Cannot PolymorphicDowncast from ScopedTensorHandle->RefTensorHandle. - // Need to PolymorphicDowncast from ITensorHandle->RefTensorHandle. - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernelData.data()); // required for ConstantTensor AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); @@ -1852,8 +1818,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl( AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs. data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padLeft; @@ -2009,12 +1973,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dDepthMul1TestImpl( armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); // required for QueueDescriptor AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernelData.data()); // required for ConstantTensor AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); @@ -2031,8 +1989,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dDepthMul1TestImpl( AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. data.m_Parameters.m_StrideX = 1; data.m_Parameters.m_StrideY = 1; data.m_Parameters.m_PadLeft = 0; @@ -2234,12 +2190,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); // required for QueueDescriptor AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernelData.data()); // required for ConstantTensor AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); @@ -2256,8 +2206,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. data.m_Parameters.m_StrideX = 2; data.m_Parameters.m_StrideY = 1; data.m_Parameters.m_PadLeft = 0; @@ -2426,12 +2374,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); // required for QueueDescriptor AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernelData.data()); // required for ConstantTensor AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); @@ -2448,8 +2390,6 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padLeft; @@ -3177,25 +3117,14 @@ LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl( armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(aclKernelDescriptor); - armnn::ScopedTensorHandle biasTensor(biasDesc); - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); AddInputToWorkload(data, info, aclKernelDescriptor, weightsHandle.get()); AddInputToWorkload(data, info, biasDesc, biasHandle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. AllocateAndCopyDataToITensorHandle(weightsHandle.get(), aclKernelData.data()); - AllocateAndCopyDataToITensorHandle(&weightsTensor, aclKernelData.data()); AllocateAndCopyDataToITensorHandle(biasHandle.get(), bias.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; data.m_Parameters.m_StrideX = strideX; data.m_Parameters.m_StrideY = strideY; data.m_Parameters.m_PadLeft = padX; @@ -3696,16 +3625,14 @@ LayerTestResult<uint8_t, 4> Convolution2dPerAxisQuantTest( std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; WorkloadInfo workloadInfo; - ScopedTensorHandle weightTensor(kernelInfo); - ScopedTensorHandle biasTensor(biasInfo); - - AllocateAndCopyDataToITensorHandle(&weightTensor, kernelData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); +// ScopedTensorHandle weightTensor(kernelInfo); +// ScopedTensorHandle biasTensor(biasInfo); +// +// AllocateAndCopyDataToITensorHandle(&weightTensor, kernelData.data()); +// AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); Convolution2dQueueDescriptor queueDescriptor; queueDescriptor.m_Parameters = descriptor; - queueDescriptor.m_Weight = &weightTensor; - queueDescriptor.m_Bias = &biasTensor; AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get()); AddInputToWorkload(queueDescriptor, workloadInfo, kernelInfo, weightsHandle.get()); @@ -3995,26 +3922,16 @@ LayerTestResult<uint8_t, 4> DepthwiseConvolution2dPerAxisQuantTest( DepthwiseConvolution2dQueueDescriptor queueDescriptor; WorkloadInfo workloadInfo; - ScopedTensorHandle weightTensor(kernelInfo); - ScopedTensorHandle biasTensor(biasInfo); AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get()); AddInputToWorkload(queueDescriptor, workloadInfo, kernelInfo, weightsHandle.get()); AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get()); AddInputToWorkload(queueDescriptor, workloadInfo, biasInfo, biasHandle.get()); - // AllocateAndCopyDataToITensorHandle() is required twice for the weights AND biases: - // See comment in DepthwiseConvolution2dAsymmetricTestImpl() for reasons. - // 1) ScopedTensorHandle (weightsTensor) required for QueueDescriptor (data.m_Weight). - // 2) ITensorHandle (converts to Backend TensorHandle) required in RefWorkload for GetTensorInfo() method. AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernelData.data()); - AllocateAndCopyDataToITensorHandle(&weightTensor, kernelData.data()); AllocateAndCopyDataToITensorHandle(biasHandle.get(), biasData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); queueDescriptor.m_Parameters = descriptor; - queueDescriptor.m_Weight = &weightTensor; - queueDescriptor.m_Bias = &biasTensor; std::unique_ptr<IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, queueDescriptor, diff --git a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp index 5555772c5a..f7519a73bc 100644 --- a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp +++ b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -44,20 +44,11 @@ LayerTestResult<T, 2> SimpleFullyConnectedTestImpl( armnn::FullyConnectedQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(weightsTensorInfo); - armnn::ScopedTensorHandle biasTensor(biasesTensorInfo); - - AllocateAndCopyDataToITensorHandle(&weightsTensor, weights.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); AddInputToWorkload(data, info, inputTensorInfo, input0Handle.get()); AddInputToWorkload(data, info, weightsTensorInfo, input1Handle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - // Need to set as layer members will be null when creating the workload because the optimization hasn't been run. - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; - data.m_Parameters.m_BiasEnabled = biasEnabled; data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; data.m_Parameters.m_ConstantWeights = constantWeights; diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp index 3a757f8820..adea733582 100644 --- a/src/backends/cl/test/ClCreateWorkloadTests.cpp +++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -477,16 +477,9 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "CreateConvolution2dClCompiledContext WorkloadInfo workloadInfo; - ScopedTensorHandle weightTensor(kernelInfo); - ScopedTensorHandle biasTensor(biasInfo); - - AllocateAndCopyDataToITensorHandle(&weightTensor, kernelData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); Convolution2dQueueDescriptor queueDescriptor; queueDescriptor.m_Parameters = descriptor; - queueDescriptor.m_Weight = &weightTensor; - queueDescriptor.m_Bias = &biasTensor; AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get()); AddInputToWorkload(queueDescriptor, workloadInfo, kernelInfo, weightsHandle.get()); diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp index 6b0a3b8352..1920f2d20b 100644 --- a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp +++ b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp @@ -88,16 +88,16 @@ ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescrip arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); arm_compute::ICLTensor& weights = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor* bias = nullptr; if (m_Data.m_Parameters.m_BiasEnabled) { - bias = &static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor(); + arm_compute::ICLTensor& bias = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor(); + m_BiasProxy = std::make_unique<ICLTensorProxy>(&bias); } // Create Proxy tensor and set the initial tensor handle to it m_InputProxy = std::make_unique<ICLTensorProxy>(&input); m_OutputProxy = std::make_unique<ICLTensorProxy>(&output); - + m_WeightsProxy = std::make_unique<ICLTensorProxy>(&weights); arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); input.info()->set_data_layout(aclDataLayout); @@ -112,8 +112,8 @@ ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescrip ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload_configure"); m_ConvolutionLayer.configure(clCompileContext, m_InputProxy.get(), - &weights, - bias, + m_WeightsProxy.get(), + m_BiasProxy.get(), m_OutputProxy.get(), padStrideInfo, arm_compute::WeightsInfo(), @@ -138,11 +138,11 @@ ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescrip detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod)); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp index 7293c830ac..2b46bb9c04 100644 --- a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp +++ b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -63,6 +63,8 @@ private: arm_compute::ConvolutionMethod m_ConvolutionMethod; std::unique_ptr<ICLTensorProxy> m_InputProxy; + std::unique_ptr<ICLTensorProxy> m_WeightsProxy; + std::unique_ptr<ICLTensorProxy> m_BiasProxy; std::unique_ptr<ICLTensorProxy> m_OutputProxy; }; diff --git a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp index 42fe400041..041cb8b0fc 100644 --- a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp +++ b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -96,10 +96,10 @@ ClDepthwiseConvolutionWorkload::ClDepthwiseConvolutionWorkload( detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details diff --git a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp b/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp index 0e1efe0239..22df04fc76 100644 --- a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp +++ b/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -65,10 +65,10 @@ ClFullyConnectedWorkload::ClFullyConnectedWorkload( detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details diff --git a/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp b/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp index 586b9c9849..879bb747a1 100644 --- a/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp +++ b/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -90,11 +90,11 @@ NeonConvolution2dWorkload::NeonConvolution2dWorkload( output.info()->set_data_layout(aclDataLayout); m_KernelTensor = std::make_unique<arm_compute::Tensor>(); - BuildArmComputeTensor(*m_KernelTensor, m_Data.m_Weight->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); + BuildArmComputeTensor(*m_KernelTensor, info.m_InputTensorInfos[1], m_Data.m_Parameters.m_DataLayout); if (m_Data.m_Parameters.m_BiasEnabled) { m_BiasTensor = std::make_unique<arm_compute::Tensor>(); - BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); + BuildArmComputeTensor(*m_BiasTensor, info.m_InputTensorInfos[2], m_Data.m_Parameters.m_DataLayout); } arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters); @@ -130,11 +130,12 @@ NeonConvolution2dWorkload::NeonConvolution2dWorkload( detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod)); + if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details @@ -146,22 +147,31 @@ NeonConvolution2dWorkload::NeonConvolution2dWorkload( m_ConvolutionLayer.reset(convolutionLayer.release()); ARMNN_ASSERT(m_ConvolutionLayer); - - InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight); + m_KernelTensorInfo = info.m_InputTensorInfos[1]; if (m_Data.m_Parameters.m_BiasEnabled) { - InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias); + m_BiasTensorInfo = info.m_InputTensorInfos[2]; } - - m_ConvolutionLayer->prepare(); - FreeTensorIfUnused(m_KernelTensor); - FreeTensorIfUnused(m_BiasTensor); } void NeonConvolution2dWorkload::Execute() const { ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution2dWorkload_Execute", this->GetGuid()); + // The constant tensors may not be fully in place until the workload is Executed + if (!prepared) + { + InitializeArmComputeTensorData(*m_KernelTensor, m_KernelTensorInfo, m_Data.m_Inputs[1]); + + if (m_Data.m_Parameters.m_BiasEnabled) + { + InitializeArmComputeTensorData(*m_BiasTensor, m_BiasTensorInfo, m_Data.m_Inputs[2]); + } + m_ConvolutionLayer->prepare(); + FreeTensorIfUnused(m_KernelTensor); + FreeTensorIfUnused(m_BiasTensor); + prepared = true; + } m_ConvolutionLayer->run(); } diff --git a/src/backends/neon/workloads/NeonConvolution2dWorkload.hpp b/src/backends/neon/workloads/NeonConvolution2dWorkload.hpp index e833f2ac66..4e5c1cf5ed 100644 --- a/src/backends/neon/workloads/NeonConvolution2dWorkload.hpp +++ b/src/backends/neon/workloads/NeonConvolution2dWorkload.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -41,10 +41,12 @@ public: private: std::unique_ptr<arm_compute::IFunction> m_ConvolutionLayer; - std::unique_ptr<arm_compute::Tensor> m_KernelTensor; - std::unique_ptr<arm_compute::Tensor> m_BiasTensor; - + mutable std::unique_ptr<arm_compute::Tensor> m_KernelTensor; + mutable std::unique_ptr<arm_compute::Tensor> m_BiasTensor; + TensorInfo m_KernelTensorInfo; + TensorInfo m_BiasTensorInfo; arm_compute::ConvolutionMethod m_ConvolutionMethod; + mutable bool prepared = false; }; } //namespace armnn diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp index e2d0a8200f..00d9d3340e 100644 --- a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp +++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -141,10 +141,10 @@ NeonDepthwiseConvolutionWorkload::NeonDepthwiseConvolutionWorkload( detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details diff --git a/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp b/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp index 0b91eb37c2..7bb23f870b 100644 --- a/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp +++ b/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -70,15 +70,15 @@ NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueue // Copy the weights' tensor into arm_compute tensor. m_WeightsTensor = std::make_unique<arm_compute::Tensor>(); - BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo()); - InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight); - + m_WeightsTensorInfo = info.m_InputTensorInfos[1]; + BuildArmComputeTensor(*m_WeightsTensor, m_WeightsTensorInfo); + if (m_Data.m_Parameters.m_BiasEnabled) { // Copy the biases tensor into arm_compute tensor. m_BiasesTensor = std::make_unique<arm_compute::Tensor>(); - BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo()); - InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias); + m_BiasesTensorInfo = info.m_InputTensorInfos[2]; + BuildArmComputeTensor(*m_BiasesTensor, m_BiasesTensorInfo); } const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); @@ -94,10 +94,10 @@ NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueue detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; - detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); } // Report Profiling Details @@ -107,14 +107,25 @@ NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueue this->GetGuid()); // Force Compute Library to perform the necessary copying and reshaping. - m_FullyConnectedLayer->prepare(); - FreeTensorIfUnused(m_WeightsTensor); - FreeTensorIfUnused(m_BiasesTensor); } void NeonFullyConnectedWorkload::Execute() const { ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonFullyConnectedWorkload_Execute", this->GetGuid()); + // The constant tensors may not be fully in place until the workload is Executed + if (!prepared) + { + InitializeArmComputeTensorData(*m_WeightsTensor, m_WeightsTensorInfo, m_Data.m_Inputs[1]); + + if (m_Data.m_Parameters.m_BiasEnabled) + { + InitializeArmComputeTensorData(*m_BiasesTensor, m_BiasesTensorInfo, m_Data.m_Inputs[2]); + } + m_FullyConnectedLayer->prepare(); + FreeTensorIfUnused(m_WeightsTensor); + FreeTensorIfUnused(m_BiasesTensor); + prepared = true; + } m_FullyConnectedLayer->run(); } diff --git a/src/backends/neon/workloads/NeonFullyConnectedWorkload.hpp b/src/backends/neon/workloads/NeonFullyConnectedWorkload.hpp index 944731d7bd..d5bb932c6e 100644 --- a/src/backends/neon/workloads/NeonFullyConnectedWorkload.hpp +++ b/src/backends/neon/workloads/NeonFullyConnectedWorkload.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -29,13 +29,16 @@ class NeonFullyConnectedWorkload : public NeonBaseWorkload<FullyConnectedQueueDe public: NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); + virtual void Execute() const override; private: std::unique_ptr<arm_compute::IFunction> m_FullyConnectedLayer; - std::unique_ptr<arm_compute::Tensor> m_WeightsTensor; - std::unique_ptr<arm_compute::Tensor> m_BiasesTensor; - + mutable std::unique_ptr<arm_compute::Tensor> m_WeightsTensor; + mutable std::unique_ptr<arm_compute::Tensor> m_BiasesTensor; + TensorInfo m_WeightsTensorInfo; + TensorInfo m_BiasesTensorInfo; + mutable bool prepared = false; }; } //namespace armnn diff --git a/src/backends/neon/workloads/NeonWorkloadUtils.hpp b/src/backends/neon/workloads/NeonWorkloadUtils.hpp index f9c3718e14..9f8bb9540e 100644 --- a/src/backends/neon/workloads/NeonWorkloadUtils.hpp +++ b/src/backends/neon/workloads/NeonWorkloadUtils.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -58,6 +58,42 @@ void CopyArmComputeTensorData(arm_compute::Tensor& dstTensor, const T* srcData) } inline void InitializeArmComputeTensorData(arm_compute::Tensor& tensor, + TensorInfo tensorInfo, + const ITensorHandle* handle) +{ + ARMNN_ASSERT(handle); + + switch(tensorInfo.GetDataType()) + { + case DataType::Float16: + CopyArmComputeTensorData(tensor, reinterpret_cast<const armnn::Half*>(handle->Map())); + break; + case DataType::Float32: + CopyArmComputeTensorData(tensor, reinterpret_cast<const float*>(handle->Map())); + break; + case DataType::QAsymmU8: + CopyArmComputeTensorData(tensor, reinterpret_cast<const uint8_t*>(handle->Map())); + break; + case DataType::QSymmS8: + case DataType::QAsymmS8: + CopyArmComputeTensorData(tensor, reinterpret_cast<const int8_t*>(handle->Map())); + break; + case DataType::Signed32: + CopyArmComputeTensorData(tensor, reinterpret_cast<const int32_t*>(handle->Map())); + break; + case DataType::QSymmS16: + CopyArmComputeTensorData(tensor, reinterpret_cast<const int16_t*>(handle->Map())); + break; + case DataType::BFloat16: + CopyArmComputeTensorData(tensor, reinterpret_cast<const armnn::BFloat16*>(handle->Map())); + break; + default: + // Throw exception; assertion not called in release build. + throw Exception("Unexpected tensor type during InitializeArmComputeTensorData()."); + } +}; + +inline void InitializeArmComputeTensorData(arm_compute::Tensor& tensor, const ConstTensorHandle* handle) { ARMNN_ASSERT(handle); |