diff options
Diffstat (limited to 'src/backends')
13 files changed, 503 insertions, 177 deletions
diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp index 74ab789402..de3a34ee08 100644 --- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp +++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp @@ -263,28 +263,13 @@ LayerType* FuseDepthwiseConvolution2dLayer(OptimizationViews& optimizationViews, ActivationDescriptor& activationDesc, std::string name) { - std::shared_ptr<ConstTensorHandle> weightHandle = baseLayer->m_Weight; - TensorInfo weightInfo = weightHandle->GetTensorInfo(); - - std::shared_ptr<ConstTensorHandle> biasHandle = baseLayer->m_Bias; - ConstTensor biasTensor; - if (!biasHandle) - { - biasTensor = ConstTensor(); - } - else - { - biasTensor = ConstTensor(biasHandle->GetTensorInfo(), biasHandle->Map(true)); - } - - IConnectableLayer* replacement = - optimizationViews.GetINetwork()-> - AddDepthwiseConvolution2dLayer(baseLayer->GetParameters(), - ConstTensor(weightInfo, weightHandle->Map(true)), - Optional<ConstTensor>(biasTensor), - name.c_str()); + IConnectableLayer* replacement = optimizationViews.GetINetwork()-> + AddDepthwiseConvolution2dLayer(baseLayer->GetParameters(), name.c_str()); LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement); + replacementLayer->m_Weight = std::move(baseLayer->m_Weight); + replacementLayer->m_Bias = std::move(baseLayer->m_Bias); + FuseLayer(optimizationViews, baseLayer, replacementLayer, diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp index d89b5899ba..7a46741964 100644 --- a/src/backends/backendsCommon/WorkloadData.cpp +++ b/src/backends/backendsCommon/WorkloadData.cpp @@ -1382,7 +1382,13 @@ void DepthwiseConvolution2dQueueDescriptor::Validate(const WorkloadInfo& workloa { const std::string descriptorName{"DepthwiseConvolution2dQueueDescriptor"}; - ValidateNumInputs(workloadInfo, descriptorName, 1); + uint32_t numInputs = 2; + if (m_Parameters.m_BiasEnabled) + { + numInputs = 3; + } + + ValidateNumInputs(workloadInfo, descriptorName, numInputs); ValidateNumOutputs(workloadInfo, descriptorName, 1); const TensorInfo& inputTensorInfo = workloadInfo.m_InputTensorInfos[0]; @@ -1391,9 +1397,7 @@ void DepthwiseConvolution2dQueueDescriptor::Validate(const WorkloadInfo& workloa ValidateTensorNumDimensions(inputTensorInfo, descriptorName, 4, "input"); ValidateTensorNumDimensions(outputTensorInfo, descriptorName, 4, "output"); - ValidatePointer(m_Weight, descriptorName, "weight"); - - const TensorInfo& weightTensorInfo = m_Weight->GetTensorInfo(); + const TensorInfo& weightTensorInfo = workloadInfo.m_InputTensorInfos[1]; ValidateTensorNumDimensions(weightTensorInfo, descriptorName, 4, "weight"); if (m_Parameters.m_DilationX < 1 || m_Parameters.m_DilationY < 1 ) @@ -1447,9 +1451,7 @@ void DepthwiseConvolution2dQueueDescriptor::Validate(const WorkloadInfo& workloa Optional<TensorInfo> optionalBiasTensorInfo; if (m_Parameters.m_BiasEnabled) { - ValidatePointer(m_Bias, descriptorName, "bias"); - - optionalBiasTensorInfo = MakeOptional<TensorInfo>(m_Bias->GetTensorInfo()); + optionalBiasTensorInfo = MakeOptional<TensorInfo>(workloadInfo.m_InputTensorInfos[2]); const TensorInfo& biasTensorInfo = optionalBiasTensorInfo.value(); ValidateBiasTensorQuantization(biasTensorInfo, inputTensorInfo, weightTensorInfo, descriptorName); diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp index 5847e8cc21..f624ee6021 100644 --- a/src/backends/backendsCommon/WorkloadFactory.cpp +++ b/src/backends/backendsCommon/WorkloadFactory.cpp @@ -324,10 +324,13 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, case LayerType::DepthwiseConvolution2d: { auto cLayer = PolymorphicDowncast<const DepthwiseConvolution2dLayer*>(&layer); - const TensorInfo& input = OverrideDataType(layer.GetInputSlot(0).GetConnection()->GetTensorInfo(), - dataType); - const TensorInfo& output = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType); - ARMNN_ASSERT(cLayer->m_Weight.get() != nullptr); + const TensorInfo& input = OverrideDataType(layer.GetInputSlot(0).GetConnection()->GetTensorInfo(), + dataType); + const TensorInfo& output = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType); + const TensorInfo& weights = OverrideDataType(layer.GetInputSlot(1).GetConnection()->GetTensorInfo(), + dataType); + + ARMNN_ASSERT(cLayer->GetInputSlot(1).GetConnection() != nullptr); const DepthwiseConvolution2dDescriptor& descriptor = cLayer->GetParameters(); @@ -335,17 +338,16 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, Optional<TensorInfo> biases; if (descriptor.m_BiasEnabled) { - biases = - OverrideDataType(cLayer->m_Bias->GetTensorInfo(), GetBiasTypeFromWeightsType(dataType)); + biases = OverrideDataType(cLayer->GetInputSlot(2).GetConnection()->GetTensorInfo(), + GetBiasTypeFromWeightsType(dataType)); } - result = layerSupportObject.IsDepthwiseConvolutionSupported( - input, - output, - descriptor, - OverrideDataType(cLayer->m_Weight->GetTensorInfo(), dataType), - biases, - reason); + result = layerSupportObject.IsDepthwiseConvolutionSupported(input, + output, + descriptor, + weights, + biases, + reason); break; } case LayerType::Dequantize: diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt index 06d230b006..991f37d17e 100644 --- a/src/backends/backendsCommon/test/CMakeLists.txt +++ b/src/backends/backendsCommon/test/CMakeLists.txt @@ -21,6 +21,7 @@ list(APPEND armnnBackendsCommonUnitTests_sources DataTypeUtils.hpp DefaultAsyncExecuteTest.cpp DepthToSpaceEndToEndTestImpl.hpp + DepthwiseConvolution2dEndToEndTests.hpp DequantizeEndToEndTestImpl.hpp DetectionPostProcessEndToEndTestImpl.hpp DynamicBackendTests.cpp diff --git a/src/backends/backendsCommon/test/DepthwiseConvolution2dEndToEndTests.hpp b/src/backends/backendsCommon/test/DepthwiseConvolution2dEndToEndTests.hpp new file mode 100644 index 0000000000..1f9b60a4f2 --- /dev/null +++ b/src/backends/backendsCommon/test/DepthwiseConvolution2dEndToEndTests.hpp @@ -0,0 +1,183 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "EndToEndTestImpl.hpp" +#include <armnnUtils/QuantizeHelper.hpp> + +#include <ResolveType.hpp> + +#include <CommonTestUtils.hpp> +#include <armnnTestUtils/DataLayoutUtils.hpp> + +#include <map> +#include <vector> + +namespace +{ + +armnn::INetworkPtr CreateDepthwiseConvolution2dNetwork(const armnn::DepthwiseConvolution2dDescriptor& descriptor, + const armnn::TensorInfo& inputInfo, + const armnn::TensorInfo& weightsInfo, + const armnn::TensorInfo& biasInfo, + const armnn::TensorInfo& outputInfo, + const armnn::ConstTensor& weights, + const armnn::ConstTensor& biases) +{ + using namespace armnn; + + INetworkPtr network(INetwork::Create()); + IConnectableLayer* input = network->AddInputLayer(0, "input"); + armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights"); + armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biases, "Bias"); + IConnectableLayer* convolution2d = network->AddDepthwiseConvolution2dLayer(descriptor, "depthwiseConvolution2d"); + IConnectableLayer* output = network->AddOutputLayer(0, "output"); + + Connect(input, convolution2d, inputInfo, 0, 0); + Connect(weightsLayer, convolution2d, weightsInfo, 0, 1); + Connect(biasLayer, convolution2d, biasInfo, 0, 2); + Connect(convolution2d, output, outputInfo, 0, 0); + + return network; +} + +} // anonymous namespace + +template<armnn::DataType ArmnnType, armnn::DataType ArmnnBType> +void DepthwiseConvolution2dEndToEnd(const std::vector<armnn::BackendId>& backends, + armnn::DataLayout dataLayout) +{ + using namespace armnn; + using T = ResolveType<ArmnnType>; + using BT = ResolveType<ArmnnBType>; + + const float qScale = IsQuantizedType<T>() ? 0.25f : 1.0f; + const int32_t qOffset = IsQuantizedType<T>() ? 50 : 0; + + unsigned int depthMultiplier = 2; + + unsigned int inputHeight = 8; + unsigned int inputWidth = 16; + unsigned int inputChannels = 2; + unsigned int inputBatchSize = 1; + + unsigned int kernelHeight = 5; + unsigned int kernelWidth = 3; + + unsigned int outputHeight = inputHeight - kernelHeight + 1 + 2; + unsigned int outputWidth = (inputWidth - kernelWidth + 1)/2; + unsigned int outputChannels = inputChannels * depthMultiplier; + unsigned int outputBatchSize = inputBatchSize; + + TensorInfo inputInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, ArmnnType, qScale, qOffset, true); + TensorInfo outputInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, ArmnnType, qScale, qOffset); + TensorInfo weightsInfo({1, kernelHeight, kernelWidth, outputChannels}, ArmnnType, qScale, qOffset, true); + TensorInfo biasesInfo({outputChannels}, ArmnnBType, qScale * qScale, 0, true); + + std::vector<float> inputData = + { + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f + }; + + std::vector<float> weightsData = + { + 1.0f, 1.0f, 1.0f, + 1.0f, -1.0f, 1.0f, + 1.0f, 1.0f, 1.0f, + 1.0f, 1.0f, 1.0f, + 1.0f, 1.0f, 1.0f, + + 2.0f, 2.0f, 2.0f, + 2.0f, 2.0f, 2.0f, + 2.0f, 2.0f, 2.0f, + 2.0f, 2.0f, 2.0f, + 2.0f, 2.0f, 2.0f, + + 0.0f, 0.0f, 0.0f, + 0.0f, -1.0f, 0.0f, + 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, + + 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, + 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f + }; + + std::vector<float> biasesData = { 0.0f, 2.0f, 1.0f, -1.0f }; + + std::vector<float> expectedOutputData = + { + 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, 3.0f, + 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, + 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.5f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, + 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 2.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, + 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 4.5f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, + 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, + 1.0f, 3.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 3.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f + }; + + DepthwiseConvolution2dDescriptor descriptor; + descriptor.m_PadLeft = 0; + descriptor.m_PadRight = 0; + descriptor.m_PadTop = 1; + descriptor.m_PadBottom = 0; + descriptor.m_StrideX = 2; + descriptor.m_StrideY = 1; + descriptor.m_BiasEnabled = true; + descriptor.m_DataLayout = dataLayout; + + // Permute input and output if NCDHW. + if (dataLayout == DataLayout::NCHW) + { + PermuteTensorNhwcToNchw(inputInfo, inputData); + PermuteTensorNhwcToNchw(outputInfo, expectedOutputData); + } + + // Quantize data + std::vector<T> qInputData = armnnUtils::QuantizedVector<T>(inputData, qScale, qOffset); + std::vector<T> qWeightsData = armnnUtils::QuantizedVector<T>(weightsData, qScale, qOffset); + std::vector<T> qExpectedOutputData = armnnUtils::QuantizedVector<T>(expectedOutputData, qScale, qOffset); + + std::vector<BT> qBiasesData = armnnUtils::QuantizedVector<BT>(biasesData, qScale * qScale, 0); + + ConstTensor weights(weightsInfo, qWeightsData); + ConstTensor biases(biasesInfo, qBiasesData); + + INetworkPtr network = CreateDepthwiseConvolution2dNetwork(descriptor, + inputInfo, + weightsInfo, + biasesInfo, + outputInfo, + weights, + biases); + + EndToEndLayerTestImpl<ArmnnType, ArmnnType>(std::move(network), + { { 0, qInputData } }, + { { 0, qExpectedOutputData } }, + backends); +} diff --git a/src/backends/backendsCommon/test/OptimizationViewsTests.cpp b/src/backends/backendsCommon/test/OptimizationViewsTests.cpp index f0f5b632de..1219ac5a33 100644 --- a/src/backends/backendsCommon/test/OptimizationViewsTests.cpp +++ b/src/backends/backendsCommon/test/OptimizationViewsTests.cpp @@ -128,20 +128,16 @@ TEST_CASE("OptimizedViewsSubgraphLayerCountUsingGetINetwork") IConnectableLayer* const inputLayer = view.GetINetwork()->AddInputLayer(0, "input"); DepthwiseConvolution2dDescriptor convDescriptor; - PreCompiledDescriptor substitutionLayerDescriptor(1, 1); + PreCompiledDescriptor substitutionLayerDescriptor(2, 1); CompiledBlobPtr blobPtr; BackendId backend = Compute::CpuRef; Layer* convLayer1 = PolymorphicDowncast<Layer*>( view.GetINetwork()->AddDepthwiseConvolution2dLayer(convDescriptor, - ConstTensor(), - Optional<ConstTensor>(), "conv1")); Layer* convLayer2 = PolymorphicDowncast<Layer*>( view.GetINetwork()->AddDepthwiseConvolution2dLayer(convDescriptor, - ConstTensor(), - Optional<ConstTensor>(), "conv2")); IConnectableLayer* const outputLayer = view.GetINetwork()->AddOutputLayer(0, "output"); diff --git a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp index bd7cc40f27..4203fed23a 100644 --- a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp +++ b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp @@ -1736,19 +1736,38 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl( } std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); + 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, kernel.data()); + AllocateAndCopyDataToITensorHandle(weightsHandle.get(), kernel.data()); // required for ConstantTensor + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); armnn::ScopedTensorHandle biasTensor(biasDesc); if (biasEnabled) { AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); + + biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); + AllocateAndCopyDataToITensorHandle(biasHandle.get(), bias.data()); + AddInputToWorkload(data, info, biasDesc, biasHandle.get()); } - armnn::DepthwiseConvolution2dQueueDescriptor data; 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; @@ -1760,12 +1779,9 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl( data.m_Parameters.m_BiasEnabled = biasEnabled; data.m_Parameters.m_DataLayout = layout; - armnn::WorkloadInfo info; - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); - AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, data, info); + inputHandle->Allocate(); outputHandle->Allocate(); @@ -1890,19 +1906,35 @@ LayerTestResult<T, 4> DepthwiseConvolution2dDepthMul1TestImpl( std::vector<T> actualOutput(outputTensorInfo.GetNumElements()); std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasV.data()); + 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()); + AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + armnn::ScopedTensorHandle biasTensor(biasDesc); + if (biasEnabled) + { + AllocateAndCopyDataToITensorHandle(&biasTensor, biasV.data()); + + biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); + AllocateAndCopyDataToITensorHandle(biasHandle.get(), biasV.data()); + 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; @@ -1916,6 +1948,7 @@ LayerTestResult<T, 4> DepthwiseConvolution2dDepthMul1TestImpl( std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, data, info); + inputHandle->Allocate(); outputHandle->Allocate(); @@ -2086,19 +2119,35 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( std::vector<T> actualOutput(outputTensorInfo.GetNumElements()); std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); - AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, biasV.data()); + 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()); + AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + armnn::ScopedTensorHandle biasTensor(biasDesc); + if (biasEnabled) + { + AllocateAndCopyDataToITensorHandle(&biasTensor, biasV.data()); + + biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); + AllocateAndCopyDataToITensorHandle(biasHandle.get(), biasV.data()); + 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; @@ -2112,6 +2161,7 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, data, info); + inputHandle->Allocate(); outputHandle->Allocate(); @@ -2247,22 +2297,34 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( } std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandle = nullptr; std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); armnn::DepthwiseConvolution2dQueueDescriptor data; armnn::WorkloadInfo info; + armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); + // 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, originalKernel.data()); // required for QueueDescriptor + AllocateAndCopyDataToITensorHandle(weightsHandle.get(), originalKernel.data()); // required for ConstantTensor - AllocateAndCopyDataToITensorHandle(&weightsTensor, originalKernel.data()); + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddInputToWorkload(data, info, kernelDesc, weightsHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - if(biasEnabled) + armnn::ScopedTensorHandle biasTensor(biasDesc); + if (biasEnabled) { AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - } - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); - AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); + AllocateAndCopyDataToITensorHandle(biasHandle.get(), bias.data()); + 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. @@ -2279,6 +2341,7 @@ LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl( std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, data, info); + inputHandle->Allocate(); outputHandle->Allocate(); @@ -2970,18 +3033,30 @@ LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl( std::vector<T> expectedOutput(outputTensorInfo.GetNumElements()); std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandle = tensorHandleFactory.CreateTensorHandle(biasDesc); std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); armnn::DepthwiseConvolution2dQueueDescriptor 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()); - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); - AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); data.m_Weight = &weightsTensor; data.m_Bias = &biasTensor; data.m_Parameters.m_StrideX = strideX; @@ -2994,11 +3069,15 @@ LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl( data.m_Parameters.m_DataLayout = layout.GetDataLayout(); std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refTensorHandleFactory.CreateTensorHandle(outputTensorInfo); + std::unique_ptr<armnn::ITensorHandle> weightsHandleRef = refTensorHandleFactory.CreateTensorHandle(kernelDesc); + std::unique_ptr<armnn::ITensorHandle> biasHandleRef = refTensorHandleFactory.CreateTensorHandle(biasDesc); std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refTensorHandleFactory.CreateTensorHandle(inputTensorInfo); armnn::DepthwiseConvolution2dQueueDescriptor refData = data; armnn::WorkloadInfo refInfo = info; SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); + SetWorkloadInput(refData, refInfo, 1, kernelDesc, weightsHandleRef.get()); + SetWorkloadInput(refData, refInfo, 2, biasDesc, biasHandleRef.get()); SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); std::unique_ptr<armnn::IWorkload> workload @@ -3007,6 +3086,8 @@ LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl( = refWorkloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, refData, refInfo); outputHandleRef->Allocate(); + weightsHandleRef->Allocate(); + biasHandleRef->Allocate(); inputHandleRef->Allocate(); inputHandle->Allocate(); @@ -3014,6 +3095,8 @@ LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl( CopyDataToITensorHandle(inputHandle.get(), input.data()); CopyDataToITensorHandle(inputHandleRef.get(), input.data()); + CopyDataToITensorHandle(weightsHandleRef.get(), kernel.data()); + CopyDataToITensorHandle(biasHandleRef.get(), bias.data()); ExecuteWorkload(*workload, memoryManager); @@ -3739,23 +3822,33 @@ LayerTestResult<uint8_t, 4> DepthwiseConvolution2dPerAxisQuantTest( descriptor.m_DataLayout = layout; std::unique_ptr<ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputInfo); + std::unique_ptr<ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelInfo); + std::unique_ptr<ITensorHandle> biasHandle = tensorHandleFactory.CreateTensorHandle(biasInfo); std::unique_ptr<ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputInfo); + 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()); - DepthwiseConvolution2dQueueDescriptor queueDescriptor; queueDescriptor.m_Parameters = descriptor; queueDescriptor.m_Weight = &weightTensor; queueDescriptor.m_Bias = &biasTensor; - AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get()); - AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get()); - std::unique_ptr<IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::DepthwiseConvolution2d, queueDescriptor, workloadInfo); diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp index 018adec781..ed6f221511 100644 --- a/src/backends/cl/ClBackend.cpp +++ b/src/backends/cl/ClBackend.cpp @@ -373,14 +373,14 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph, if (baseLayer->GetParameters().m_BiasEnabled) { - biases = baseLayer->m_Bias->GetTensorInfo(); + biases = baseLayer->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo(); } arm_compute::Status status = ClDepthwiseConvolutionWorkloadValidate( baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), baseLayer->GetParameters(), - baseLayer->m_Weight->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), biases, &activationDesc); diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp index 7089f23efa..7a258c38e0 100644 --- a/src/backends/neon/NeonBackend.cpp +++ b/src/backends/neon/NeonBackend.cpp @@ -225,14 +225,14 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph if (baseLayer->GetParameters().m_BiasEnabled) { - biases = baseLayer->m_Bias->GetTensorInfo(); + biases = baseLayer->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo(); } arm_compute::Status status = NeonDepthwiseConvolutionWorkloadValidate( baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), baseLayer->GetParameters(), - baseLayer->m_Weight->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), biases, &activationDesc); diff --git a/src/backends/neon/NeonTensorHandle.hpp b/src/backends/neon/NeonTensorHandle.hpp index 9445cb1c75..fcae77cdaa 100644 --- a/src/backends/neon/NeonTensorHandle.hpp +++ b/src/backends/neon/NeonTensorHandle.hpp @@ -241,6 +241,7 @@ private: break; case arm_compute::DataType::QSYMM8: case arm_compute::DataType::QASYMM8_SIGNED: + case arm_compute::DataType::QSYMM8_PER_CHANNEL: armcomputetensorutils::CopyArmComputeITensorData(static_cast<const int8_t*>(memory), this->GetTensor()); break; diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp index 496b11db91..29df20f834 100644 --- a/src/backends/reference/test/RefLayerTests.cpp +++ b/src/backends/reference/test/RefLayerTests.cpp @@ -329,9 +329,9 @@ ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dUint8, DepthwiseConvolution2 ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2d, DepthwiseConvolution2dTest, false, DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dUint8, - DepthwiseConvolution2dUint8Test, - false, - DataLayout::NCHW) + DepthwiseConvolution2dUint8Test, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dQSymm16, DepthwiseConvolution2dInt16Test, true, DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dNhwc, DepthwiseConvolution2dTest, true, DataLayout::NHWC) @@ -339,148 +339,176 @@ ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dUint8Nhwc, DepthwiseConvolut ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dNhwc, DepthwiseConvolution2dTest, false, DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dUint8Nhwc, - DepthwiseConvolution2dUint8Test, - false, - DataLayout::NHWC) + DepthwiseConvolution2dUint8Test, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthNhwc, DepthwiseConvolution2dDepthNhwcTest, false) ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleDepthwiseConvolution2d3x3Dilation3x3Nhwc, - SimpleDepthwiseConvolution2d3x3Dilation3x3NhwcTest) + SimpleDepthwiseConvolution2d3x3Dilation3x3NhwcTest) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3Nhwc, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3BFloat16, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3NhwcBFloat16, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3Int8, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3NhwcInt8, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3Uint8, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3NhwcUint8, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3Int16, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d3x3Dilation3x3NhwcInt16, - DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3Nhwc, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::Float32, DataType::Float32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3BFloat16, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3NhwcBFloat16, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::BFloat16, DataType::BFloat16>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3Int8, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3NhwcInt8, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmS8, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3Uint8, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3NhwcUint8, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QAsymmU8, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3Int16, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, - false, - DataLayout::NCHW) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d2x3x3Dilation3x3NhwcInt16, - DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, - false, - DataLayout::NHWC) + DepthwiseConvolution2d2x3x3Dilation3x3Test<DataType::QSymmS16, DataType::Signed32>, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dMult4, - DepthwiseConvolution2dMult4Test<armnn::DataType::Float32, armnn::DataType::Float32>, - false, - armnn::DataLayout::NCHW) + DepthwiseConvolution2dMult4Test<armnn::DataType::Float32, armnn::DataType::Float32>, + false, + armnn::DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dMult2, - DepthwiseConvolution2dMult2Test<armnn::DataType::Float32, armnn::DataType::Float32>, - false, - armnn::DataLayout::NCHW) + DepthwiseConvolution2dMult2Test<armnn::DataType::Float32, armnn::DataType::Float32>, + false, + armnn::DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dMult4BFloat16, - DepthwiseConvolution2dMult4Test<armnn::DataType::BFloat16, armnn::DataType::BFloat16>, - false, - armnn::DataLayout::NCHW) + DepthwiseConvolution2dMult4Test<armnn::DataType::BFloat16, armnn::DataType::BFloat16>, + false, + armnn::DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dMult2BFloat16, - DepthwiseConvolution2dMult2Test<armnn::DataType::BFloat16, armnn::DataType::BFloat16>, - false, - armnn::DataLayout::NCHW) + DepthwiseConvolution2dMult2Test<armnn::DataType::BFloat16, armnn::DataType::BFloat16>, + false, + armnn::DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1, - DepthwiseConvolution2dDepthMul1Test, true, DataLayout::NCHW) + DepthwiseConvolution2dDepthMul1Test, + true, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1Uint8, - DepthwiseConvolution2dDepthMul1Uint8Test, true, DataLayout::NCHW) + DepthwiseConvolution2dDepthMul1Uint8Test, + true, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1Int16, - DepthwiseConvolution2dDepthMul1Int16Test, true, DataLayout::NCHW) + DepthwiseConvolution2dDepthMul1Int16Test, + true, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dDepthMul1, - DepthwiseConvolution2dDepthMul1Test, false, DataLayout::NCHW) + DepthwiseConvolution2dDepthMul1Test, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, - DepthwiseConvolution2dDepthMul1Uint8Test, false, DataLayout::NCHW) + DepthwiseConvolution2dDepthMul1Uint8Test, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1Nhwc, - DepthwiseConvolution2dDepthMul1Test, true, DataLayout::NHWC) + DepthwiseConvolution2dDepthMul1Test, + true, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1Uint8Nhwc, - DepthwiseConvolution2dDepthMul1Uint8Test, true, DataLayout::NHWC) + DepthwiseConvolution2dDepthMul1Uint8Test, + true, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dDepthMul1Nhwc, - DepthwiseConvolution2dDepthMul1Test, false, DataLayout::NHWC) + DepthwiseConvolution2dDepthMul1Test, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dDepthMul1Uint8Nhwc, - DepthwiseConvolution2dDepthMul1Uint8Test, false, DataLayout::NHWC) + DepthwiseConvolution2dDepthMul1Uint8Test, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dAsymmetric, - DepthwiseConvolution2dAsymmetricTest, true, DataLayout::NCHW) + DepthwiseConvolution2dAsymmetricTest, + true, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dAsymmetric, - DepthwiseConvolution2dAsymmetricTest, false, DataLayout::NCHW) + DepthwiseConvolution2dAsymmetricTest, + false, + DataLayout::NCHW) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dAsymmetricNhwc, - DepthwiseConvolution2dAsymmetricTest, true, DataLayout::NHWC) + DepthwiseConvolution2dAsymmetricTest, + true, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dAsymmetricNhwc, - DepthwiseConvolution2dAsymmetricTest, false, DataLayout::NHWC) + DepthwiseConvolution2dAsymmetricTest, + false, + DataLayout::NHWC) ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul64, DepthwiseConvolution2dDepthMul64Test); -ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dPerAxisQuantTestNchw, DepthwiseConvolution2dPerAxisQuantTest, - DataLayout::NCHW); -ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dPerAxisQuantTestNhwc, DepthwiseConvolution2dPerAxisQuantTest, - DataLayout::NHWC); +ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dPerAxisQuantTestNchw, + DepthwiseConvolution2dPerAxisQuantTest, + DataLayout::NCHW); +ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dPerAxisQuantTestNhwc, + DepthwiseConvolution2dPerAxisQuantTest, + DataLayout::NHWC); // [ Pooling 2D //MaxPooling diff --git a/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.cpp b/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.cpp index ad5edde7e6..c1c3916292 100644 --- a/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.cpp +++ b/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.cpp @@ -19,16 +19,41 @@ RefDepthwiseConvolution2dWorkload::RefDepthwiseConvolution2dWorkload( const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) : RefBaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info) { - m_Weight = std::make_unique<ScopedTensorHandle>(*(descriptor.m_Weight)); - const TensorInfo& rFilterInfo = m_Weight->GetTensorInfo(); - m_FilterShape = rFilterInfo.GetShape(); - m_FilterDecoder = MakeDecoder<float>(rFilterInfo, m_Weight->Map(true)); + WorkloadInfo detailsInfo; + detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; + detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; + detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - m_Bias = std::make_unique<ScopedTensorHandle>(*(descriptor.m_Bias)); - const TensorInfo& biasInfo = m_Bias->GetTensorInfo(); - m_BiasDecoder = MakeDecoder<float>(biasInfo, m_Bias->Map(true)); + detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]); + } + + // Report Profiling Details + ARMNN_REPORT_PROFILING_WORKLOAD_DESC("RefDepthwiseConvolution2dWorkload_Construct", + descriptor.m_Parameters, + detailsInfo, + this->GetGuid()); +} + +void RefDepthwiseConvolution2dWorkload::PostAllocationConfigure() +{ + PostAllocationConfigure(m_Data.m_Inputs, m_Data.m_Outputs); +} + +void RefDepthwiseConvolution2dWorkload::PostAllocationConfigure(std::vector<ITensorHandle*> inputs, + std::vector<ITensorHandle*> outputs) +{ + IgnoreUnused(outputs); + + const TensorInfo& rFilterInfo = GetTensorInfo(inputs[1]); + m_FilterShape = rFilterInfo.GetShape(); + m_FilterDecoder = MakeDecoder<float>(rFilterInfo); + + if (m_Data.m_Parameters.m_BiasEnabled) + { + const TensorInfo& biasInfo = GetTensorInfo(inputs[2]); + m_BiasDecoder = MakeDecoder<float>(biasInfo); } } @@ -39,6 +64,8 @@ void RefDepthwiseConvolution2dWorkload::Execute() const void RefDepthwiseConvolution2dWorkload::ExecuteAsync(WorkingMemDescriptor &workingMemDescriptor) { + PostAllocationConfigure(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs); + Execute(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs); } @@ -54,6 +81,12 @@ void RefDepthwiseConvolution2dWorkload::Execute(std::vector<ITensorHandle*> inpu const TensorShape& inputShape = GetTensorInfo(inputs[0]).GetShape(); const TensorShape& outputShape = GetTensorInfo(outputs[0]).GetShape(); + m_FilterDecoder->Reset(inputs[1]->Map()); + if (m_Data.m_Parameters.m_BiasEnabled) + { + m_BiasDecoder->Reset(inputs[2]->Map()); + } + Convolve(inputShape, *inputDecoder, outputShape, *OutputEncoder, m_FilterShape, *m_FilterDecoder, m_Data.m_Parameters.m_BiasEnabled, m_BiasDecoder.get(), m_Data.m_Parameters.m_DataLayout, m_Data.m_Parameters.m_PadTop, m_Data.m_Parameters.m_PadLeft, diff --git a/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.hpp b/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.hpp index 5d4b483fa7..1c7de29b37 100644 --- a/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.hpp +++ b/src/backends/reference/workloads/RefDepthwiseConvolution2dWorkload.hpp @@ -17,11 +17,13 @@ public: explicit RefDepthwiseConvolution2dWorkload(const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info); + void PostAllocationConfigure() override; void Execute() const override; void ExecuteAsync(WorkingMemDescriptor& workingMemDescriptor) override; private: + void PostAllocationConfigure(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs); void Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const; std::unique_ptr <ScopedTensorHandle> m_Weight; |