From ec67a0f08e0f96a5aebf3cac65331c67f6649f5e Mon Sep 17 00:00:00 2001 From: Mike Kelly Date: Fri, 25 Nov 2022 13:55:24 +0000 Subject: IVGCVSW-7209 Remove deprecated code due to be removed in 23.02 * Removed weights and bias from Convolution, DepthwiseConv & FullyConnected layers * Removed the weight and bias ConstTensorHandles from the QueueDescriptors * Updated Workloads to take tensors from WorkloadInfo rather than the QueueDescriptors * Removed unused RedirectMembersToConstantInputs optimization and tests. Signed-off-by: Teresa Charlin Signed-off-by: Mike Kelly Change-Id: I9ffcdc4a1c0dff725539dd69fc435b700bd98a56 --- Android.mk | 1 - CMakeLists.txt | 1 - include/armnn/backends/WorkloadData.hpp | 29 +---- src/armnn/LoadedNetwork.cpp | 2 +- src/armnn/Network.cpp | 139 +-------------------- src/armnn/Network.hpp | 43 +------ src/armnn/layers/Convolution2dLayer.cpp | 29 +---- src/armnn/layers/Convolution2dLayer.hpp | 17 ++- src/armnn/layers/DepthwiseConvolution2dLayer.cpp | 28 +---- src/armnn/layers/DepthwiseConvolution2dLayer.hpp | 18 ++- src/armnn/layers/FullyConnectedLayer.cpp | 25 +--- src/armnn/layers/FullyConnectedLayer.hpp | 12 +- src/armnn/optimizations/All.hpp | 1 - src/armnn/optimizations/FoldPadIntoLayer2d.hpp | 19 +-- src/armnn/optimizations/FuseBatchNorm.hpp | 8 +- .../RedirectMembersToConstantInputs.hpp | 90 ------------- src/armnn/test/OptimizerTests.cpp | 35 +++--- .../ConvertConstantsFloatToHalfTests.cpp | 21 ++-- .../ConvertConstantsHalfToFloatTests.cpp | 17 ++- .../Fp32NetworkToFp16ConverterTests.cpp | 15 ++- .../test/optimizations/FuseBatchNormTests.cpp | 27 ++-- .../RedirectMembersToConstantInputsTests.cpp | 85 ------------- src/armnnTestUtils/CommonTestUtils.hpp | 11 +- src/armnnTestUtils/CreateWorkload.hpp | 39 +----- src/backends/aclCommon/ArmComputeSubgraphUtils.hpp | 8 +- .../backendsCommon/test/DynamicBackendTests.hpp | 4 +- .../test/IsLayerSupportedTestImpl.hpp | 36 +++--- .../test/LayerReleaseConstantDataTest.cpp | 126 +++++++++---------- .../test/OptimizeSubgraphViewTests.cpp | 124 ++++++++---------- .../backendsCommon/test/WorkloadDataValidation.cpp | 27 +--- .../test/layerTests/Conv2dTestImpl.cpp | 111 +++------------- .../test/layerTests/FullyConnectedTestImpl.cpp | 11 +- src/backends/cl/test/ClCreateWorkloadTests.cpp | 9 +- .../cl/workloads/ClConvolution2dWorkload.cpp | 14 +-- .../cl/workloads/ClConvolution2dWorkload.hpp | 4 +- .../workloads/ClDepthwiseConvolutionWorkload.cpp | 6 +- .../cl/workloads/ClFullyConnectedWorkload.cpp | 6 +- .../neon/workloads/NeonConvolution2dWorkload.cpp | 34 +++-- .../neon/workloads/NeonConvolution2dWorkload.hpp | 10 +- .../workloads/NeonDepthwiseConvolutionWorkload.cpp | 6 +- .../neon/workloads/NeonFullyConnectedWorkload.cpp | 33 +++-- .../neon/workloads/NeonFullyConnectedWorkload.hpp | 11 +- src/backends/neon/workloads/NeonWorkloadUtils.hpp | 38 +++++- 43 files changed, 374 insertions(+), 956 deletions(-) delete mode 100644 src/armnn/optimizations/RedirectMembersToConstantInputs.hpp delete mode 100644 src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp diff --git a/Android.mk b/Android.mk index e9f70c9417..0e4e6f90d4 100644 --- a/Android.mk +++ b/Android.mk @@ -450,7 +450,6 @@ LOCAL_SRC_FILES := \ src/armnn/test/optimizations/OptimizeInversePermutesTests.cpp \ src/armnn/test/optimizations/PermuteAndBatchToSpaceAsDepthToSpaceTests.cpp \ src/armnn/test/optimizations/PermuteAsReshapeTests.cpp \ - src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp \ src/armnn/test/optimizations/ReduceMultipleAxesTests.cpp \ src/armnn/test/optimizations/SquashEqualSiblingsTests.cpp \ src/armnn/test/optimizations/TransposeAsReshapeTests.cpp \ diff --git a/CMakeLists.txt b/CMakeLists.txt index 76fb958f07..fc4207b258 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -591,7 +591,6 @@ if(BUILD_UNIT_TESTS) src/armnn/test/optimizations/OptimizeInversePermutesTests.cpp src/armnn/test/optimizations/PermuteAndBatchToSpaceAsDepthToSpaceTests.cpp src/armnn/test/optimizations/PermuteAsReshapeTests.cpp - src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp src/armnn/test/optimizations/ReduceMultipleAxesTests.cpp src/armnn/test/optimizations/SquashEqualSiblingsTests.cpp src/armnn/test/optimizations/TransposeAsReshapeTests.cpp diff --git a/include/armnn/backends/WorkloadData.hpp b/include/armnn/backends/WorkloadData.hpp index 4fbb6d423a..7a0a765f8e 100644 --- a/include/armnn/backends/WorkloadData.hpp +++ b/include/armnn/backends/WorkloadData.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2021-2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -179,15 +179,6 @@ struct FillQueueDescriptor : QueueDescriptorWithParameters // Fully connected layer workload data. struct FullyConnectedQueueDescriptor : QueueDescriptorWithParameters { - FullyConnectedQueueDescriptor() - : m_Weight(nullptr) - , m_Bias(nullptr) - { - } - - const ConstTensorHandle* m_Weight; - const ConstTensorHandle* m_Bias; - void Validate(const WorkloadInfo& workloadInfo) const; }; @@ -213,15 +204,6 @@ struct Pooling3dQueueDescriptor : QueueDescriptorWithParameters { - Convolution2dQueueDescriptor() - : m_Weight(nullptr) - , m_Bias(nullptr) - { - } - - const ConstTensorHandle* m_Weight; - const ConstTensorHandle* m_Bias; - void Validate(const WorkloadInfo& workloadInfo) const; }; @@ -246,15 +228,6 @@ struct Convolution3dQueueDescriptor : QueueDescriptorWithParameters { - DepthwiseConvolution2dQueueDescriptor() - : m_Weight(nullptr) - , m_Bias(nullptr) - { - } - - const ConstTensorHandle* m_Weight; - const ConstTensorHandle* m_Bias; - void Validate(const WorkloadInfo& workloadInfo) const; }; diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp index 40fbde8ac8..7b24fd77b8 100644 --- a/src/armnn/LoadedNetwork.cpp +++ b/src/armnn/LoadedNetwork.cpp @@ -421,7 +421,7 @@ LoadedNetwork::LoadedNetwork(std::unique_ptr net, ConstWorkloads.push_back(m_WorkloadQueue.back().get()); } } - // release the constant data in the layer.. + // release the constant data in the layer. layer->ReleaseConstantData(); break; } diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp index a61624fb0a..158142f48e 100644 --- a/src/armnn/Network.cpp +++ b/src/armnn/Network.cpp @@ -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 // @@ -1714,9 +1714,6 @@ IOptimizedNetworkPtr Optimize(const Graph& inGraph, ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "Optimizer_ConvertConstants"); Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsFloatToHalf())); Optimizer::Pass(optGraph, MakeOptimizations(ConvertConstantsHalfToFloat())); - - // Once the constants are converted we can now safely call RedirectMembersToConstantInputs - Optimizer::Pass(optGraph, MakeOptimizations(RedirectMembersToConstantInputs())); } // This must occur after all topological changes to the graph and any redirection of variables @@ -1860,82 +1857,6 @@ IConnectableLayer* NetworkImpl::AddFullyConnectedLayer(const FullyConnectedDescr return m_Graph->AddLayer(fullyConnectedDescriptor, name); } -IConnectableLayer* NetworkImpl::AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, - const Optional& weights, - const Optional& biases, - const char* name) -{ - ConstantLayer* weightsLayer = nullptr; - ConstantLayer* biasLayer = nullptr; - unsigned int numInputs = fullyConnectedDescriptor.GetNumInputs(); - - // Add a constant layer for weights - if (weights.has_value()) - { - weightsLayer = m_Graph->AddLayer("Weights"); - weightsLayer->m_LayerOutput = std::make_shared(weights.value()); - - TensorInfo weightsInfo = weightsLayer->m_LayerOutput->GetTensorInfo(); - weightsInfo.SetConstant(); - - weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); - } - else if (fullyConnectedDescriptor.m_ConstantWeights) - { - throw InvalidArgumentException("AddFullyConnectedLayer: Constant weights tensor is empty."); - } - - // Add a constant layer for biases - if (biases.has_value() && fullyConnectedDescriptor.m_BiasEnabled) - { - biasLayer = m_Graph->AddLayer("Biases"); - biasLayer->m_LayerOutput = std::make_shared(biases.value()); - - TensorInfo biasInfo = biasLayer->m_LayerOutput->GetTensorInfo(); - biasInfo.SetConstant(); - - biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); - } - - if (numInputs < 2) - { - throw InvalidArgumentException("AddFullyConnectedLayer: Requires at least 2 input tensors: Input, Weights"); - } - - auto layer = m_Graph->AddLayer(fullyConnectedDescriptor, name); - - if (weightsLayer) - { - // Connect weights layer - weightsLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); - } - - if ( fullyConnectedDescriptor.m_BiasEnabled && numInputs == 3 ) - { - if (biasLayer) - { - // Connect bias layer - biasLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(2)); - } - } - else if ( !fullyConnectedDescriptor.m_BiasEnabled && numInputs == 2 ) - { - // Bias is disabled - layer->m_Bias = nullptr; - } - else - { - throw InvalidArgumentException(fmt::format( - "AddFullyConnectedLayer: Value mismatch. When bias is enabled in the " - "descriptor the number of inputs is expected to be 3 otherwise 2. " - "BiasEnabled={}, numInputs={}", - fullyConnectedDescriptor.m_BiasEnabled, - numInputs)); - } - - return layer; -} - IConnectableLayer* NetworkImpl::AddConcatLayer(const ConcatDescriptor& concatDescriptor, const char* name) { @@ -1948,32 +1869,6 @@ IConnectableLayer* NetworkImpl::AddConvolution2dLayer(const Convolution2dDescrip return m_Graph->AddLayer(convolution2dDescriptor, name); } -IConnectableLayer* NetworkImpl::AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional& biases, - const char* name) -{ - auto layer = m_Graph->AddLayer(convolution2dDescriptor, name); - // Add a constant layer for weights - ConstantLayer* weightsLayer = m_Graph->AddLayer("Weights"); - auto weightsTensorHandle = std::make_shared(weights); - weightsLayer->m_LayerOutput = weightsTensorHandle; - layer->m_Weight = weightsTensorHandle; - weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo()); - weightsLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); - // Add a constant layer for biases - if (biases.has_value() && convolution2dDescriptor.m_BiasEnabled) - { - ConstantLayer* biasLayer = m_Graph->AddLayer("Bias"); - auto biasTensorHandle = std::make_shared(biases.value()); - biasLayer->m_LayerOutput = biasTensorHandle; - layer->m_Bias = biasTensorHandle; - biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo()); - biasLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(2)); - } - return layer; -} - IConnectableLayer* NetworkImpl::AddConvertFp16ToFp32Layer(const char* name) { return m_Graph->AddLayer(name); @@ -2003,38 +1898,6 @@ IConnectableLayer* NetworkImpl::AddDepthwiseConvolution2dLayer( return m_Graph->AddLayer(convolution2dDescriptor, name); } -IConnectableLayer* NetworkImpl::AddDepthwiseConvolution2dLayer( - const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional& biases, - const char* name) -{ - auto layer = m_Graph->AddLayer(convolution2dDescriptor, name); - - // Add a constant layer for weights - ConstantLayer* weightsLayer = m_Graph->AddLayer("Weights"); - auto weightsTensorHandle = std::make_shared(weights); - weightsLayer->m_LayerOutput = weightsTensorHandle; - layer->m_Weight = weightsTensorHandle; - - weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo()); - weightsLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); - - // Add a constant layer for biases - if (biases.has_value() && convolution2dDescriptor.m_BiasEnabled) - { - ConstantLayer* biasLayer = m_Graph->AddLayer("Bias"); - auto biasTensorHandle = std::make_shared(biases.value()); - biasLayer->m_LayerOutput = biasTensorHandle; - layer->m_Bias = biasTensorHandle; - - biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo()); - biasLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(2)); - } - - return layer; -} - IConnectableLayer* NetworkImpl::AddDetectionPostProcessLayer(const armnn::DetectionPostProcessDescriptor& descriptor, const ConstTensor& anchors, const char* name) { diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp index 5ca16e2968..a37a4be218 100644 --- a/src/armnn/Network.hpp +++ b/src/armnn/Network.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 @@ -76,23 +76,6 @@ public: IConnectableLayer* AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, const char* name = nullptr); - ARMNN_DEPRECATED_MSG_REMOVAL_DATE("This AddConvolution2dLayer overload is deprecated", "23.02") - IConnectableLayer* AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional& biases, - const char* name = nullptr); - - ARMNN_DEPRECATED_MSG_REMOVAL_DATE("This AddConvolution2dLayer overload is deprecated", "23.02") - IConnectableLayer* AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const char* name = nullptr); - - ARMNN_DEPRECATED_MSG_REMOVAL_DATE("This AddConvolution2dLayer overload is deprecated", "23.02") - IConnectableLayer* AddConvolution2dLayer(const Convolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const ConstTensor& biases, - const char* name = nullptr); - IConnectableLayer* AddConvolution3dLayer(const Convolution3dDescriptor& convolution3dDescriptor, const char* name = nullptr); @@ -101,23 +84,14 @@ public: IConnectableLayer* AddDepthToSpaceLayer(const DepthToSpaceDescriptor& depthToSpaceDescriptor, const char* name = nullptr); - IConnectableLayer* AddDepthwiseConvolution2dLayer( - const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, - const char* name = nullptr); - - ARMNN_DEPRECATED_MSG("This AddDepthwiseConvolution2dLayer overload is deprecated") - IConnectableLayer* AddDepthwiseConvolution2dLayer( - const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional& biases, - const char* name = nullptr); + IConnectableLayer* AddDepthwiseConvolution2dLayer(const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, + const char* name = nullptr); IConnectableLayer* AddDequantizeLayer(const char* name = nullptr); - IConnectableLayer* AddDetectionPostProcessLayer( - const DetectionPostProcessDescriptor& descriptor, - const ConstTensor& anchors, - const char* name = nullptr); + IConnectableLayer* AddDetectionPostProcessLayer(const DetectionPostProcessDescriptor& descriptor, + const ConstTensor& anchors, + const char* name = nullptr); IConnectableLayer* AddDivisionLayer(const char* name = nullptr); @@ -134,11 +108,6 @@ public: IConnectableLayer* AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, const char* name = nullptr); - IConnectableLayer* AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor, - const Optional& weights, - const Optional& biases, - const char* name = nullptr); - IConnectableLayer* AddGatherLayer(const GatherDescriptor& gatherDescriptor, const char* name = nullptr); diff --git a/src/armnn/layers/Convolution2dLayer.cpp b/src/armnn/layers/Convolution2dLayer.cpp index d0233976c4..e06b45acb0 100644 --- a/src/armnn/layers/Convolution2dLayer.cpp +++ b/src/armnn/layers/Convolution2dLayer.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 // @@ -48,18 +48,8 @@ void Convolution2dLayer::SerializeLayerParameters(ParameterStringifyFunction& fn std::unique_ptr Convolution2dLayer::CreateWorkload(const IWorkloadFactory& factory) const { - // on this level constant data should not be released.. ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "Convolution2dLayer_CreateWorkload"); Convolution2dQueueDescriptor descriptor; - if (m_Weight) - { - descriptor.m_Weight = m_Weight.get(); - } - if (m_Param.m_BiasEnabled && m_Bias) - { - descriptor.m_Bias = m_Bias.get(); - } - SetAdditionalInfo(descriptor); return factory.CreateWorkload(LayerType::Convolution2d, descriptor, PrepInfoAndDesc(descriptor)); @@ -68,14 +58,6 @@ std::unique_ptr Convolution2dLayer::CreateWorkload(const IWorkloadFac Convolution2dLayer* Convolution2dLayer::Clone(Graph& graph) const { auto layer = CloneBase(graph, m_Param, GetName()); - - layer->m_Weight = m_Weight ? m_Weight : nullptr; - - if (layer->m_Param.m_BiasEnabled) - { - layer->m_Bias = m_Bias ? m_Bias : nullptr; - } - return std::move(layer); } @@ -140,14 +122,7 @@ void Convolution2dLayer::ValidateTensorShapesFromInputs() Layer::ConstantTensors Convolution2dLayer::GetConstantTensorsByRef() { Layer::ConstantTensors tensors = GetConnectedConstantAsInputTensors(); - - if (!tensors.empty()) - { - return tensors; - } - - // For API stability DO NOT ALTER order and add new members to the end of vector - return {m_Weight, m_Bias}; + return tensors; } void Convolution2dLayer::ExecuteStrategy(IStrategy& strategy) const diff --git a/src/armnn/layers/Convolution2dLayer.hpp b/src/armnn/layers/Convolution2dLayer.hpp index 02ae05f83b..f7e4dec72f 100644 --- a/src/armnn/layers/Convolution2dLayer.hpp +++ b/src/armnn/layers/Convolution2dLayer.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 @@ -16,13 +16,6 @@ class Convolution2dLayer : public LayerWithParameters { public: - /// A unique pointer to store Weight values. - /// @Note: Deprecated. Removal date is 23.02. Weights are stored in ConstantLayers now. - std::shared_ptr m_Weight; - /// A unique pointer to store Bias values. - /// @Note: Deprecated. Removal date is 23.02. Bias are stored in ConstantLayers now. - std::shared_ptr m_Bias; - /// Makes a workload for the Convolution2d type. /// @param [in] graph The graph where this layer can be found. /// @param [in] factory The workload factory which will create the workload. @@ -48,6 +41,10 @@ public: void SerializeLayerParameters(ParameterStringifyFunction& fn) const override; + /// This layer does not have any data stored, weights and bias are now stored in constant layers. + /// We do not want to release the data in the constant layer, that is why we override with an empty function. + void ReleaseConstantData() override {} + protected: /// Constructor to create a Convolution2dLayer. /// @param [in] param Convolution2dDescriptor to configure the convolution2d operation. @@ -57,8 +54,8 @@ protected: /// Default destructor ~Convolution2dLayer() = default; - /// @Note Deprecated. GetConstantTensorsByRef is deprecated. m_Weights and m_Bias - /// should be connected to layer as Constant Layers instead." + /// Retrieve the handles to the constant values connected to the layer. + /// @return A vector of the constant tensors connected to the layer. ConstantTensors GetConstantTensorsByRef() override; }; diff --git a/src/armnn/layers/DepthwiseConvolution2dLayer.cpp b/src/armnn/layers/DepthwiseConvolution2dLayer.cpp index dcd800e367..4c97437a1c 100644 --- a/src/armnn/layers/DepthwiseConvolution2dLayer.cpp +++ b/src/armnn/layers/DepthwiseConvolution2dLayer.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 // @@ -50,16 +50,6 @@ void DepthwiseConvolution2dLayer::SerializeLayerParameters(ParameterStringifyFun std::unique_ptr DepthwiseConvolution2dLayer::CreateWorkload(const IWorkloadFactory& factory) const { DepthwiseConvolution2dQueueDescriptor descriptor; - - if (m_Weight) - { - descriptor.m_Weight = m_Weight.get(); - } - if (m_Param.m_BiasEnabled && m_Bias) - { - descriptor.m_Bias = m_Bias.get(); - } - SetAdditionalInfo(descriptor); return factory.CreateWorkload(LayerType::DepthwiseConvolution2d, descriptor, PrepInfoAndDesc(descriptor)); @@ -68,13 +58,6 @@ std::unique_ptr DepthwiseConvolution2dLayer::CreateWorkload(const IWo DepthwiseConvolution2dLayer* DepthwiseConvolution2dLayer::Clone(Graph& graph) const { auto layer = CloneBase(graph, m_Param, GetName()); - layer->m_Weight = m_Weight ? m_Weight : nullptr; - - if (layer->m_Param.m_BiasEnabled) - { - layer->m_Bias = m_Bias ? m_Bias : nullptr; - } - return std::move(layer); } @@ -143,14 +126,7 @@ void DepthwiseConvolution2dLayer::ValidateTensorShapesFromInputs() Layer::ConstantTensors DepthwiseConvolution2dLayer::GetConstantTensorsByRef() { Layer::ConstantTensors tensors = GetConnectedConstantAsInputTensors(); - - if (!tensors.empty()) - { - return tensors; - } - - // For API stability DO NOT ALTER order and add new members to the end of vector - return {m_Weight, m_Bias}; + return tensors; } void DepthwiseConvolution2dLayer::ExecuteStrategy(IStrategy& strategy) const diff --git a/src/armnn/layers/DepthwiseConvolution2dLayer.hpp b/src/armnn/layers/DepthwiseConvolution2dLayer.hpp index baae7f122a..ef7410f1d3 100644 --- a/src/armnn/layers/DepthwiseConvolution2dLayer.hpp +++ b/src/armnn/layers/DepthwiseConvolution2dLayer.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 @@ -15,12 +15,6 @@ class ScopedTensorHandle; class DepthwiseConvolution2dLayer : public LayerWithParameters { public: - /// A unique pointer to store Weight values. - /// @Note Deprecated. Removal date is 23.02. Weights are stored in ConstantLayers now. - std::shared_ptr m_Weight; - /// A unique pointer to store Bias values. - /// @Note Deprecated. Removal date is 23.02. Bias are stored in ConstantLayers now. - std::shared_ptr m_Bias; /// Makes a workload for the DepthwiseConvolution2d type. /// @param [in] graph The graph where this layer can be found. @@ -47,6 +41,10 @@ public: void SerializeLayerParameters(ParameterStringifyFunction& fn) const override; + /// This layer does not have any data stored, weights and bias are now stored in constant layers. + /// We do not want to release the data in the constant layer, that is why we override with an empty function. + void ReleaseConstantData() override {} + protected: /// Constructor to create a DepthwiseConvolution2dLayer. /// @param [in] param DepthwiseConvolution2dDescriptor to configure the depthwise convolution2d. @@ -56,10 +54,8 @@ protected: /// Default destructor ~DepthwiseConvolution2dLayer() = default; - /// Retrieve the handles to the constant values stored by the layer. - /// @return A vector of the constant tensors stored by this layer. - /// @Note Deprecated. GetConstantTensorsByRef is deprecated. m_Weights and m_Bias - /// should be connected to layer as Constant Layers instead." + /// Retrieve the handles to the constant values connected to the layer. + /// @return A vector of the constant tensors connected to the layer. ConstantTensors GetConstantTensorsByRef() override; }; diff --git a/src/armnn/layers/FullyConnectedLayer.cpp b/src/armnn/layers/FullyConnectedLayer.cpp index c20bc8d167..05c53961e3 100644 --- a/src/armnn/layers/FullyConnectedLayer.cpp +++ b/src/armnn/layers/FullyConnectedLayer.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 // #include "FullyConnectedLayer.hpp" @@ -22,27 +22,13 @@ FullyConnectedLayer::FullyConnectedLayer(const FullyConnectedDescriptor& param, std::unique_ptr FullyConnectedLayer::CreateWorkload(const IWorkloadFactory& factory) const { FullyConnectedQueueDescriptor descriptor; - if (m_Weight) - { - descriptor.m_Weight = m_Weight.get(); - } - if (m_Param.m_BiasEnabled && m_Bias) - { - descriptor.m_Bias = m_Bias.get(); - } SetAdditionalInfo(descriptor); - return factory.CreateWorkload(LayerType::FullyConnected, descriptor, PrepInfoAndDesc(descriptor)); } FullyConnectedLayer* FullyConnectedLayer::Clone(Graph& graph) const { auto layer = CloneBase(graph, m_Param, GetName()); - layer->m_Weight = m_Weight ? m_Weight : nullptr; - if (layer->m_Param.m_BiasEnabled) - { - layer->m_Bias = m_Bias ? m_Bias : nullptr; - } return std::move(layer); } @@ -78,14 +64,7 @@ void FullyConnectedLayer::ValidateTensorShapesFromInputs() Layer::ConstantTensors FullyConnectedLayer::GetConstantTensorsByRef() { Layer::ConstantTensors tensors = GetConnectedConstantAsInputTensors(); - - if (!tensors.empty()) - { - return tensors; - } - - // For API stability DO NOT ALTER order and add new members to the end of vector - return {m_Weight, m_Bias}; + return tensors; } void FullyConnectedLayer::ExecuteStrategy(IStrategy& strategy) const diff --git a/src/armnn/layers/FullyConnectedLayer.hpp b/src/armnn/layers/FullyConnectedLayer.hpp index 07f4a936f9..f3ca696b62 100644 --- a/src/armnn/layers/FullyConnectedLayer.hpp +++ b/src/armnn/layers/FullyConnectedLayer.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 @@ -15,12 +15,6 @@ class ScopedTensorHandle; class FullyConnectedLayer : public LayerWithParameters { public: - /// A unique pointer to store Weight values. - /// @Note: Deprecated. Removal date is 23.02. Weights are stored in ConstantLayers now. - std::shared_ptr m_Weight; - /// A unique pointer to store Bias values. - /// @Note: Deprecated. Removal date is 23.02. Bias are stored in ConstantLayers now. - std::shared_ptr m_Bias; /// Makes a workload for the FullyConnected type. /// @param [in] graph The graph where this layer can be found. @@ -45,6 +39,10 @@ public: void ExecuteStrategy(IStrategy& strategy) const override; + /// This layer does not have any data stored, weights and bias are now stored in constant layers. + /// We do not want to release the data in the constant layer, that is why we override with an empty function. + void ReleaseConstantData() override {} + protected: /// Constructor to create a FullyConnectedLayer. /// @param [in] param FullyConnectedDescriptor to configure the fully connected operation. diff --git a/src/armnn/optimizations/All.hpp b/src/armnn/optimizations/All.hpp index a11dec9446..0e67516193 100644 --- a/src/armnn/optimizations/All.hpp +++ b/src/armnn/optimizations/All.hpp @@ -20,6 +20,5 @@ #include "PermuteAsReshape.hpp" #include "PermuteAndBatchToSpaceAsDepthToSpace.hpp" #include "PermuteDepthwiseConv2dWeights.hpp" -#include "RedirectMembersToConstantInputs.hpp" #include "SquashEqualSiblings.hpp" #include "TransposeAsReshape.hpp" \ No newline at end of file diff --git a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp b/src/armnn/optimizations/FoldPadIntoLayer2d.hpp index 7a50c4ac06..874749fda9 100644 --- a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp +++ b/src/armnn/optimizations/FoldPadIntoLayer2d.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 // @@ -196,21 +196,14 @@ public: if (newConv2dLayer != nullptr) { const auto conv2dLayer = PolymorphicDowncast(&connection.GetOwningLayer()); - // Copy weights and bias to the new convolution layer ARMNN_ASSERT_MSG(newConv2dLayer->GetInputSlot(1).GetConnection() != nullptr, "FoldPadIntoConvolution2d: New convolution layer is missing connection to weights layer"); - // Deprecated. Removal date is 23.02. - newConv2dLayer->m_Weight = std::move(conv2dLayer->m_Weight); - if (conv2dLayer->GetParameters().m_BiasEnabled) { ARMNN_ASSERT_MSG(newConv2dLayer->GetInputSlot(2).GetConnection() != nullptr, "FoldPadIntoConvolution2d: New convolution layer is missing " "connection to bias layer."); - - // Deprecated. Removal date is 23.02. - newConv2dLayer->m_Bias = std::move(conv2dLayer->m_Bias); } } } @@ -230,24 +223,18 @@ public: if (newConv2dLayer != nullptr) { const auto conv2dLayer = PolymorphicDowncast(&connection.GetOwningLayer()); - // Copy weights and bias to the new convolution layer ARMNN_ASSERT_MSG(newConv2dLayer->GetInputSlot(1).GetConnection() != nullptr, - "FoldPadIntoDepthwiseConvolution2d: New convolution layer is missing connection to weights layer"); - - // Deprecated. Removal date is 23.02. - newConv2dLayer->m_Weight = std::move(conv2dLayer->m_Weight); + "FoldPadIntoDepthwiseConvolution2d: New convolution layer is missing " + "connection to weights layer"); if (conv2dLayer->GetParameters().m_BiasEnabled) { ARMNN_ASSERT_MSG(newConv2dLayer->GetInputSlot(2).GetConnection() != nullptr, "FoldPadIntoConvolution2d: New convolution layer is missing " "connection to bias layer."); - // Deprecated. Removal date is 23.02. - newConv2dLayer->m_Bias = std::move(conv2dLayer->m_Bias); } } } - protected: FoldPadIntoDepthwiseConvolution2dImpl() = default; ~FoldPadIntoDepthwiseConvolution2dImpl() = default; diff --git a/src/armnn/optimizations/FuseBatchNorm.hpp b/src/armnn/optimizations/FuseBatchNorm.hpp index bca0c7d00a..88ac97cd0c 100644 --- a/src/armnn/optimizations/FuseBatchNorm.hpp +++ b/src/armnn/optimizations/FuseBatchNorm.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020,2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -167,8 +167,6 @@ public: auto& newConv2dLayer = *graph.InsertNewLayer(base.GetInputSlot(0), convDescriptor, name.c_str()); - newConv2dLayer.m_Weight = std::make_unique(fusedWeightsTensor); - newConv2dLayer.m_Bias = std::make_unique(ConstTensor(fusedBiasTensor)); // Connect weights and bias from old to new Conv2d layer // This optimization will always have 3 input slots on the Conv2d base layer @@ -177,7 +175,7 @@ public: // Remove old connection and connect to new layer2d weightLayer->GetOutputSlot(0).Disconnect(base.GetInputSlot(1)); weightLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(1)); - weightLayer->m_LayerOutput = newConv2dLayer.m_Weight; + weightLayer->m_LayerOutput = std::make_unique(fusedWeightsTensor); // Move bias const layers as normal if it was enabled before the optimisation ConstantLayer* biasLayer; @@ -198,7 +196,7 @@ public: biasLayer->GetOutputSlot(0).SetTensorInfo(fusedBiasTensor.GetInfo()); biasLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(2)); } - biasLayer->m_LayerOutput = newConv2dLayer.m_Bias; + biasLayer->m_LayerOutput = std::make_unique(ConstTensor(fusedBiasTensor)); } diff --git a/src/armnn/optimizations/RedirectMembersToConstantInputs.hpp b/src/armnn/optimizations/RedirectMembersToConstantInputs.hpp deleted file mode 100644 index a2bad710e6..0000000000 --- a/src/armnn/optimizations/RedirectMembersToConstantInputs.hpp +++ /dev/null @@ -1,90 +0,0 @@ -// -// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include "Optimization.hpp" - -#include -#include - -namespace armnn -{ -namespace optimizations -{ - -class RedirectMembersToConstantInputsImpl -{ -public: - /// Search for layers with ConstantLayers as inputs. If the inputs are constant redirect the layers member - /// variable for ConstTensors (e.g. m_weights) to the data stored in the ConstantLayer it is connected to. - void Run(Graph& graph, Layer& layer) const - { - IgnoreUnused(graph); - - switch (layer.GetType()) - { - case LayerType::BatchNormalization: - break; - case LayerType::Convolution2d: - RedirectWeightsAndBiases(&layer); - break; - case LayerType::DepthwiseConvolution2d: - RedirectWeightsAndBiases(&layer); - break; - case LayerType::DetectionPostProcess: - break; - case LayerType::FullyConnected: - RedirectWeightsAndBiases(&layer); - break; - case LayerType::Lstm: - break; - case LayerType::TransposeConvolution2d: - break; - default: - break; - } - } - -protected: - RedirectMembersToConstantInputsImpl() = default; - ~RedirectMembersToConstantInputsImpl() = default; - -private: - template - static LayerT* RedirectWeightsAndBiases(Layer* layer) - { - LayerT* layerPtr = PolymorphicDowncast(layer); - - // Loop through input slots to check for constant weights and biases layers. - // Weights index = 1, Biases index = 2. - for (unsigned int inputSlotIndex = 1; inputSlotIndex != layerPtr->GetNumInputSlots(); ++inputSlotIndex) - { - OutputSlot* outputSlot = layerPtr->GetInputSlot(inputSlotIndex).GetConnectedOutputSlot(); - // Debug layers should not be inserted in optimize process yet - ARMNN_ASSERT(outputSlot->GetOwningLayer().GetType() != LayerType::Debug); - if (outputSlot->GetOwningLayer().GetType() == LayerType::Constant) - { - // Get constant layer and redirect base layer member variables. - ConstantLayer& constantLayer = dynamic_cast(outputSlot->GetOwningLayer()); - if (inputSlotIndex == 1) - { - layerPtr->m_Weight = constantLayer.m_LayerOutput; - } - else if (inputSlotIndex == 2) - { - layerPtr->m_Bias = constantLayer.m_LayerOutput; - } - } - } - - return layerPtr; - } -}; - -using RedirectMembersToConstantInputs = OptimizeForType; - -} // namespace optimizations -} // namespace armnn diff --git a/src/armnn/test/OptimizerTests.cpp b/src/armnn/test/OptimizerTests.cpp index b78863dddc..f83900404b 100644 --- a/src/armnn/test/OptimizerTests.cpp +++ b/src/armnn/test/OptimizerTests.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 // @@ -441,16 +441,15 @@ void CreateConvolution2dGraph(Graph &graph, const unsigned int* inputShape, Layer* input = graph.AddLayer(0, "input"); input->GetOutputSlot().SetTensorInfo(inputInfo); - ConstantLayer* weightsLayer = nullptr; - weightsLayer = graph.AddLayer("Weights"); + ConstantLayer* weightsLayer = graph.AddLayer("Weights"); weightsLayer->m_LayerOutput = std::make_shared(weights); weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsLayer->m_LayerOutput->GetTensorInfo()); Convolution2dLayer* layer = graph.AddLayer(desc, "conv2d"); - layer->m_Weight = std::make_unique(weights); layer->GetOutputSlot().SetTensorInfo(outputInfo); Layer* output = graph.AddLayer(0, "output"); + input->GetOutputSlot().Connect(layer->GetInputSlot(0)); layer->GetOutputSlot().Connect(output->GetInputSlot(0)); weightsLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); @@ -908,11 +907,10 @@ TEST_CASE("OptimizeForExclusiveConnectionsFuseTest") { std::vector biasVector = { 11 }; ConstTensor bias(TensorInfo(1, outputChannelSize, DataType::Float32, 0.0f, 0, true), biasVector); - biasLayer =graph.AddLayer("Bias"); + biasLayer = graph.AddLayer("Bias"); biasLayer->m_LayerOutput = std::make_shared(bias); biasLayer->GetOutputSlot(0).SetTensorInfo(biasLayer->m_LayerOutput->GetTensorInfo()); biasLayer->GetOutputSlot(0).Connect(conv->GetInputSlot(2)); - conv->m_Bias = biasLayer->m_LayerOutput; } // Connect layers @@ -921,9 +919,6 @@ TEST_CASE("OptimizeForExclusiveConnectionsFuseTest") conv->GetOutputSlot(0).Connect(batchNorm->GetInputSlot(0)); batchNorm->GetOutputSlot(0).Connect(output->GetInputSlot(0)); - // Temporary workaround to ensure the descriptor weights are populated - conv->m_Weight = weightsLayer->m_LayerOutput; - if (convolution2dDescriptor.m_BiasEnabled) { CHECK(6 == graph.GetNumLayers()); @@ -983,22 +978,22 @@ TEST_CASE("OptimizeForExclusiveConnectionsWithoutFuseTest") batchNorm->GetOutputSlot(0).Connect(output->GetInputSlot(0)); conv->GetOutputSlot(0).Connect(output2->GetInputSlot(0)); - CHECK(5 == graph.GetNumLayers()); + CHECK((5 == graph.GetNumLayers())); CHECK(CheckSequence(graph.cbegin(), graph.cend(), - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType)); + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType)); // Optimize graph armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(FuseBatchNormIntoConvolution2DFloat32())); CHECK(5 == graph.GetNumLayers()); CHECK(CheckSequence(graph.cbegin(), graph.cend(), - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType)); + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType)); } } // Optimizer TestSuite diff --git a/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp b/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp index 34e5f6d3b6..118907e703 100644 --- a/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp +++ b/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp @@ -1,12 +1,12 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include -#include #include +#include #include @@ -25,33 +25,38 @@ TEST_CASE("ConvertConstantsFloatToHalfTest") // Create const tensor from fp32 data unsigned int dims[] = { 4, 1, 1, 1 }; std::vector floatWeights{ 1.0f, 2.0f, 3.0f, 4.0f }; - armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true), floatWeights); + armnn::TensorInfo weightsInfo = armnn::TensorInfo(4, dims, armnn::DataType::Float32, 0.0f, 0, true); + armnn::ConstTensor weights(weightsInfo, floatWeights); // Create simple test network auto input = graph.AddLayer(0, "input"); input->GetOutputSlot().SetTensorInfo(info); auto fc = graph.AddLayer(armnn::FullyConnectedDescriptor(), "fc"); - fc->m_Weight = std::make_unique(weights); fc->GetOutputSlot().SetTensorInfo(info); + auto weightsLayer = graph.AddLayer("weights"); + weightsLayer->m_LayerOutput = std::make_unique(weights); + weightsLayer->GetOutputSlot().SetTensorInfo(weightsInfo); + auto output = graph.AddLayer(1, "output"); // Connect up the layers input->GetOutputSlot().Connect(fc->GetInputSlot(0)); + weightsLayer->GetOutputSlot().Connect(fc->GetInputSlot(1)); fc->GetOutputSlot().Connect(output->GetInputSlot(0)); // Check tensor data type before conversion - CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float32); + CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32); // Run the optimizer armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsFloatToHalf())); // Check tensor data type after conversion - CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float16); + CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16); // Check whether data matches expected fp16 data - const Half* data = fc->m_Weight->GetConstTensor(); + const Half* data = weightsLayer->m_LayerOutput->GetConstTensor(); CHECK(data[0] == Half(1.0f)); CHECK(data[1] == Half(2.0f)); CHECK(data[2] == Half(3.0f)); @@ -100,12 +105,14 @@ TEST_CASE("ConvertConstantsFloatToHalfTest_constant") fcLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); // Check tensor data type before conversion + CHECK(5 == graph.GetNumLayers()); CHECK(weights->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32); // Run the optimizer armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsFloatToHalf())); // Check tensor data type after conversion + CHECK(5 == graph.GetNumLayers()); CHECK(weights->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16); // Check whether weights data matches expected fp16 data diff --git a/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp b/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp index 4c453cc799..778d7b0814 100644 --- a/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp +++ b/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -25,33 +25,38 @@ TEST_CASE("ConvertConstantsHalfToFloatTest") std::vector halfWeights(4); armnnUtils::FloatingPointConverter::ConvertFloat32To16(convWeightsData.data(), convWeightsData.size(), halfWeights.data()); - armnn::ConstTensor weights(armnn::TensorInfo(4, dims, armnn::DataType::Float16, 0.0f, 0, true), halfWeights); + armnn::TensorInfo weightInfo = armnn::TensorInfo(4, dims, armnn::DataType::Float16, 0.0f, 0, true); + armnn::ConstTensor weights(weightInfo, halfWeights); //Create the simple test network auto input = graph.AddLayer(0, "input"); input->GetOutputSlot().SetTensorInfo(info); auto fc = graph.AddLayer(armnn::FullyConnectedDescriptor(), "fc"); - fc->m_Weight = std::make_unique(weights); fc->GetOutputSlot().SetTensorInfo(info); + auto weightsLayer = graph.AddLayer("weights"); + weightsLayer->m_LayerOutput = std::make_unique(weights); + weightsLayer->GetOutputSlot(0).SetTensorInfo(weightInfo); + auto output = graph.AddLayer(1, "output"); //Connect up the layers input->GetOutputSlot().Connect(fc->GetInputSlot(0)); + weightsLayer->GetOutputSlot().Connect(fc->GetInputSlot(1)); fc->GetOutputSlot().Connect(output->GetInputSlot(0)); //Test the tensor info is correct. - CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float16); + CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float16); // Run the optimizer armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(ConvertConstantsHalfToFloat())); //Test the tensor info is correct. - CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float32); + CHECK(weightsLayer->m_LayerOutput->GetTensorInfo().GetDataType() == armnn::DataType::Float32); // Now test the data matches float32 data - const float* data = fc->m_Weight->GetConstTensor(); + const float* data = weightsLayer->m_LayerOutput->GetConstTensor(); CHECK(1.0f == data[0]); CHECK(2.0f == data[1]); CHECK(3.0f == data[2]); diff --git a/src/armnn/test/optimizations/Fp32NetworkToFp16ConverterTests.cpp b/src/armnn/test/optimizations/Fp32NetworkToFp16ConverterTests.cpp index bc8839948b..0a4a4fafde 100644 --- a/src/armnn/test/optimizations/Fp32NetworkToFp16ConverterTests.cpp +++ b/src/armnn/test/optimizations/Fp32NetworkToFp16ConverterTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -33,14 +33,21 @@ TEST_CASE("Fp32NetworkToFp16OptimizationTest") floor->GetOutputSlot().Connect(output->GetInputSlot(0)); CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType, - &IsLayerOfType, &IsLayerOfType)); + &IsLayerOfType, + &IsLayerOfType)); // Run the optimizer armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(Fp32NetworkToFp16Converter())); CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType, - &IsLayerOfType, &IsLayerOfType, - &IsLayerOfType, &IsLayerOfType)); + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType, + &IsLayerOfType)); + + CHECK(floor->GetDataType() == armnn::DataType::Float16); + CHECK(floor->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo().GetDataType() == armnn::DataType::Float16); + CHECK(floor->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16); } } \ No newline at end of file diff --git a/src/armnn/test/optimizations/FuseBatchNormTests.cpp b/src/armnn/test/optimizations/FuseBatchNormTests.cpp index 54cbbce89f..5cbd17fb6a 100644 --- a/src/armnn/test/optimizations/FuseBatchNormTests.cpp +++ b/src/armnn/test/optimizations/FuseBatchNormTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -27,13 +27,8 @@ public: static IConnectableLayer *AddConvolution(INetwork *network, const Convolution2dDescriptor &descriptor, - const ConstTensor &weights, - const Optional &biases, const char *name) { - IgnoreUnused(weights); - IgnoreUnused(biases); - return network->AddConvolution2dLayer(descriptor, name); } @@ -65,12 +60,8 @@ public: static IConnectableLayer* AddConvolution(INetwork* network, const DepthwiseConvolution2dDescriptor& descriptor, - const ConstTensor& weights, - const Optional& biases, const char* name) { - IgnoreUnused(weights); - IgnoreUnused(biases); return network->AddDepthwiseConvolution2dLayer(descriptor, name); } @@ -171,8 +162,6 @@ INetworkPtr CreateNetwork(bool depthwise, bool preventFusing) IConnectableLayer* convLayer = Conv2dTest::AddConvolution(network.get(), convolution2dDescriptor, - weights, - Optional(), "convolution"); IConnectableLayer* batchNormLayer = network->AddBatchNormalizationLayer(batchNormDescriptor, @@ -243,13 +232,21 @@ void FuseBatchNormIntoConvTest(bool depthwise, float tolerance, armnn::Compute b return IsLayerOfType(layer) && (layer->GetNameStr() == "fused-batchNorm-into-convolution"); }; - + auto checkConstant = [ ](const armnn::Layer* const layer) -> bool + { + const ConstantLayer* constLayer = PolymorphicDowncast(layer); + auto tensor = ConstTensor(constLayer->m_LayerOutput->GetTensorInfo(), + constLayer->m_LayerOutput->Map(true)); + const auto* buffer = static_cast(tensor.GetMemoryArea()); + std::vector vector(buffer, buffer + tensor.GetNumElements()); + return IsLayerOfType(layer); + }; CHECK(5 == graphFused.GetNumLayers()); CHECK(CheckSequence(graphFused.cbegin(), graphFused.cend(), &IsLayerOfType, - &IsLayerOfType, - &IsLayerOfType, + checkConstant, + checkConstant, checkFusedConv2d, &IsLayerOfType)); diff --git a/src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp b/src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp deleted file mode 100644 index b3f9ed8780..0000000000 --- a/src/armnn/test/optimizations/RedirectMembersToConstantInputsTests.cpp +++ /dev/null @@ -1,85 +0,0 @@ -// -// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include - -#include - -#include - -TEST_SUITE("Optimizer") -{ -using namespace armnn::optimizations; - -TEST_CASE("RedirectMembersToConstantInputsFullyConnectedTest") -{ - armnn::Graph graph; - - const armnn::TensorInfo inputInfo ({ 1, 2, 2, 3 }, armnn::DataType::Float32); - const armnn::TensorInfo outputInfo ({ 1, 2, 2, 3 }, armnn::DataType::Float32); - const armnn::TensorInfo weightsInfo({ 4 }, armnn::DataType::Float32, 0.0f, 0, true); - const armnn::TensorInfo biasesInfo ({ 2 }, armnn::DataType::Float32, 0.0f, 0, true); - - // Check if isConstant is enabled for weights and biases tensor info. - CHECK(weightsInfo.IsConstant()); - CHECK(biasesInfo.IsConstant()); - - armnn::FullyConnectedDescriptor desc; - desc.m_BiasEnabled = true; - desc.m_ConstantWeights = false; - - // Create the simple test network with Weights and Biases as inputs to a FullyConnected layer. - auto input = graph.AddLayer(0, "Input"); - auto weights = graph.AddLayer("Weights"); - auto biases = graph.AddLayer("Biases"); - auto fcLayer = graph.AddLayer(desc, "FullyConnected"); - auto output = graph.AddLayer(1, "Output"); - - float expectedWeightsData[] = { 1.0f, 1.0f, 1.0f, 1.0f }; - float expectedBiasesData[] = { 2.0f, 2.0f }; - - // Set the m_LayerOutput for the optimizer to point to. - armnn::ConstTensor weightsTensor(weightsInfo, &expectedWeightsData); - armnn::ConstTensor biasesTensor(biasesInfo, &expectedBiasesData); - weights->m_LayerOutput = std::make_unique(weightsTensor); - biases->m_LayerOutput = std::make_unique(biasesTensor); - - input->GetOutputSlot().SetTensorInfo(inputInfo); - weights->GetOutputSlot().SetTensorInfo(weightsInfo); - biases->GetOutputSlot().SetTensorInfo(biasesInfo); - fcLayer->GetOutputSlot().SetTensorInfo(outputInfo); - - // Connect up the layers - input->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(0)); - weights->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(1)); - biases->GetOutputSlot(0).Connect(fcLayer->GetInputSlot(2)); - fcLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); - - // Member variables should be null before optimization. - CHECK(fcLayer->m_Weight == nullptr); - CHECK(fcLayer->m_Bias == nullptr); - - // Run the optimizer - armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(RedirectMembersToConstantInputs())); - - // Check if member variables are not null and shape is set correctly. - CHECK(fcLayer->m_Weight != nullptr); - CHECK(fcLayer->m_Bias != nullptr); - CHECK(fcLayer->m_Weight->GetTensorInfo().GetShape() == weightsInfo.GetShape()); - CHECK(fcLayer->m_Bias->GetTensorInfo().GetShape() == biasesInfo.GetShape()); - - // Check whether data matches expected float data - const float* weightsData = fcLayer->m_Weight->GetConstTensor(); - CHECK(weightsData[0] == expectedWeightsData[0]); - CHECK(weightsData[1] == expectedWeightsData[1]); - CHECK(weightsData[2] == expectedWeightsData[2]); - CHECK(weightsData[3] == expectedWeightsData[3]); - - const float* biasesData = fcLayer->m_Bias->GetConstTensor(); - CHECK(biasesData[0] == expectedBiasesData[0]); - CHECK(biasesData[1] == expectedBiasesData[1]); -} - -} \ No newline at end of file diff --git a/src/armnnTestUtils/CommonTestUtils.hpp b/src/armnnTestUtils/CommonTestUtils.hpp index 5b4b356247..4262ed5180 100644 --- a/src/armnnTestUtils/CommonTestUtils.hpp +++ b/src/armnnTestUtils/CommonTestUtils.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 // @@ -70,15 +70,6 @@ inline bool Compare(T a, T b, float tolerance = 0.000001f) return std::fabs(static_cast(a) - static_cast(b)) <= tolerance; } -template -void SetWeightAndBias(ConvolutionLayer* layer, const armnn::TensorInfo& weightInfo, const armnn::TensorInfo& biasInfo) -{ - layer->m_Weight = std::make_unique(weightInfo); - layer->m_Bias = std::make_unique(biasInfo); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); -} armnn::SubgraphView::InputSlots CreateInputsFrom(armnn::Layer* layer, std::vector ignoreSlots = {}); diff --git a/src/armnnTestUtils/CreateWorkload.hpp b/src/armnnTestUtils/CreateWorkload.hpp index 905b8fa50b..0846d21388 100644 --- a/src/armnnTestUtils/CreateWorkload.hpp +++ b/src/armnnTestUtils/CreateWorkload.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 @@ -495,10 +495,6 @@ std::unique_ptr CreateConvolution2dWorkloadTest(armnn::IW TensorShape inputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3}; TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2}; - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo(weightShape, DataType)); - layer->m_Weight->Allocate(); - armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale); weightsTensorInfo.SetConstant(); @@ -562,12 +558,6 @@ std::unique_ptr CreateConvolution2dFusedActivationWithBlo TensorShape weightShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 5, 3} : TensorShape{2, 5, 3, 3}; TensorShape inputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3}; TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2}; - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo(weightShape, DataType)); - layer->m_Bias = std::make_unique(TensorInfo({2}, GetBiasDataType(DataType))); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale); weightsTensorInfo.SetConstant(); @@ -662,12 +652,6 @@ std::unique_ptr CreateConvolution2dWorkloadFastMathTest(a TensorShape biasShape = TensorShape{ 32 }; TensorShape inputShape = TensorShape{ 1, 32, 149, 149 }; TensorShape outputShape = TensorShape{ 1, 32, 147, 147 }; - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo(weightShape, DataType)); - layer->m_Bias = std::make_unique(TensorInfo(biasShape, GetBiasDataType(DataType))); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale); weightsTensorInfo.SetConstant(); @@ -1132,12 +1116,6 @@ std::unique_ptr CreateDirectConvolution2dWorkloadTest(arm armnn::TensorInfo biasTensorInfo(biasShape, GetBiasDataType(DataType), inputsQScale); biasTensorInfo.SetConstant(); - layer->m_Weight = std::make_unique(weightsTensorInfo); - layer->m_Bias = std::make_unique(biasTensorInfo); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); - // Creates extra layers. Layer* const input = graph.AddLayer(0, "input"); auto const weights = graph.AddLayer("weights"); @@ -1170,8 +1148,6 @@ std::unique_ptr CreateDirectConvolution2dWorkloadTest(arm CHECK(queueDescriptor.m_Inputs.size() == 3); CHECK(queueDescriptor.m_Outputs.size() == 1); - CHECK((queueDescriptor.m_Weight->GetTensorInfo() == weightsTensorInfo)); - CHECK((queueDescriptor.m_Bias->GetTensorInfo() == biasTensorInfo)); // Returns so we can do extra, backend-specific tests. return workload; @@ -1203,9 +1179,6 @@ std::unique_ptr CreateDepthwiseConvolutio DepthwiseConvolution2dLayer* const layer = graph.AddLayer(layerDesc, "layer"); - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo(weightShape, DataType)); // [ 1, H, W, I*M ] - layer->m_Weight->Allocate(); // Creates extra layers. Layer* const input = graph.AddLayer(0, "input"); @@ -1252,10 +1225,6 @@ std::unique_ptr CreateFullyConnectedWorkloadTest(armnn:: float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0; float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0; - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo({7, 20}, DataType, inputsQScale, 0)); - layer->m_Weight->Allocate(); - armnn::TensorInfo weightsTensorInfo({7, 20}, DataType, inputsQScale); weightsTensorInfo.SetConstant(); @@ -1301,12 +1270,6 @@ std::unique_ptr CreateFullyConnectedWithBlobWorkloadTest float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0; float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0; - // As optimization isn't run member variables need to be updated. - layer->m_Weight = std::make_unique(TensorInfo({7, 20}, DataType, inputsQScale, 0)); - layer->m_Bias = std::make_unique(TensorInfo({7}, GetBiasDataType(DataType), inputsQScale)); - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); - armnn::TensorInfo weightsTensorInfo({7, 20}, DataType, inputsQScale); armnn::TensorInfo biasesTensorInfo({7}, GetBiasDataType(DataType), inputsQScale); weightsTensorInfo.SetConstant(); 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(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(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(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(desc, ""); - m_Layer->m_Weight = std::make_unique( - armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); - m_Layer->m_Bias = std::make_unique( - armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); } ~DummyConvolutionLayer() @@ -268,10 +264,28 @@ struct DummyLayer { }; -template<> +// Note: When m_Weight and m_Bias are removed from TransposeConvolution, Transpose can use DummyConvolutionLayer +template <> struct DummyLayer - : public DummyConvolutionLayer { + DummyLayer() + { + typename armnn::TransposeConvolution2dLayer::DescriptorType desc; + desc.m_StrideX = 1; + desc.m_StrideY = 1; + m_Layer = dummyGraph.AddLayer(desc, ""); + m_Layer->m_Weight = std::make_unique( + armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32)); + m_Layer->m_Bias = std::make_unique( + 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::DescriptorType desc; m_Layer = dummyGraph.AddLayer(desc, ""); - m_Layer->m_Weight = std::make_unique( - 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(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 -#include - 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(layerDesc, "layer"); - - layer->m_Weight = std::make_unique(TensorInfo({ 2, 3, 5, 3 }, - armnn::DataType::Float32)); - layer->m_Bias = std::make_unique - (TensorInfo({ 2 }, GetBiasDataType(armnn::DataType::Float32))); - - layer->m_Weight->Allocate(); - layer->m_Bias->Allocate(); - - ConstantLayer* weightsLayer = graph.AddLayer("Weights"); - ConstantLayer* biasLayer = graph.AddLayer("Bias"); + auto* const convolutionLayer = graph.AddLayer(layerDesc, "convolution"); + auto* const weightsLayer = graph.AddLayer("weights"); + auto* const biasLayer = graph.AddLayer("bias"); - weightsLayer->m_LayerOutput = std::make_shared(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( - 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(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared(biasInfo); weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); @@ -116,21 +100,21 @@ TEST_CASE("ReleaseConvolution2dLayerConstantDataTest") Layer* const output = graph.AddLayer(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(layerDesc, "layer"); + auto* const convolutionLayer = graph.AddLayer(layerDesc, "convolution"); + auto* const weightsLayer = graph.AddLayer("weights"); + auto* const biasLayer = graph.AddLayer("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(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared(biasInfo); - layer->m_Weight = std::make_unique( - TensorInfo({3, 3, 5, 3}, DataType::Float32)); - layer->m_Bias = std::make_unique( - 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(0, "input"); Layer* const output = graph.AddLayer(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(layerDesc, "layer"); + auto* const fullyConnectedLayer = graph.AddLayer(layerDesc, "layer"); + auto* const weightsLayer = graph.AddLayer("weights"); + auto* const biasLayer = graph.AddLayer("bias"); float inputsQScale = 1.0f; float outputQScale = 2.0f; - layer->m_Weight = std::make_unique( - TensorInfo({7, 20}, DataType::QAsymmU8, inputsQScale, 0)); - layer->m_Bias = std::make_unique( - 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(weightsInfo); + biasLayer->m_LayerOutput = std::make_shared(biasInfo); + + weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); + biasLayer->GetOutputSlot(0).SetTensorInfo(biasInfo); // create extra layers Layer* const input = graph.AddLayer(0, "input"); Layer* const output = graph.AddLayer(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(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 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 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 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 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 SimpleConvolution2dTestImpl( armnn::Convolution2dQueueDescriptor data; armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(kernelDesc); - armnn::ScopedTensorHandle biasTensor(biasDesc); // Permute the kernel if necessary std::vector kernel = originalKernel; @@ -322,12 +320,6 @@ LayerTestResult 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 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 SimpleConvolution2dNhwcTestImpl( std::unique_ptr weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelDesc); std::unique_ptr 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 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 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 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 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 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 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 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 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 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 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 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 Convolution2dPerAxisQuantTest( std::unique_ptr 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 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 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 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(m_Data.m_Inputs[0])->GetTensor(); arm_compute::ICLTensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); arm_compute::ICLTensor& weights = static_cast(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor* bias = nullptr; if (m_Data.m_Parameters.m_BiasEnabled) { - bias = &static_cast(m_Data.m_Inputs[2])->GetTensor(); + arm_compute::ICLTensor& bias = static_cast(m_Data.m_Inputs[2])->GetTensor(); + m_BiasProxy = std::make_unique(&bias); } // Create Proxy tensor and set the initial tensor handle to it m_InputProxy = std::make_unique(&input); m_OutputProxy = std::make_unique(&output); - + m_WeightsProxy = std::make_unique(&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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); detailsInfo.m_ConvolutionMethod = armnn::Optional(GetConvolutionMethodString(m_ConvolutionMethod)); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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 m_InputProxy; + std::unique_ptr m_WeightsProxy; + std::unique_ptr m_BiasProxy; std::unique_ptr 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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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(); - 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(); - 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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); detailsInfo.m_ConvolutionMethod = armnn::Optional(GetConvolutionMethodString(m_ConvolutionMethod)); + if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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 m_ConvolutionLayer; - std::unique_ptr m_KernelTensor; - std::unique_ptr m_BiasTensor; - + mutable std::unique_ptr m_KernelTensor; + mutable std::unique_ptr 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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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(); - 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(); - 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(descriptor.m_Weight->GetTensorInfo()); + detailsInfo.m_WeightsTensorInfo = armnn::Optional(info.m_InputTensorInfos[1]); if (descriptor.m_Parameters.m_BiasEnabled) { - detailsInfo.m_BiasTensorInfo = armnn::Optional(descriptor.m_Bias->GetTensorInfo()); + detailsInfo.m_BiasTensorInfo = armnn::Optional(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& memoryManager); + virtual void Execute() const override; private: std::unique_ptr m_FullyConnectedLayer; - std::unique_ptr m_WeightsTensor; - std::unique_ptr m_BiasesTensor; - + mutable std::unique_ptr m_WeightsTensor; + mutable std::unique_ptr 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 @@ -57,6 +57,42 @@ void CopyArmComputeTensorData(arm_compute::Tensor& dstTensor, const T* srcData) CopyArmComputeITensorData(srcData, dstTensor); } +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(handle->Map())); + break; + case DataType::Float32: + CopyArmComputeTensorData(tensor, reinterpret_cast(handle->Map())); + break; + case DataType::QAsymmU8: + CopyArmComputeTensorData(tensor, reinterpret_cast(handle->Map())); + break; + case DataType::QSymmS8: + case DataType::QAsymmS8: + CopyArmComputeTensorData(tensor, reinterpret_cast(handle->Map())); + break; + case DataType::Signed32: + CopyArmComputeTensorData(tensor, reinterpret_cast(handle->Map())); + break; + case DataType::QSymmS16: + CopyArmComputeTensorData(tensor, reinterpret_cast(handle->Map())); + break; + case DataType::BFloat16: + CopyArmComputeTensorData(tensor, reinterpret_cast(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) { -- cgit v1.2.1