diff options
| author | Mike Kelly <mike.kelly@arm.com> | 2023-03-08 13:47:17 +0000 |
|---|---|---|
| committer | Francis Murtagh <francis.murtagh@arm.com> | 2023-03-14 16:40:09 +0000 |
| commit | 3ec3077b4eaedcc0c20ab5774bdbe365da541445 (patch) | |
| tree | d601d2000897dec8691bf64cbddc9036f26b8034 | |
| parent | a088cd00b3cce672d26cdcb4965fc2a86b48f339 (diff) | |
| download | armnn-3ec3077b4eaedcc0c20ab5774bdbe365da541445.tar.gz | |
IVGCVSW-3808 Add ElementwiseBinaryLayer
!android-nn-driver:9329
* Added ElementwiseBinaryLayer that can represent all ElementwiseBinary
operations including Add, Div, Sub, Maximum, Mul and Minimum.
* Updated Delegate to use ElementwiseBinaryLayer instead of the Add,
Div, Sub, Maximum, Mul and Minimum layers.
* Updated Deserializer to use ElementwiseBinaryLayer instead of the Add,
Div, Sub, Maximum, Mul and Minimum layers.
* Updated OnnxParser to use ElementwiseBinaryLayer instead of the Add
layer.
* Updated TfLiteParser to use ElementwiseBinaryLayer instead of the Add,
Div, Sub, Maximum, Mul and Minimum layers.
* Updated CL and Neon tests to use ElementwiseBinaryLayer.
* Updated CL and Neon Backend Specific Optimizations to accept
ElementBinaryLayers as well as Add, Div, Mul, Sub, Maximum and Minimum
layers.
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Change-Id: I7cbb96b60eb01f0e2b57b0541016d48a08b86c75
69 files changed, 1546 insertions, 243 deletions
diff --git a/Android.mk b/Android.mk index 0e4e6f90d4..884624ee63 100644 --- a/Android.mk +++ b/Android.mk @@ -1,5 +1,5 @@ # -# Copyright © 2017 ARM Ltd. All rights reserved. +# Copyright © 2017-2023 ARM Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT # @@ -227,6 +227,7 @@ LOCAL_SRC_FILES := \ src/armnn/layers/DetectionPostProcessLayer.cpp \ src/armnn/layers/DivisionLayer.cpp \ src/armnn/layers/ElementwiseBaseLayer.cpp \ + src/armnn/layers/ElementwiseBinaryLayer.cpp \ src/armnn/layers/ElementwiseUnaryLayer.cpp \ src/armnn/layers/FakeQuantizationLayer.cpp \ src/armnn/layers/FillLayer.cpp \ diff --git a/CMakeLists.txt b/CMakeLists.txt index 3b788cd056..be6c4c6a6c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,6 +1,6 @@ # -# Copyright © 2018-2023 Arm Ltd and Contributors. All rights reserved. -# Copyright 2020 NXP +# Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. +# Copyright © 2020 NXP # SPDX-License-Identifier: MIT # cmake_minimum_required (VERSION 3.0.2) # 3.0.2 required for return() statement used in AddDllCopyCommands.cmake @@ -222,6 +222,8 @@ list(APPEND armnn_sources src/armnn/layers/DetectionPostProcessLayer.cpp src/armnn/layers/ElementwiseBaseLayer.hpp src/armnn/layers/ElementwiseBaseLayer.cpp + src/armnn/layers/ElementwiseBinaryLayer.hpp + src/armnn/layers/ElementwiseBinaryLayer.cpp src/armnn/layers/ElementwiseUnaryLayer.hpp src/armnn/layers/ElementwiseUnaryLayer.cpp src/armnn/layers/FakeQuantizationLayer.hpp diff --git a/delegate/src/ElementwiseBinary.hpp b/delegate/src/ElementwiseBinary.hpp index 52c6b2434b..fa9021b5c1 100644 --- a/delegate/src/ElementwiseBinary.hpp +++ b/delegate/src/ElementwiseBinary.hpp @@ -27,15 +27,17 @@ TfLiteStatus ValidateAddOperator(DelegateData& delegateData, bool isSupported = false; auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported) { + std::vector<armnn::TensorInfo> infos { inputInfo1, inputInfo2, outputInfo }; FORWARD_LAYER_SUPPORT_FUNC("ADD", tfLiteContext, - IsAdditionSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputInfo, + armnn::BinaryOperation::Add); }; validateFunc(outputInfo, isSupported); @@ -54,13 +56,14 @@ TfLiteStatus ValidateDivOperator(DelegateData& delegateData, { FORWARD_LAYER_SUPPORT_FUNC("DIV", tfLiteContext, - IsDivisionSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputTensorInfo, + armnn::BinaryOperation::Div); }; validateFunc(outputInfo, isSupported); @@ -107,13 +110,14 @@ TfLiteStatus ValidateMaximumOperator(DelegateData& delegateData, { FORWARD_LAYER_SUPPORT_FUNC("MAXIMUM", tfLiteContext, - IsMaximumSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputTensorInfo, + armnn::BinaryOperation::Maximum); }; validateFunc(outputInfo, isSupported); @@ -131,13 +135,14 @@ TfLiteStatus ValidateMinimumOperator(DelegateData& delegateData, { FORWARD_LAYER_SUPPORT_FUNC("MINIMUM", tfLiteContext, - IsMinimumSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputTensorInfo, + armnn::BinaryOperation::Minimum); }; validateFunc(outputInfo, isSupported); @@ -155,13 +160,14 @@ TfLiteStatus ValidateMulOperator(DelegateData& delegateData, { FORWARD_LAYER_SUPPORT_FUNC("MUL", tfLiteContext, - IsMultiplicationSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputTensorInfo, + armnn::BinaryOperation::Mul); }; validateFunc(outputInfo, isSupported); @@ -179,13 +185,14 @@ TfLiteStatus ValidateSubOperator(DelegateData& delegateData, { FORWARD_LAYER_SUPPORT_FUNC("SUB", tfLiteContext, - IsSubtractionSupported, + IsElementwiseBinarySupported, delegateData.m_Backends, isSupported, armnn::BackendId(), inputInfo1, inputInfo2, - outputTensorInfo); + outputTensorInfo, + armnn::BinaryOperation::Sub); }; validateFunc(outputInfo, isSupported); @@ -196,7 +203,8 @@ std::pair<armnn::IConnectableLayer*, armnn::IConnectableLayer*> AddFloorDivLayer DelegateData& delegateData, const armnn::TensorInfo& outputTensorInfo) { - armnn::IConnectableLayer* divisionLayer = delegateData.m_Network->AddDivisionLayer(); + armnn::IConnectableLayer* divisionLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Div); // if the output of the div is Signed32 the Floor layer is not required if (armnn::DataType::Signed32 == outputTensorInfo.GetDataType()) { @@ -330,10 +338,12 @@ TfLiteStatus VisitElementwiseBinaryOperator(DelegateData& delegateData, switch(elementwiseBinaryOperatorCode) { case kTfLiteBuiltinAdd: - elementwiseBinaryLayer = delegateData.m_Network->AddAdditionLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Add); break; case kTfLiteBuiltinDiv: - elementwiseBinaryLayer = delegateData.m_Network->AddDivisionLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Div); break; case kTfLiteBuiltinFloorDiv: { @@ -343,16 +353,20 @@ TfLiteStatus VisitElementwiseBinaryOperator(DelegateData& delegateData, } break; case kTfLiteBuiltinMaximum: - elementwiseBinaryLayer = delegateData.m_Network->AddMaximumLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Maximum); break; case kTfLiteBuiltinMinimum: - elementwiseBinaryLayer = delegateData.m_Network->AddMinimumLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Minimum); break; case kTfLiteBuiltinMul: - elementwiseBinaryLayer = delegateData.m_Network->AddMultiplicationLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Mul); break; case kTfLiteBuiltinSub: - elementwiseBinaryLayer = delegateData.m_Network->AddSubtractionLayer(); + elementwiseBinaryLayer = delegateData.m_Network->AddElementwiseBinaryLayer( + armnn::BinaryOperation::Sub); break; default: return kTfLiteError; diff --git a/include/armnn/BackendHelper.hpp b/include/armnn/BackendHelper.hpp index cf60b01ddd..85aabe0dd2 100644 --- a/include/armnn/BackendHelper.hpp +++ b/include/armnn/BackendHelper.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2019,2021-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -161,6 +161,12 @@ public: const TensorInfo& output, Optional<std::string&> reasonIfUnsupported = EmptyOptional()); + bool IsElementwiseBinarySupported(const TensorInfo& input0, + const TensorInfo& input1, + const TensorInfo& output, + const ElementwiseBinaryDescriptor& descriptor, + Optional<std::string&> reasonIfUnsupported = EmptyOptional()); + bool IsElementwiseUnarySupported(const TensorInfo& input, const TensorInfo& output, const ElementwiseUnaryDescriptor& descriptor, diff --git a/include/armnn/Descriptors.hpp b/include/armnn/Descriptors.hpp index 7c6942ce58..9ff894f1b0 100644 --- a/include/armnn/Descriptors.hpp +++ b/include/armnn/Descriptors.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -105,6 +105,26 @@ struct ComparisonDescriptor : BaseDescriptor ComparisonOperation m_Operation; }; +/// A ElementwiseBinaryDescriptor for the ElementwiseBinaryLayer +struct ElementwiseBinaryDescriptor : BaseDescriptor +{ + ElementwiseBinaryDescriptor() + : ElementwiseBinaryDescriptor(BinaryOperation::Add) + {} + + ElementwiseBinaryDescriptor(BinaryOperation operation) + : m_Operation(operation) + {} + + bool operator ==(const ElementwiseBinaryDescriptor &rhs) const + { + return m_Operation == rhs.m_Operation; + } + + /// Specifies the elementwiseBinary operation to execute + BinaryOperation m_Operation; +}; + /// A ElementwiseUnaryDescriptor for the ElementwiseUnaryLayer struct ElementwiseUnaryDescriptor : BaseDescriptor { diff --git a/include/armnn/DescriptorsFwd.hpp b/include/armnn/DescriptorsFwd.hpp index 08833504bb..2c25a49f00 100644 --- a/include/armnn/DescriptorsFwd.hpp +++ b/include/armnn/DescriptorsFwd.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -20,6 +20,7 @@ struct Convolution2dDescriptor; struct Convolution3dDescriptor; struct DepthwiseConvolution2dDescriptor; struct DetectionPostProcessDescriptor; +struct ElementwiseBinaryDescriptor; struct ElementwiseUnaryDescriptor; struct FakeQuantizationDescriptor; struct FillDescriptor; diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp index c944d09d41..4eac0cfe2a 100644 --- a/include/armnn/INetwork.hpp +++ b/include/armnn/INetwork.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -349,9 +349,16 @@ public: const ConstTensor& anchors, const char* name = nullptr); + /// Add an ElementwiseBinary layer to the network. + /// @param name - Optional name for the layer. + /// @param desc - Descriptor for the elementwiseBinary operations. + /// @return - Interface for configuring the layer. + IConnectableLayer* AddElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& elementwiseUnaryDescriptor, + const char* name = nullptr); + /// Add an ElementwiseUnary layer to the network. /// @param name - Optional name for the layer. - /// @param desc - Descriptor for the elementwiseUnary operation. + /// @param desc - Descriptor for the elementwiseUnary operations. /// @return - Interface for configuring the layer. IConnectableLayer* AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor& elementwiseUnaryDescriptor, const char* name = nullptr); diff --git a/include/armnn/Types.hpp b/include/armnn/Types.hpp index 4dc613bef3..3c64e82a1e 100644 --- a/include/armnn/Types.hpp +++ b/include/armnn/Types.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2018-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -133,6 +133,16 @@ enum class UnaryOperation Sin = 7 }; +enum class BinaryOperation +{ + Add = 0, + Div = 1, + Maximum = 2, + Minimum = 3, + Mul = 4, + Sub = 5 +}; + enum class PoolingAlgorithm { Max = 0, @@ -385,6 +395,7 @@ using InferenceTimingPair = std::pair<HighResolutionClock, HighResolutionClock>; /// This list uses X macro technique. /// See https://en.wikipedia.org/wiki/X_Macro for more info +// New layers should be added at last position to minimize instability. #define LIST_OF_LAYER_TYPE \ X(Activation) \ X(Addition) \ @@ -458,8 +469,9 @@ using InferenceTimingPair = std::pair<HighResolutionClock, HighResolutionClock>; X(Pooling3d) \ X(GatherNd) \ X(BatchMatMul) \ + X(ElementwiseBinary) \ -// New layers should be added at last to minimize instability. +// New layers should be added at last position to minimize instability. /// When adding a new layer, adapt also the LastLayer enum value in the /// enum class LayerType below @@ -469,7 +481,7 @@ enum class LayerType LIST_OF_LAYER_TYPE #undef X FirstLayer = Activation, - LastLayer = BatchMatMul + LastLayer = ElementwiseBinary }; const char* GetLayerTypeAsCString(LayerType type); diff --git a/include/armnn/TypesUtils.hpp b/include/armnn/TypesUtils.hpp index 4dd5360591..fbe9e8f670 100644 --- a/include/armnn/TypesUtils.hpp +++ b/include/armnn/TypesUtils.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2018-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -72,6 +72,20 @@ constexpr char const* GetComparisonOperationAsCString(ComparisonOperation operat } } +constexpr char const* GetBinaryOperationAsCString(BinaryOperation operation) +{ + switch (operation) + { + case BinaryOperation::Add: return "Add"; + case BinaryOperation::Div: return "Div"; + case BinaryOperation::Maximum: return "Maximum"; + case BinaryOperation::Minimum: return "Minimum"; + case BinaryOperation::Mul: return "Mul"; + case BinaryOperation::Sub: return "Sub"; + default: return "Unknown"; + } +} + constexpr char const* GetUnaryOperationAsCString(UnaryOperation operation) { switch (operation) diff --git a/include/armnn/backends/WorkloadData.hpp b/include/armnn/backends/WorkloadData.hpp index 7a0a765f8e..2abd26740a 100644 --- a/include/armnn/backends/WorkloadData.hpp +++ b/include/armnn/backends/WorkloadData.hpp @@ -663,6 +663,11 @@ struct ComparisonQueueDescriptor : QueueDescriptorWithParameters<ComparisonDescr void Validate(const WorkloadInfo& workloadInfo) const; }; +struct ElementwiseBinaryQueueDescriptor : QueueDescriptorWithParameters<ElementwiseBinaryDescriptor> +{ + void Validate(const WorkloadInfo& workloadInfo) const; +}; + struct ElementwiseUnaryQueueDescriptor : QueueDescriptorWithParameters<ElementwiseUnaryDescriptor> { void Validate(const WorkloadInfo& workloadInfo) const; diff --git a/include/armnnTestUtils/MockBackend.hpp b/include/armnnTestUtils/MockBackend.hpp index e5378bf74a..05df1b46b1 100644 --- a/include/armnnTestUtils/MockBackend.hpp +++ b/include/armnnTestUtils/MockBackend.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -320,6 +320,11 @@ public: reasonIfUnsupported); } } + case LayerType::ElementwiseBinary: + { + auto elementwiseDesc = *(PolymorphicDowncast<const ElementwiseBinaryDescriptor*>(&descriptor)); + return (elementwiseDesc.m_Operation == BinaryOperation::Add); + } default: return false; } diff --git a/src/armnn/BackendHelper.cpp b/src/armnn/BackendHelper.cpp index af38ce8839..580c52c568 100644 --- a/src/armnn/BackendHelper.cpp +++ b/src/armnn/BackendHelper.cpp @@ -600,6 +600,22 @@ bool LayerSupportHandle::IsDivisionSupported(const TensorInfo& input0, reasonIfUnsupported); } +bool LayerSupportHandle::IsElementwiseBinarySupported(const TensorInfo &input0, + const TensorInfo &input1, + const TensorInfo &output, + const ElementwiseBinaryDescriptor &descriptor, + Optional<std::string &> reasonIfUnsupported) +{ + TensorInfos infos{input0, input1, output}; + + return m_LayerSupport->IsLayerSupported(LayerType::ElementwiseBinary, + infos, + descriptor, + EmptyOptional(), + EmptyOptional(), + reasonIfUnsupported); +} + bool LayerSupportHandle::IsElementwiseUnarySupported(const TensorInfo& input, const TensorInfo& output, const ElementwiseUnaryDescriptor& descriptor, diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp index 43862d5072..f634272316 100644 --- a/src/armnn/LayersFwd.hpp +++ b/src/armnn/LayersFwd.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -27,6 +27,7 @@ #include "layers/DequantizeLayer.hpp" #include "layers/DetectionPostProcessLayer.hpp" #include "layers/DivisionLayer.hpp" +#include "layers/ElementwiseBinaryLayer.hpp" #include "layers/ElementwiseUnaryLayer.hpp" #include "layers/FakeQuantizationLayer.hpp" #include "layers/FillLayer.hpp" @@ -127,6 +128,7 @@ DECLARE_LAYER(DepthwiseConvolution2d) DECLARE_LAYER(Dequantize) DECLARE_LAYER(DetectionPostProcess) DECLARE_LAYER(Division) +DECLARE_LAYER(ElementwiseBinary) DECLARE_LAYER(ElementwiseUnary) DECLARE_LAYER(FakeQuantization) DECLARE_LAYER(Fill) diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp index 08d3280cfe..9ebb67b593 100644 --- a/src/armnn/Network.cpp +++ b/src/armnn/Network.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017, 2022-2023 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -20,7 +20,6 @@ #include <backendsCommon/TensorHandleFactoryRegistry.hpp> #include <armnn/Exceptions.hpp> -#include <armnn/Utils.hpp> #include <armnn/TypesUtils.hpp> #include <armnn/BackendRegistry.hpp> #include <armnn/Logging.hpp> @@ -36,10 +35,8 @@ #include <fcntl.h> #include <algorithm> -#include <fstream> #include <memory> #include <vector> -#include <algorithm> namespace armnn { @@ -58,7 +55,6 @@ IConnectableLayer* INetwork::AddInputLayer(LayerBindingId id, const char* name) return pNetworkImpl->AddInputLayer(id, name); } - IConnectableLayer* INetwork::AddArgMinMaxLayer(const ArgMinMaxDescriptor& desc, const char* name) { @@ -126,6 +122,11 @@ IConnectableLayer* INetwork::AddDetectionPostProcessLayer( return pNetworkImpl->AddDetectionPostProcessLayer(descriptor, anchors, name); } +IConnectableLayer* INetwork::AddElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& elementwiseBinaryDescriptor, + const char* name) +{ + return pNetworkImpl->AddElementwiseBinaryLayer(elementwiseBinaryDescriptor, name); +} IConnectableLayer* INetwork::AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor& elementwiseUnaryDescriptor, const char* name) @@ -133,7 +134,6 @@ IConnectableLayer* INetwork::AddElementwiseUnaryLayer(const ElementwiseUnaryDesc return pNetworkImpl->AddElementwiseUnaryLayer(elementwiseUnaryDescriptor, name); } - IConnectableLayer* INetwork::AddFillLayer(const FillDescriptor& fillDescriptor, const char* name) { @@ -1853,6 +1853,12 @@ IConnectableLayer* NetworkImpl::AddComparisonLayer(const ComparisonDescriptor& c return m_Graph->AddLayer<ComparisonLayer>(comparisonDescriptor, name); } +IConnectableLayer* NetworkImpl::AddElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& elementwiseBinaryDesc, + const char* name) +{ + return m_Graph->AddLayer<ElementwiseBinaryLayer>(elementwiseBinaryDesc, name); +} + IConnectableLayer* NetworkImpl::AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor& elementwiseUnaryDescriptor, const char* name) { diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp index a37a4be218..03642ce993 100644 --- a/src/armnn/Network.hpp +++ b/src/armnn/Network.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -95,6 +95,9 @@ public: IConnectableLayer* AddDivisionLayer(const char* name = nullptr); + IConnectableLayer* AddElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& elementwiseBinaryDescriptor, + const char* name = nullptr); + IConnectableLayer* AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor& elementwiseUnaryDescriptor, const char* name = nullptr); diff --git a/src/armnn/layers/ElementwiseBinaryLayer.cpp b/src/armnn/layers/ElementwiseBinaryLayer.cpp new file mode 100644 index 0000000000..ae1813f33a --- /dev/null +++ b/src/armnn/layers/ElementwiseBinaryLayer.cpp @@ -0,0 +1,89 @@ +// +// Copyright © 2023 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "ElementwiseBinaryLayer.hpp" + +#include "LayerCloneBase.hpp" + +namespace armnn +{ + +ElementwiseBinaryLayer::ElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& param, const char* name) + : LayerWithParameters(2, 1, LayerType::ElementwiseBinary, param, name) +{ +} + +std::unique_ptr<IWorkload> ElementwiseBinaryLayer::CreateWorkload(const IWorkloadFactory& factory) const +{ + ElementwiseBinaryQueueDescriptor descriptor; + SetAdditionalInfo(descriptor); + + return factory.CreateWorkload(LayerType::ElementwiseBinary, descriptor, PrepInfoAndDesc(descriptor)); +} + +ElementwiseBinaryLayer* ElementwiseBinaryLayer::Clone(Graph& graph) const +{ + return CloneBase<ElementwiseBinaryLayer>(graph, m_Param, GetName()); +} + +std::vector<TensorShape> ElementwiseBinaryLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const +{ + ARMNN_ASSERT(inputShapes.size() == 2); + TensorShape input0 = inputShapes[0]; + TensorShape input1 = inputShapes[1]; + + if (inputShapes[0].GetNumDimensions() < inputShapes[1].GetNumDimensions()) + { + input1 = inputShapes[0]; + input0 = inputShapes[1]; + } + + unsigned int numDims = input0.GetNumDimensions(); + unsigned int shiftedDims = input0.GetNumDimensions() - input1.GetNumDimensions(); + + // Get the max of the inputs. + std::vector<unsigned int> dims(numDims); + for (unsigned int i = shiftedDims; i < numDims; i++) + { + unsigned int dim0 = input0[i]; + unsigned int dim1 = input1[i - shiftedDims]; + + // Validate inputs are broadcast compatible. + ARMNN_ASSERT_MSG(dim0 == dim1 || dim0 == 1 || dim1 == 1, + "Dimensions should either match or one should be of size 1."); + + dims[i] = std::max(dim0, dim1); + } + + // Fill in the rest of the shifted dimensions. + for (unsigned int i = 0; i < shiftedDims; i++) + { + dims[i] = input0[i]; + } + + return std::vector<TensorShape>({ TensorShape(numDims, dims.data()) }); +} + +void ElementwiseBinaryLayer::ValidateTensorShapesFromInputs() +{ + VerifyLayerConnections(2, CHECK_LOCATION()); + + const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape(); + + VerifyShapeInferenceType(outputShape, m_ShapeInferenceMethod); + + auto inferredShapes = InferOutputShapes({ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(), + GetInputSlot(1).GetConnection()->GetTensorInfo().GetShape() }); + + ARMNN_ASSERT(inferredShapes.size() == 1); + + ValidateAndCopyShape(outputShape, inferredShapes[0], m_ShapeInferenceMethod, GetLayerTypeAsCString(GetType())); +} + +void ElementwiseBinaryLayer::ExecuteStrategy(IStrategy& strategy) const +{ + strategy.ExecuteStrategy(this, GetParameters(), {}, GetName()); +} +} // namespace armnn diff --git a/src/armnn/layers/ElementwiseBinaryLayer.hpp b/src/armnn/layers/ElementwiseBinaryLayer.hpp new file mode 100644 index 0000000000..78e3f41f9e --- /dev/null +++ b/src/armnn/layers/ElementwiseBinaryLayer.hpp @@ -0,0 +1,48 @@ +// +// Copyright © 2023 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "LayerWithParameters.hpp" + +namespace armnn +{ + +/// This layer represents a elementwiseBinary operation. +class ElementwiseBinaryLayer : public LayerWithParameters<ElementwiseBinaryDescriptor> +{ +public: + /// Makes a workload for the elementwiseBinary type + /// @param [in] graph The graph where this layer can be found + /// @param [in] factory The workload factory which will create the workload + /// @return A pointer to the created workload, or nullptr if not created + virtual std::unique_ptr<IWorkload> CreateWorkload(const IWorkloadFactory& factory) const override; + + /// Creates a dynamically-allocated copy of this layer + /// @param [in] graph The graph into which this layer is being cloned + ElementwiseBinaryLayer* Clone(Graph& graph) const override; + + /// Returns inputShapes by default. + /// @param [in] inputShapes The input shapes layer has. + /// @return A vector to the inferred output shape. + std::vector<TensorShape> InferOutputShapes(const std::vector<TensorShape>& inputShapes) const override; + + /// Check if the input tensor shape(s) will lead to a valid configuration + /// of @ref ElementwiseBinaryLayer + void ValidateTensorShapesFromInputs() override; + + void ExecuteStrategy(IStrategy& strategy) const override; + +protected: + /// Constructor to create a ElementwiseBinaryLayer + /// @param [in] param ElementwiseBinaryDescriptor to configure the ElementwiseBinaryLayer + /// @param [in] name Optional name for the layer + ElementwiseBinaryLayer(const ElementwiseBinaryDescriptor& param, const char* name); + + /// Default destructor + ~ElementwiseBinaryLayer() = default; +}; + +} // namespace armnn diff --git a/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp b/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp index b9e8584a5a..dbde72b917 100644 --- a/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp +++ b/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -18,7 +18,7 @@ namespace optimizations static const std::set<armnn::LayerType> broadcastOps{ LayerType::Addition, LayerType::Division, LayerType::Maximum, LayerType::Minimum, LayerType::Multiplication, LayerType::Prelu, - LayerType::Subtraction }; + LayerType::Subtraction, LayerType::ElementwiseBinary }; class AddBroadcastReshapeLayerImpl { diff --git a/src/armnn/optimizations/MovePermuteUp.hpp b/src/armnn/optimizations/MovePermuteUp.hpp index ae8a28cba4..19078b3bc6 100644 --- a/src/armnn/optimizations/MovePermuteUp.hpp +++ b/src/armnn/optimizations/MovePermuteUp.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017-2018,2020,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -68,6 +68,12 @@ private: case LayerType::MemCopy: case LayerType::Multiplication: return true; + case LayerType::ElementwiseBinary: + { + auto descriptor = PolymorphicDowncast<const ElementwiseBinaryDescriptor*>(&base.GetParameters()); + return (descriptor->m_Operation == BinaryOperation::Add || + descriptor->m_Operation == BinaryOperation::Mul); + } default: return false; } diff --git a/src/armnn/optimizations/MoveTransposeUp.hpp b/src/armnn/optimizations/MoveTransposeUp.hpp index 999a4ebe42..40f6b9cdac 100644 --- a/src/armnn/optimizations/MoveTransposeUp.hpp +++ b/src/armnn/optimizations/MoveTransposeUp.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd. All rights reserved. +// Copyright © 2020,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -68,6 +68,12 @@ private: case LayerType::MemCopy: case LayerType::Multiplication: return true; + case LayerType::ElementwiseBinary: + { + auto descriptor = PolymorphicDowncast<const ElementwiseBinaryDescriptor*>(&base.GetParameters()); + return (descriptor->m_Operation == BinaryOperation::Add || + descriptor->m_Operation == BinaryOperation::Mul); + } default: return false; } diff --git a/src/armnn/test/GraphTests.cpp b/src/armnn/test/GraphTests.cpp index eea7ae824a..b1b1a84ec9 100644 --- a/src/armnn/test/GraphTests.cpp +++ b/src/armnn/test/GraphTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include <GraphUtils.hpp> @@ -36,7 +36,7 @@ TEST_CASE("TopologicalSort") CHECK_NOTHROW(graph.AddLayer<armnn::InputLayer>(0, "layerA")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerB")); - CHECK_NOTHROW(graph.AddLayer<armnn::AdditionLayer>("layerC")); + CHECK_NOTHROW(graph.AddLayer<armnn::ElementwiseBinaryLayer>(armnn::BinaryOperation::Add, "layerC")); CHECK_NOTHROW(graph.AddLayer<armnn::OutputLayer>(0, "output")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerD")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerE")); @@ -82,7 +82,7 @@ TEST_CASE("InsertNewLayerBefore") CHECK_NOTHROW(graph.AddLayer<armnn::InputLayer>(0, "layerA")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerB")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerC")); - CHECK_NOTHROW(graph.AddLayer<armnn::AdditionLayer>("layerD")); + CHECK_NOTHROW(graph.AddLayer<armnn::ElementwiseBinaryLayer>(armnn::BinaryOperation::Add, "layerD")); CHECK_NOTHROW(graph.AddLayer<armnn::OutputLayer>(0, "output")); armnn::Layer* const layerA = GetFirstLayerWithName(graph, "layerA"); @@ -168,7 +168,7 @@ TEST_CASE("InsertNewLayerAfter") CHECK_NOTHROW(graph.AddLayer<armnn::InputLayer>(0, "layerA")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerB")); CHECK_NOTHROW(graph.AddLayer<armnn::ActivationLayer>(activationDefaults, "layerC")); - CHECK_NOTHROW(graph.AddLayer<armnn::AdditionLayer>("layerD")); + CHECK_NOTHROW(graph.AddLayer<armnn::ElementwiseBinaryLayer>(armnn::BinaryOperation::Add, "layerD")); CHECK_NOTHROW(graph.AddLayer<armnn::OutputLayer>(0, "output")); armnn::Layer* const layerA = GetFirstLayerWithName(graph, "layerA"); @@ -548,7 +548,7 @@ TEST_CASE_FIXTURE(CopyLayersFixture, "CopyLayersAddedBetweenSameLayersHaveDiffer armnn::SplitterLayer* const splitterLayer = graph.AddLayer<armnn::SplitterLayer>(splitterDesc, "splitter"); splitterLayer->SetBackendId(armnn::Compute::GpuAcc); - armnn::AdditionLayer* const additionLayer = graph.AddLayer<armnn::AdditionLayer>("addition"); + auto* const additionLayer = graph.AddLayer<armnn::ElementwiseBinaryLayer>(armnn::BinaryOperation::Add, "addition"); additionLayer->SetBackendId(armnn::Compute::CpuRef); armnn::OutputLayer* const outputLayer = graph.AddLayer<armnn::OutputLayer>(0, "output"); diff --git a/src/armnn/test/NetworkTests.cpp b/src/armnn/test/NetworkTests.cpp index 058f079e46..0bfad4d4d1 100644 --- a/src/armnn/test/NetworkTests.cpp +++ b/src/armnn/test/NetworkTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -32,7 +32,7 @@ TEST_CASE("LayerGuids") { armnn::NetworkImpl net; LayerGuid inputId = net.AddInputLayer(0)->GetGuid(); - LayerGuid addId = net.AddAdditionLayer()->GetGuid(); + LayerGuid addId = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Add)->GetGuid(); LayerGuid outputId = net.AddOutputLayer(0)->GetGuid(); CHECK(inputId != addId); @@ -50,7 +50,7 @@ TEST_CASE("LayerNamesAreOptionalForINetwork") { armnn::INetworkPtr inet(armnn::INetwork::Create()); inet->AddInputLayer(0); - inet->AddAdditionLayer(); + inet->AddElementwiseBinaryLayer(armnn::BinaryOperation::Add); inet->AddActivationLayer(armnn::ActivationDescriptor()); inet->AddOutputLayer(0); } @@ -59,7 +59,7 @@ TEST_CASE("LayerNamesAreOptionalForNetwork") { armnn::NetworkImpl net; net.AddInputLayer(0); - net.AddAdditionLayer(); + net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Add); net.AddActivationLayer(armnn::ActivationDescriptor()); net.AddOutputLayer(0); } @@ -136,13 +136,15 @@ TEST_CASE("NetworkModification") softmaxLayer->GetOutputSlot(0).Connect(batchNormalizationLayer->GetInputSlot(0)); - armnn::IConnectableLayer* const additionLayer = net.AddAdditionLayer("addition"); + armnn::IConnectableLayer* const additionLayer = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Add, + "addition"); CHECK(additionLayer); batchNormalizationLayer->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0)); batchNormalizationLayer->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(1)); - armnn::IConnectableLayer* const multiplicationLayer = net.AddMultiplicationLayer("multiplication"); + armnn::IConnectableLayer* const multiplicationLayer = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Mul, + "multiplication"); CHECK(multiplicationLayer); additionLayer->GetOutputSlot(0).Connect(multiplicationLayer->GetInputSlot(0)); @@ -338,7 +340,7 @@ TEST_CASE("NetworkModification_SplitterAddition") splitterLayer->GetOutputSlot(1).Connect(softmax2Layer->GetInputSlot(0)); // Adds addition layer. - layer = net.AddAdditionLayer("add layer"); + layer = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Add, "add layer"); CHECK(layer); softmax1Layer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); @@ -382,7 +384,7 @@ TEST_CASE("NetworkModification_SplitterMultiplication") splitterLayer->GetOutputSlot(1).Connect(softmax2Layer->GetInputSlot(0)); // Adds multiplication layer. - layer = net.AddMultiplicationLayer("multiplication layer"); + layer = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Mul, "multiplication layer"); CHECK(layer); softmax1Layer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); @@ -634,15 +636,27 @@ TEST_CASE("ObtainConv2DDescriptorFromIConnectableLayer") CHECK(originalDescriptor.m_DataLayout == armnn::DataLayout::NCHW); } -TEST_CASE("CheckNullDescriptor") +TEST_CASE("CheckNotNullDescriptor") { armnn::NetworkImpl net; - armnn::IConnectableLayer* const addLayer = net.AddAdditionLayer(); + armnn::IConnectableLayer* const addLayer = net.AddElementwiseBinaryLayer(armnn::BinaryOperation::Add); CHECK(addLayer); const armnn::BaseDescriptor& descriptor = addLayer->GetParameters(); // additional layer has no descriptor so a NullDescriptor will be returned + CHECK(descriptor.IsNull() == false); +} + +TEST_CASE("CheckNullDescriptor") +{ + armnn::NetworkImpl net; + armnn::IConnectableLayer* const addLayer = net.AddPreluLayer(); + + CHECK(addLayer); + + const armnn::BaseDescriptor& descriptor = addLayer->GetParameters(); + // Prelu has no descriptor so a NullDescriptor will be returned CHECK(descriptor.IsNull() == true); } diff --git a/src/armnn/test/TestNameAndDescriptorLayerVisitor.cpp b/src/armnn/test/TestNameAndDescriptorLayerVisitor.cpp index cfdaaf529b..b8607d1b5f 100644 --- a/src/armnn/test/TestNameAndDescriptorLayerVisitor.cpp +++ b/src/armnn/test/TestNameAndDescriptorLayerVisitor.cpp @@ -1,12 +1,10 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include "TestNameAndDescriptorLayerVisitor.hpp" #include "Network.hpp" -#include <armnn/Exceptions.hpp> - #include <doctest/doctest.h> namespace @@ -84,6 +82,12 @@ armnn::ConcatDescriptor GetDescriptor<armnn::ConcatDescriptor>() } template<> +armnn::ElementwiseBinaryDescriptor GetDescriptor<armnn::ElementwiseBinaryDescriptor>() +{ + return armnn::ElementwiseBinaryDescriptor(armnn::BinaryOperation::Add); +} + +template<> armnn::ElementwiseUnaryDescriptor GetDescriptor<armnn::ElementwiseUnaryDescriptor>() { return armnn::ElementwiseUnaryDescriptor(armnn::UnaryOperation::Abs); @@ -273,12 +277,14 @@ armnn::TransposeDescriptor GetDescriptor<armnn::TransposeDescriptor>() TEST_SUITE("TestNameAndDescriptorLayerVisitor") { -TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Activation, CheckAdditionLayerVisitorNameAndDescriptor) +TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Activation, CheckActivationLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(ArgMinMax, CheckArgMinMaxLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(DepthToSpace, CheckDepthToSpaceLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(BatchToSpaceNd, CheckBatchToSpaceNdLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Comparison, CheckComparisonLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Concat, CheckConcatLayerVisitorNameAndDescriptor) +TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(ElementwiseBinary, + CheckElementwiseBinaryLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(ElementwiseUnary, CheckElementwiseUnaryLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Fill, CheckFillLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Gather, CheckGatherLayerVisitorNameAndDescriptor) @@ -304,7 +310,7 @@ TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(StridedSlice, CheckStridedSlic TEST_CASE_CHECK_LAYER_VISITOR_NAME_AND_DESCRIPTOR(Transpose, CheckTransposeLayerVisitorNameAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(Activation, - CheckAdditionLayerVisitorNameNullptrAndDescriptor) + CheckActivationLayerVisitorNameNullptrAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(ArgMinMax, CheckArgMinMaxLayerVisitorNameNullptrAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(DepthToSpace, @@ -315,6 +321,8 @@ TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(Comparison, CheckComparisonLayerVisitorNameNullptrAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(Concat, CheckConcatLayerVisitorNameNullptrAndDescriptor) +TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(ElementwiseBinary, + CheckElementwiseBinaryLayerVisitorNameNullptrAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(ElementwiseUnary, CheckElementwiseUnaryLayerVisitorNameNullptrAndDescriptor) TEST_CASE_CHECK_LAYER_VISITOR_NAME_NULLPTR_AND_DESCRIPTOR(Fill, diff --git a/src/armnn/test/TestNameAndDescriptorLayerVisitor.hpp b/src/armnn/test/TestNameAndDescriptorLayerVisitor.hpp index b1f9512655..988903518d 100644 --- a/src/armnn/test/TestNameAndDescriptorLayerVisitor.hpp +++ b/src/armnn/test/TestNameAndDescriptorLayerVisitor.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -63,6 +63,7 @@ DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(BatchToSpaceNd) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(Comparison) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(Concat) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(DepthToSpace) +DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(ElementwiseBinary) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(ElementwiseUnary) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(Fill) DECLARE_TEST_NAME_AND_DESCRIPTOR_LAYER_VISITOR_CLASS(Gather) diff --git a/src/armnn/test/optimizations/AddBroadcastReshapeLayerTests.cpp b/src/armnn/test/optimizations/AddBroadcastReshapeLayerTests.cpp index 0636a00234..59dfb862a0 100644 --- a/src/armnn/test/optimizations/AddBroadcastReshapeLayerTests.cpp +++ b/src/armnn/test/optimizations/AddBroadcastReshapeLayerTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -27,7 +27,7 @@ void AddBroadcastReshapeLayerOptimizerTest(const TensorInfo& info0, auto input0 = graph.AddLayer<InputLayer>(0, "input0"); auto input1 = graph.AddLayer<InputLayer>(1, "input1"); - auto add = graph.AddLayer<AdditionLayer>("add"); + auto add = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Add, "add"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input0->GetOutputSlot().SetTensorInfo(info0); input1->GetOutputSlot().SetTensorInfo(info1); @@ -40,7 +40,7 @@ void AddBroadcastReshapeLayerOptimizerTest(const TensorInfo& info0, CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<AdditionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -48,19 +48,19 @@ void AddBroadcastReshapeLayerOptimizerTest(const TensorInfo& info0, // Broadcast reshape layer has been added to the graph correctly CHECK(CheckSequence(graph.cbegin(), graph.cend(), - &IsLayerOfType<InputLayer>, - &IsLayerOfType<InputLayer>, - &IsLayerOfType<ReshapeLayer>, - &IsLayerOfType<AdditionLayer>, - &IsLayerOfType<OutputLayer>)); + &IsLayerOfType<InputLayer>, + &IsLayerOfType<InputLayer>, + &IsLayerOfType<ReshapeLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, + &IsLayerOfType<OutputLayer>)); Layer* const reshapeLayer = GetFirstLayerWithName(graph, reshapeLayerName); CHECK(reshapeLayer); auto addedReshapeTensorInfo = reshapeLayer->GetOutputSlot().GetTensorInfo(); // Tensorshape and the data type are correct - CHECK((addedReshapeTensorInfo.GetShape() == expectedReshapeShape)); - CHECK((addedReshapeTensorInfo.GetDataType() == expectedDataType)); + CHECK_EQ(addedReshapeTensorInfo.GetShape(), expectedReshapeShape); + CHECK_EQ(addedReshapeTensorInfo.GetDataType(), expectedDataType); } TEST_CASE("AddBroadcastReshapeLayerSimpleTest") @@ -121,7 +121,7 @@ TEST_CASE("AddBroadcastReshapeLayerSubtractionTest") auto input0 = graph.AddLayer<InputLayer>(0, "input0"); auto input1 = graph.AddLayer<InputLayer>(1, "input1"); - auto sub = graph.AddLayer<SubtractionLayer>("sub"); + auto sub = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Sub, "sub"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input0->GetOutputSlot().SetTensorInfo(info0); input1->GetOutputSlot().SetTensorInfo(info1); @@ -134,7 +134,7 @@ TEST_CASE("AddBroadcastReshapeLayerSubtractionTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<SubtractionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -145,7 +145,7 @@ TEST_CASE("AddBroadcastReshapeLayerSubtractionTest") &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, &IsLayerOfType<ReshapeLayer>, - &IsLayerOfType<SubtractionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); Layer* const reshapeLayer = GetFirstLayerWithName(graph, "Reshape_for:sub-0"); @@ -153,8 +153,8 @@ TEST_CASE("AddBroadcastReshapeLayerSubtractionTest") auto addedReshapeTensorInfo = reshapeLayer->GetOutputSlot().GetTensorInfo(); // Tensorshape and the data type are correct - CHECK((addedReshapeTensorInfo.GetShape() == TensorShape({ 1, 1, 1, 5 }))); - CHECK((addedReshapeTensorInfo.GetDataType() == DataType::Float32)); + CHECK_EQ(addedReshapeTensorInfo.GetShape(), TensorShape({ 1, 1, 1, 5 })); + CHECK_EQ(addedReshapeTensorInfo.GetDataType(), DataType::Float32); } TEST_CASE("AddBroadcastReshapeLayerDivisionTest") @@ -166,7 +166,7 @@ TEST_CASE("AddBroadcastReshapeLayerDivisionTest") auto input0 = graph.AddLayer<InputLayer>(0, "input0"); auto input1 = graph.AddLayer<InputLayer>(1, "input1"); - auto div = graph.AddLayer<DivisionLayer>("div"); + auto div = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Div, "div"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input0->GetOutputSlot().SetTensorInfo(info0); input1->GetOutputSlot().SetTensorInfo(info1); @@ -179,7 +179,7 @@ TEST_CASE("AddBroadcastReshapeLayerDivisionTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<DivisionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -190,7 +190,7 @@ TEST_CASE("AddBroadcastReshapeLayerDivisionTest") &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, &IsLayerOfType<ReshapeLayer>, - &IsLayerOfType<DivisionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); Layer* const reshapeLayer = GetFirstLayerWithName(graph, "Reshape_for:div-0"); @@ -198,8 +198,8 @@ TEST_CASE("AddBroadcastReshapeLayerDivisionTest") auto addedReshapeTensorInfo = reshapeLayer->GetOutputSlot().GetTensorInfo(); // Tensorshape and the data type are correct - CHECK((addedReshapeTensorInfo.GetShape() == TensorShape({ 1, 1, 4, 5 }))); - CHECK((addedReshapeTensorInfo.GetDataType() == DataType::QAsymmS8)); + CHECK_EQ(addedReshapeTensorInfo.GetShape(), TensorShape({ 1, 1, 4, 5 })); + CHECK_EQ(addedReshapeTensorInfo.GetDataType(), DataType::QAsymmS8); } TEST_CASE("AddBroadcastReshapeLayerMultiplicationTest") @@ -211,7 +211,7 @@ TEST_CASE("AddBroadcastReshapeLayerMultiplicationTest") auto input0 = graph.AddLayer<InputLayer>(0, "input0"); auto input1 = graph.AddLayer<InputLayer>(1, "input1"); - auto mul = graph.AddLayer<MultiplicationLayer>("mul"); + auto mul = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Mul, "mul"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input0->GetOutputSlot().SetTensorInfo(info0); input1->GetOutputSlot().SetTensorInfo(info1); @@ -224,7 +224,7 @@ TEST_CASE("AddBroadcastReshapeLayerMultiplicationTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -235,7 +235,7 @@ TEST_CASE("AddBroadcastReshapeLayerMultiplicationTest") &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, &IsLayerOfType<ReshapeLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); Layer* const reshapeLayer = GetFirstLayerWithName(graph, "Reshape_for:mul-0"); @@ -243,8 +243,8 @@ TEST_CASE("AddBroadcastReshapeLayerMultiplicationTest") auto addedReshapeTensorInfo = reshapeLayer->GetOutputSlot().GetTensorInfo(); // Tensorshape and the data type are correct - CHECK((addedReshapeTensorInfo.GetShape() == TensorShape({ 1, 1, 3, 5 }))); - CHECK((addedReshapeTensorInfo.GetDataType() == DataType::QAsymmU8)); + CHECK_EQ(addedReshapeTensorInfo.GetShape(), TensorShape({ 1, 1, 3, 5 })); + CHECK_EQ(addedReshapeTensorInfo.GetDataType(), DataType::QAsymmU8); } TEST_CASE("AddNoBroadcastReshapeLayerTest") @@ -256,7 +256,7 @@ TEST_CASE("AddNoBroadcastReshapeLayerTest") auto input0 = graph.AddLayer<InputLayer>(0, "input0"); auto input1 = graph.AddLayer<InputLayer>(1, "input1"); - auto mul = graph.AddLayer<MultiplicationLayer>("mul"); + auto mul = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Mul, "mul"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input0->GetOutputSlot().SetTensorInfo(info0); input1->GetOutputSlot().SetTensorInfo(info1); @@ -269,7 +269,7 @@ TEST_CASE("AddNoBroadcastReshapeLayerTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -279,7 +279,7 @@ TEST_CASE("AddNoBroadcastReshapeLayerTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<InputLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); Layer* const reshapeLayer = GetFirstLayerWithName(graph, "Reshape_for:mul-0"); @@ -295,7 +295,7 @@ TEST_CASE("ReshapeParentConstLayerTest") auto input = graph.AddLayer<InputLayer>(0, "input"); auto constant = graph.AddLayer<ConstantLayer>("constant"); - auto mul = graph.AddLayer<MultiplicationLayer>("mul"); + auto mul = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Mul, "mul"); auto output = graph.AddLayer<OutputLayer>(0, "output"); uint8_t tensor[] = { 1, 1, 1, 1, 1 }; @@ -313,7 +313,7 @@ TEST_CASE("ReshapeParentConstLayerTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<ConstantLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -323,7 +323,7 @@ TEST_CASE("ReshapeParentConstLayerTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<ConstantLayer>, - &IsLayerOfType<MultiplicationLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); TensorShape expectedShape = TensorShape{ 1, 1, 1, 5 }; @@ -351,8 +351,8 @@ TEST_CASE("ReshapeParentConstAddLayerMultipleConnectionsTest") auto input = graph.AddLayer<InputLayer>(0, "input"); auto constant = graph.AddLayer<ConstantLayer>("constant"); - auto add1 = graph.AddLayer<AdditionLayer>("add1"); - auto add2 = graph.AddLayer<AdditionLayer>("add2"); + auto add1 = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Add, "add1"); + auto add2 = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Add, "add2"); auto output = graph.AddLayer<OutputLayer>(0, "output"); input->GetOutputSlot().SetTensorInfo(inputInfo); @@ -371,8 +371,8 @@ TEST_CASE("ReshapeParentConstAddLayerMultipleConnectionsTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<InputLayer>, &IsLayerOfType<ConstantLayer>, - &IsLayerOfType<AdditionLayer>, - &IsLayerOfType<AdditionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Run optimizer @@ -384,8 +384,8 @@ TEST_CASE("ReshapeParentConstAddLayerMultipleConnectionsTest") &IsLayerOfType<ConstantLayer>, &IsLayerOfType<ReshapeLayer>, &IsLayerOfType<ReshapeLayer>, - &IsLayerOfType<AdditionLayer>, - &IsLayerOfType<AdditionLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, + &IsLayerOfType<ElementwiseBinaryLayer>, &IsLayerOfType<OutputLayer>)); // Ensure the output shape of the constant hasn't changed. diff --git a/src/armnn/test/optimizations/MovePermuteUpTests.cpp b/src/armnn/test/optimizations/MovePermuteUpTests.cpp index 152e79925b..018286c70d 100644 --- a/src/armnn/test/optimizations/MovePermuteUpTests.cpp +++ b/src/armnn/test/optimizations/MovePermuteUpTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -36,7 +36,7 @@ TEST_CASE("MovePermuteUpTest") head = graph.InsertNewLayer<armnn::ActivationLayer>(head->GetInputSlot(0), armnn::ActivationDescriptor{}, ""); head->GetOutputHandler().SetTensorInfo(info); - head = graph.InsertNewLayer<armnn::AdditionLayer>(head->GetInputSlot(0), ""); + head = graph.InsertNewLayer<armnn::ElementwiseBinaryLayer>(head->GetInputSlot(0), armnn::BinaryOperation::Add, ""); head->GetOutputHandler().SetTensorInfo(info); // Inserts input for 2nd input of Addition. @@ -54,7 +54,7 @@ TEST_CASE("MovePermuteUpTest") head = graph.InsertNewLayer<armnn::MemCopyLayer>(head->GetInputSlot(0), ""); head->GetOutputHandler().SetTensorInfo(info); - head = graph.InsertNewLayer<armnn::MultiplicationLayer>(head->GetInputSlot(0), ""); + head = graph.InsertNewLayer<armnn::ElementwiseBinaryLayer>(head->GetInputSlot(0), armnn::BinaryOperation::Mul, ""); head->GetOutputHandler().SetTensorInfo(info); // Inserts input for 2nd input of Multiplication. @@ -69,9 +69,9 @@ TEST_CASE("MovePermuteUpTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, - &IsLayerOfType<armnn::MultiplicationLayer>, &IsLayerOfType<armnn::MemCopyLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::MemCopyLayer>, &IsLayerOfType<armnn::FloorLayer>, &IsLayerOfType<armnn::FakeQuantizationLayer>, - &IsLayerOfType<armnn::AdditionLayer>, &IsLayerOfType<armnn::ActivationLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::ActivationLayer>, &IsLayerOfType<armnn::PermuteLayer>, &IsLayerOfType<armnn::OutputLayer>)); armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(MovePermuteUp())); @@ -80,10 +80,11 @@ TEST_CASE("MovePermuteUpTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::PermuteLayer>, &IsLayerOfType<armnn::PermuteLayer>, - &IsLayerOfType<armnn::PermuteLayer>, &IsLayerOfType<armnn::MultiplicationLayer>, + &IsLayerOfType<armnn::PermuteLayer>, &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::MemCopyLayer>, &IsLayerOfType<armnn::FloorLayer>, - &IsLayerOfType<armnn::FakeQuantizationLayer>, &IsLayerOfType<armnn::AdditionLayer>, - &IsLayerOfType<armnn::ActivationLayer>, &IsLayerOfType<armnn::OutputLayer>)); + &IsLayerOfType<armnn::FakeQuantizationLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::ActivationLayer>, + &IsLayerOfType<armnn::OutputLayer>)); std::list<std::string> testRelatedLayers = { permuteLayerName }; diff --git a/src/armnn/test/optimizations/MoveTransposeUpTests.cpp b/src/armnn/test/optimizations/MoveTransposeUpTests.cpp index 09bf9ae7d9..6a6010cb4d 100644 --- a/src/armnn/test/optimizations/MoveTransposeUpTests.cpp +++ b/src/armnn/test/optimizations/MoveTransposeUpTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd. All rights reserved. +// Copyright © 2020-2021,2023 Arm Ltd. All rights reserved. // SPDX-License-Identifier: MIT // @@ -37,7 +37,7 @@ TEST_CASE("MoveTransposeUpTest") head = graph.InsertNewLayer<armnn::ActivationLayer>(head->GetInputSlot(0), armnn::ActivationDescriptor{}, ""); head->GetOutputHandler().SetTensorInfo(info); - head = graph.InsertNewLayer<armnn::AdditionLayer>(head->GetInputSlot(0), ""); + head = graph.InsertNewLayer<armnn::ElementwiseBinaryLayer>(head->GetInputSlot(0), armnn::BinaryOperation::Add, ""); head->GetOutputHandler().SetTensorInfo(info); // Inserts input for 2nd input of Addition. @@ -55,7 +55,7 @@ TEST_CASE("MoveTransposeUpTest") head = graph.InsertNewLayer<armnn::MemCopyLayer>(head->GetInputSlot(0), ""); head->GetOutputHandler().SetTensorInfo(info); - head = graph.InsertNewLayer<armnn::MultiplicationLayer>(head->GetInputSlot(0), ""); + head = graph.InsertNewLayer<armnn::ElementwiseBinaryLayer>(head->GetInputSlot(0), armnn::BinaryOperation::Mul, ""); head->GetOutputHandler().SetTensorInfo(info); // Inserts input for 2nd input of Multiplication. @@ -70,9 +70,9 @@ TEST_CASE("MoveTransposeUpTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, - &IsLayerOfType<armnn::MultiplicationLayer>, &IsLayerOfType<armnn::MemCopyLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::MemCopyLayer>, &IsLayerOfType<armnn::FloorLayer>, &IsLayerOfType<armnn::FakeQuantizationLayer>, - &IsLayerOfType<armnn::AdditionLayer>, &IsLayerOfType<armnn::ActivationLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::ActivationLayer>, &IsLayerOfType<armnn::TransposeLayer>, &IsLayerOfType<armnn::OutputLayer>)); armnn::Optimizer::Pass(graph, armnn::MakeOptimizations(MoveTransposeUp())); @@ -81,10 +81,11 @@ TEST_CASE("MoveTransposeUpTest") CHECK(CheckSequence(graph.cbegin(), graph.cend(), &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::InputLayer>, &IsLayerOfType<armnn::TransposeLayer>, &IsLayerOfType<armnn::TransposeLayer>, - &IsLayerOfType<armnn::TransposeLayer>, &IsLayerOfType<armnn::MultiplicationLayer>, + &IsLayerOfType<armnn::TransposeLayer>, &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::MemCopyLayer>, &IsLayerOfType<armnn::FloorLayer>, - &IsLayerOfType<armnn::FakeQuantizationLayer>, &IsLayerOfType<armnn::AdditionLayer>, - &IsLayerOfType<armnn::ActivationLayer>, &IsLayerOfType<armnn::OutputLayer>)); + &IsLayerOfType<armnn::FakeQuantizationLayer>, + &IsLayerOfType<armnn::ElementwiseBinaryLayer>, &IsLayerOfType<armnn::ActivationLayer>, + &IsLayerOfType<armnn::OutputLayer>)); std::list<std::string> testRelatedLayers = { transposeLayerName }; diff --git a/src/armnnDeserializer/Deserializer.cpp b/src/armnnDeserializer/Deserializer.cpp index 702b060512..ed921880e0 100644 --- a/src/armnnDeserializer/Deserializer.cpp +++ b/src/armnnDeserializer/Deserializer.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -229,6 +229,7 @@ m_ParserFunctions(Layer_MAX+1, &IDeserializer::DeserializerImpl::ParseUnsupporte m_ParserFunctions[Layer_DequantizeLayer] = &DeserializerImpl::ParseDequantize; m_ParserFunctions[Layer_DetectionPostProcessLayer] = &DeserializerImpl::ParseDetectionPostProcess; m_ParserFunctions[Layer_DivisionLayer] = &DeserializerImpl::ParseDivision; + m_ParserFunctions[Layer_ElementwiseBinaryLayer] = &DeserializerImpl::ParseElementwiseBinary; m_ParserFunctions[Layer_ElementwiseUnaryLayer] = &DeserializerImpl::ParseElementwiseUnary; m_ParserFunctions[Layer_EqualLayer] = &DeserializerImpl::ParseEqual; m_ParserFunctions[Layer_FullyConnectedLayer] = &DeserializerImpl::ParseFullyConnected; @@ -325,6 +326,8 @@ LayerBaseRawPtr IDeserializer::DeserializerImpl::GetBaseLayer(const GraphPtr& gr return graphPtr->layers()->Get(layerIndex)->layer_as_DivisionLayer()->base(); case Layer::Layer_EqualLayer: return graphPtr->layers()->Get(layerIndex)->layer_as_EqualLayer()->base(); + case Layer::Layer_ElementwiseBinaryLayer: + return graphPtr->layers()->Get(layerIndex)->layer_as_ElementwiseBinaryLayer()->base(); case Layer::Layer_ElementwiseUnaryLayer: return graphPtr->layers()->Get(layerIndex)->layer_as_ElementwiseUnaryLayer()->base(); case Layer::Layer_FullyConnectedLayer: @@ -562,7 +565,28 @@ armnn::LogicalBinaryOperation ToLogicalBinaryOperation(armnnSerializer::LogicalB } } -armnn::UnaryOperation ToUnaryOperation(armnnSerializer::UnaryOperation operation) +armnn::BinaryOperation ToElementwiseBinaryOperation(armnnSerializer::BinaryOperation operation) +{ + switch (operation) + { + case armnnSerializer::BinaryOperation::BinaryOperation_Add: + return armnn::BinaryOperation::Add; + case armnnSerializer::BinaryOperation::BinaryOperation_Div: + return armnn::BinaryOperation::Div; + case armnnSerializer::BinaryOperation::BinaryOperation_Maximum: + return armnn::BinaryOperation::Maximum; + case armnnSerializer::BinaryOperation::BinaryOperation_Minimum: + return armnn::BinaryOperation::Minimum; + case armnnSerializer::BinaryOperation::BinaryOperation_Mul: + return armnn::BinaryOperation::Mul; + case armnnSerializer::BinaryOperation::BinaryOperation_Sub: + return armnn::BinaryOperation::Sub; + default: + throw armnn::InvalidArgumentException("Binary operation unknown"); + } +} + +armnn::UnaryOperation ToElementwiseUnaryOperation(armnnSerializer::UnaryOperation operation) { switch (operation) { @@ -1226,7 +1250,8 @@ void IDeserializer::DeserializerImpl::ParseAdd(GraphPtr graph, unsigned int laye CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddAdditionLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Add); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); @@ -1745,7 +1770,8 @@ void IDeserializer::DeserializerImpl::ParseDivision(GraphPtr graph, unsigned int CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddDivisionLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Div); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); @@ -1935,7 +1961,8 @@ void IDeserializer::DeserializerImpl::ParseMinimum(GraphPtr graph, unsigned int CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddMinimumLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Minimum); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); @@ -1955,7 +1982,8 @@ void IDeserializer::DeserializerImpl::ParseMaximum(GraphPtr graph, unsigned int CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddMaximumLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Maximum); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); @@ -2030,6 +2058,33 @@ void IDeserializer::DeserializerImpl::ParseComparison(GraphPtr graph, unsigned i RegisterOutputSlots(graph, layerIndex, layer); } +void IDeserializer::DeserializerImpl::ParseElementwiseBinary(GraphPtr graph, unsigned int layerIndex) +{ + CHECK_LAYERS(graph, 0, layerIndex); + CHECK_LOCATION(); + + auto inputs = GetInputs(graph, layerIndex); + CHECK_VALID_SIZE(inputs.size(), 2); + + auto outputs = GetOutputs(graph, layerIndex); + CHECK_VALID_SIZE(outputs.size(), 1); + + auto fbLayer = graph->layers()->Get(layerIndex)->layer_as_ElementwiseBinaryLayer(); + auto fbDescriptor = fbLayer->descriptor(); + + armnn::ElementwiseBinaryDescriptor descriptor; + descriptor.m_Operation = ToElementwiseBinaryOperation(fbDescriptor->operation()); + + const std::string& layerName = GetLayerName(graph, layerIndex); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); + + armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); + layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); + + RegisterInputSlots(graph, layerIndex, layer); + RegisterOutputSlots(graph, layerIndex, layer); +} + void IDeserializer::DeserializerImpl::ParseElementwiseUnary(GraphPtr graph, unsigned int layerIndex) { CHECK_LAYERS(graph, 0, layerIndex); @@ -2045,7 +2100,7 @@ void IDeserializer::DeserializerImpl::ParseElementwiseUnary(GraphPtr graph, unsi auto fbDescriptor = fbLayer->descriptor(); armnn::ElementwiseUnaryDescriptor descriptor; - descriptor.m_Operation = ToUnaryOperation(fbDescriptor->operation()); + descriptor.m_Operation = ToElementwiseUnaryOperation(fbDescriptor->operation()); const std::string& layerName = GetLayerName(graph, layerIndex); IConnectableLayer* layer = m_Network->AddElementwiseUnaryLayer(descriptor, layerName.c_str()); @@ -2106,7 +2161,8 @@ void IDeserializer::DeserializerImpl::ParseMultiplication(GraphPtr graph, unsign CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddMultiplicationLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Mul); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); @@ -3023,7 +3079,8 @@ void IDeserializer::DeserializerImpl::ParseSubtraction(GraphPtr graph, unsigned CHECK_VALID_SIZE(outputs.size(), 1); auto layerName = GetLayerName(graph, layerIndex); - IConnectableLayer* layer = m_Network->AddSubtractionLayer(layerName.c_str()); + armnn::ElementwiseBinaryDescriptor descriptor(armnn::BinaryOperation::Sub); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(descriptor, layerName.c_str()); armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]); layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); diff --git a/src/armnnDeserializer/Deserializer.hpp b/src/armnnDeserializer/Deserializer.hpp index bd01a35431..b75c68db0c 100644 --- a/src/armnnDeserializer/Deserializer.hpp +++ b/src/armnnDeserializer/Deserializer.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -103,6 +103,7 @@ private: void ParseDequantize(GraphPtr graph, unsigned int layerIndex); void ParseDetectionPostProcess(GraphPtr graph, unsigned int layerIndex); void ParseDivision(GraphPtr graph, unsigned int layerIndex); + void ParseElementwiseBinary(GraphPtr graph, unsigned int layerIndex); void ParseElementwiseUnary(GraphPtr graph, unsigned int layerIndex); void ParseEqual(GraphPtr graph, unsigned int layerIndex); void ParseFill(GraphPtr graph, unsigned int layerIndex); diff --git a/src/armnnOnnxParser/OnnxParser.cpp b/src/armnnOnnxParser/OnnxParser.cpp index 936216fb00..26e2ceecc1 100644 --- a/src/armnnOnnxParser/OnnxParser.cpp +++ b/src/armnnOnnxParser/OnnxParser.cpp @@ -1558,7 +1558,7 @@ void OnnxParserImpl::ParseAdd(const onnx::NodeProto& node) } - IConnectableLayer* layer = m_Network->AddAdditionLayer(node.name().c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Add, node.name().c_str()); ARMNN_ASSERT(layer != nullptr); auto outputInfo = ComputeOutputInfo({ node.output(0) }, layer, diff --git a/src/armnnSerializer/ArmnnSchema.fbs b/src/armnnSerializer/ArmnnSchema.fbs index 2dbfd85b23..fb3bc018bb 100644 --- a/src/armnnSerializer/ArmnnSchema.fbs +++ b/src/armnnSerializer/ArmnnSchema.fbs @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -183,6 +183,7 @@ enum LayerType : uint { Pooling3d = 66, GatherNd = 67, BatchMatMul = 68, + ElementwiseBinary = 69, } // Base layer table to be used as part of other layers @@ -324,6 +325,24 @@ table DivisionLayer { base:LayerBase; } +enum BinaryOperation : byte { + Add = 0, + Div = 1, + Maximum = 2, + Minimum = 3, + Mul = 4, + Sub = 5 +} + +table ElementwiseBinaryDescriptor { + operation:BinaryOperation; +} + +table ElementwiseBinaryLayer { + base:LayerBase; + descriptor:ElementwiseBinaryDescriptor; +} + enum UnaryOperation : byte { Abs = 0, Rsqrt = 1, @@ -1094,6 +1113,7 @@ union Layer { Pooling3dLayer, GatherNdLayer, BatchMatMulLayer, + ElementwiseBinaryLayer, } table AnyLayer { diff --git a/src/armnnSerializer/Serializer.cpp b/src/armnnSerializer/Serializer.cpp index c9a3022b8d..5a095d9887 100644 --- a/src/armnnSerializer/Serializer.cpp +++ b/src/armnnSerializer/Serializer.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include "Serializer.hpp" @@ -533,6 +533,21 @@ void SerializerStrategy::SerializeDivisionLayer(const armnn::IConnectableLayer* CreateAnyLayer(fbDivisionLayer.o, serializer::Layer::Layer_DivisionLayer); } +void SerializerStrategy::SerializeElementwiseBinaryLayer(const armnn::IConnectableLayer* layer, + const armnn::ElementwiseBinaryDescriptor& descriptor, + const char* name) +{ + IgnoreUnused(name); + + auto fbBaseLayer = CreateLayerBase(layer, serializer::LayerType::LayerType_ElementwiseBinary); + auto fbDescriptor = serializer::CreateElementwiseBinaryDescriptor( + m_flatBufferBuilder, + GetFlatBufferBinaryOperation(descriptor.m_Operation)); + + auto fbLayer = serializer::CreateElementwiseBinaryLayer(m_flatBufferBuilder, fbBaseLayer, fbDescriptor); + CreateAnyLayer(fbLayer.o, serializer::Layer::Layer_ElementwiseBinaryLayer); +} + void SerializerStrategy::SerializeElementwiseUnaryLayer(const armnn::IConnectableLayer* layer, const armnn::ElementwiseUnaryDescriptor& descriptor, const char* name) @@ -2119,6 +2134,13 @@ void SerializerStrategy::ExecuteStrategy(const armnn::IConnectableLayer* layer, SerializeDivisionLayer(layer, name); break; } + case armnn::LayerType::ElementwiseBinary : + { + const armnn::ElementwiseBinaryDescriptor& layerDescriptor = + static_cast<const armnn::ElementwiseBinaryDescriptor&>(descriptor); + SerializeElementwiseBinaryLayer(layer, layerDescriptor, name); + break; + } case armnn::LayerType::ElementwiseUnary : { const armnn::ElementwiseUnaryDescriptor& layerDescriptor = diff --git a/src/armnnSerializer/Serializer.hpp b/src/armnnSerializer/Serializer.hpp index 60fed4f6df..4d5e806db1 100644 --- a/src/armnnSerializer/Serializer.hpp +++ b/src/armnnSerializer/Serializer.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -172,6 +172,10 @@ private: void SerializeDivisionLayer(const armnn::IConnectableLayer* layer, const char* name = nullptr); + void SerializeElementwiseBinaryLayer(const armnn::IConnectableLayer* layer, + const armnn::ElementwiseBinaryDescriptor& descriptor, + const char* name = nullptr); + void SerializeElementwiseUnaryLayer(const armnn::IConnectableLayer* layer, const armnn::ElementwiseUnaryDescriptor& descriptor, const char* name = nullptr); diff --git a/src/armnnSerializer/SerializerUtils.cpp b/src/armnnSerializer/SerializerUtils.cpp index 49ce7217dc..703f56ff5c 100644 --- a/src/armnnSerializer/SerializerUtils.cpp +++ b/src/armnnSerializer/SerializerUtils.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2019-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -107,9 +107,30 @@ armnnSerializer::DataLayout GetFlatBufferDataLayout(armnn::DataLayout dataLayout } } -armnnSerializer::UnaryOperation GetFlatBufferUnaryOperation(armnn::UnaryOperation comparisonOperation) +armnnSerializer::BinaryOperation GetFlatBufferBinaryOperation(armnn::BinaryOperation binaryOperation) { - switch (comparisonOperation) + switch (binaryOperation) + { + case armnn::BinaryOperation::Add: + return armnnSerializer::BinaryOperation::BinaryOperation_Add; + case armnn::BinaryOperation::Div: + return armnnSerializer::BinaryOperation::BinaryOperation_Div; + case armnn::BinaryOperation::Maximum: + return armnnSerializer::BinaryOperation::BinaryOperation_Maximum; + case armnn::BinaryOperation::Minimum: + return armnnSerializer::BinaryOperation::BinaryOperation_Minimum; + case armnn::BinaryOperation::Mul: + return armnnSerializer::BinaryOperation::BinaryOperation_Mul; + case armnn::BinaryOperation::Sub: + return armnnSerializer::BinaryOperation::BinaryOperation_Sub; + default: + throw armnn::InvalidArgumentException("Elementwise Binary operation unknown"); + } +} + +armnnSerializer::UnaryOperation GetFlatBufferUnaryOperation(armnn::UnaryOperation unaryOperation) +{ + switch (unaryOperation) { case armnn::UnaryOperation::Abs: return armnnSerializer::UnaryOperation::UnaryOperation_Abs; @@ -128,7 +149,7 @@ armnnSerializer::UnaryOperation GetFlatBufferUnaryOperation(armnn::UnaryOperatio case armnn::UnaryOperation::Sin: return armnnSerializer::UnaryOperation::UnaryOperation_Sin; default: - throw armnn::InvalidArgumentException("Unary operation unknown"); + throw armnn::InvalidArgumentException("Elementwise Unary operation unknown"); } } diff --git a/src/armnnSerializer/SerializerUtils.hpp b/src/armnnSerializer/SerializerUtils.hpp index 07cdc2a491..628385eea9 100644 --- a/src/armnnSerializer/SerializerUtils.hpp +++ b/src/armnnSerializer/SerializerUtils.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017,2019-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -18,6 +18,8 @@ armnnSerializer::DataType GetFlatBufferDataType(armnn::DataType dataType); armnnSerializer::DataLayout GetFlatBufferDataLayout(armnn::DataLayout dataLayout); +armnnSerializer::BinaryOperation GetFlatBufferBinaryOperation(armnn::BinaryOperation binaryOperation); + armnnSerializer::UnaryOperation GetFlatBufferUnaryOperation(armnn::UnaryOperation unaryOperation); armnnSerializer::PoolingAlgorithm GetFlatBufferPoolingAlgorithm(armnn::PoolingAlgorithm poolingAlgorithm); diff --git a/src/armnnSerializer/test/SerializerTests.cpp b/src/armnnSerializer/test/SerializerTests.cpp index 3573a8195c..6ddc971f36 100644 --- a/src/armnnSerializer/test/SerializerTests.cpp +++ b/src/armnnSerializer/test/SerializerTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2020-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -331,6 +331,7 @@ TEST_CASE("SerializeConstant") case armnn::LayerType::Input: break; case armnn::LayerType::Output: break; case armnn::LayerType::Addition: break; + case armnn::LayerType::ElementwiseBinary: break; default: { this->VerifyNameAndConnections(layer, name); @@ -972,6 +973,47 @@ TEST_CASE("SerializeDeserializeComparisonEqual") deserializedNetwork->ExecuteStrategy(verifier); } +void SerializeElementwiseBinaryTest(armnn::BinaryOperation binaryOperation) +{ + auto layerName = GetBinaryOperationAsCString(binaryOperation); + const armnn::TensorInfo tensorInfo({ 1, 5, 2, 3 }, armnn::DataType::Float32); + armnn::ElementwiseBinaryDescriptor descriptor(binaryOperation); + + armnn::INetworkPtr network = armnn::INetwork::Create(); + armnn::IConnectableLayer* const inputLayer0 = network->AddInputLayer(0); + armnn::IConnectableLayer* const inputLayer1 = network->AddInputLayer(1); + armnn::IConnectableLayer* const elementwiseBinaryLayer = network->AddElementwiseBinaryLayer(descriptor, + layerName); + armnn::IConnectableLayer* const outputLayer = network->AddOutputLayer(0); + + inputLayer0->GetOutputSlot(0).Connect(elementwiseBinaryLayer->GetInputSlot(0)); + inputLayer1->GetOutputSlot(0).Connect(elementwiseBinaryLayer->GetInputSlot(1)); + elementwiseBinaryLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + + inputLayer0->GetOutputSlot(0).SetTensorInfo(tensorInfo); + inputLayer1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + elementwiseBinaryLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + std::string serializedNetwork = SerializeNetwork(*network); + armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(serializedNetwork); + CHECK(deserializedNetwork); + + LayerVerifierBaseWithDescriptor<armnn::ElementwiseBinaryDescriptor> + verifier(layerName, { tensorInfo, tensorInfo }, { tensorInfo }, descriptor); + deserializedNetwork->ExecuteStrategy(verifier); +} + +TEST_CASE("SerializeElementwiseBinary") +{ + using op = armnn::BinaryOperation; + std::initializer_list<op> allBinaryOperations = {op::Add, op::Div, op::Maximum, op::Minimum, op::Mul, op::Sub}; + + for (auto binaryOperation : allBinaryOperations) + { + SerializeElementwiseBinaryTest(binaryOperation); + } +} + void SerializeElementwiseUnaryTest(armnn::UnaryOperation unaryOperation) { auto layerName = GetUnaryOperationAsCString(unaryOperation); @@ -2883,6 +2925,7 @@ TEST_CASE("SerializeDeserializeNonLinearNetwork") CompareConstTensor(constants.at(0), m_LayerInput); break; } + case armnn::LayerType::ElementwiseBinary: break; default: { throw armnn::Exception("Unexpected layer type in test model"); diff --git a/src/armnnTestUtils/CreateWorkload.hpp b/src/armnnTestUtils/CreateWorkload.hpp index 0846d21388..691adbff9d 100644 --- a/src/armnnTestUtils/CreateWorkload.hpp +++ b/src/armnnTestUtils/CreateWorkload.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017,2021-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -125,7 +125,41 @@ std::unique_ptr<WorkloadType> CreateElementwiseWorkloadTest(armnn::IWorkloadFact // Makes the workload and checks it. auto workload = MakeAndCheckWorkload<WorkloadType>(*layer, factory); - DescriptorType queueDescriptor = workload->GetData(); + auto queueDescriptor = workload->GetData(); + CHECK(queueDescriptor.m_Inputs.size() == 2); + CHECK(queueDescriptor.m_Outputs.size() == 1); + + // Returns so we can do extra, backend-specific tests. + return workload; +} + +template <typename WorkloadType, armnn::DataType DataType> +std::unique_ptr<WorkloadType> CreateElementwiseBinaryWorkloadTest(armnn::IWorkloadFactory & factory, + armnn::Graph & graph, + armnn::BinaryOperation binaryOperation) +{ + // Creates the layer we're testing. + ElementwiseBinaryDescriptor descriptor(binaryOperation); + //ElementwiseBinaryDescriptor descriptor = ElementwiseBinaryDescriptor(binaryOperation); + + Layer* const layer = graph.AddLayer<ElementwiseBinaryLayer>(descriptor, "layer"); + + // Creates extra layers. + Layer* const input1 = graph.AddLayer<InputLayer>(1, "input1"); + Layer* const input2 = graph.AddLayer<InputLayer>(2, "input2"); + Layer* const output = graph.AddLayer<OutputLayer>(0, "output"); + + // Connects up. + armnn::TensorInfo tensorInfo({2, 3}, DataType); + Connect(input1, layer, tensorInfo, 0, 0); + Connect(input2, layer, tensorInfo, 0, 1); + Connect(layer, output, tensorInfo); + CreateTensorHandles(graph, factory); + + // Makes the workload and checks it. + auto workload = MakeAndCheckWorkload<WorkloadType>(*layer, factory); + + auto queueDescriptor = workload->GetData(); CHECK(queueDescriptor.m_Inputs.size() == 2); CHECK(queueDescriptor.m_Outputs.size() == 1); @@ -191,6 +225,7 @@ std::unique_ptr<WorkloadType> CreateSubtractionWithBlobWorkloadTest(armnn::IWork return workload; } + template<typename WorkloadType, typename DescriptorType, armnn::DataType DataType> diff --git a/src/armnnTestUtils/MockBackend.cpp b/src/armnnTestUtils/MockBackend.cpp index 5dfe9a3b8b..7441d0c487 100644 --- a/src/armnnTestUtils/MockBackend.cpp +++ b/src/armnnTestUtils/MockBackend.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -69,6 +69,7 @@ bool IsLayerSupported(const armnn::Layer* layer) case armnn::LayerType::Constant: case armnn::LayerType::Addition: case armnn::LayerType::Convolution2d: + case armnn::LayerType::ElementwiseBinary: // Layer supported return true; default: diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp index b7e2762ade..dc5afca30e 100644 --- a/src/armnnTfLiteParser/TfLiteParser.cpp +++ b/src/armnnTfLiteParser/TfLiteParser.cpp @@ -1838,7 +1838,7 @@ void TfLiteParserImpl::ParseMaximum(size_t subgraphIndex, size_t operatorIndex) TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); CheckMatchingQuantization(inputTensorInfo, input1TensorInfo, layerName, "Input 0", "Input 1"); - IConnectableLayer* layer = m_Network->AddMaximumLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Maximum, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -1868,7 +1868,7 @@ void TfLiteParserImpl::ParseMinimum(size_t subgraphIndex, size_t operatorIndex) TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); CheckMatchingQuantization(inputTensorInfo, input1TensorInfo, layerName, "Input 0", "Input 1"); - IConnectableLayer* layer = m_Network->AddMinimumLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Minimum, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -2384,7 +2384,7 @@ void TfLiteParserImpl::ParseSub(size_t subgraphIndex, size_t operatorIndex) armnn::TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); auto layerName = fmt::format("Sub:{}:{}", subgraphIndex, operatorIndex); - IConnectableLayer* layer = m_Network->AddSubtractionLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Sub, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -2416,7 +2416,7 @@ void TfLiteParserImpl::ParseDiv(size_t subgraphIndex, size_t operatorIndex) armnn::TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); auto layerName = fmt::format("Div:{}:{}", subgraphIndex, operatorIndex); - IConnectableLayer* layer = m_Network->AddDivisionLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Div, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -2444,7 +2444,7 @@ void TfLiteParserImpl::ParseFloorDiv(size_t subgraphIndex, size_t operatorIndex) armnn::TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); auto layerName = fmt::format("Div:{}:{}", subgraphIndex, operatorIndex); - IConnectableLayer* layer = m_Network->AddDivisionLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Div, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -2475,7 +2475,7 @@ void TfLiteParserImpl::ParseAdd(size_t subgraphIndex, size_t operatorIndex) armnn::TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); auto layerName = fmt::format("Add:{}:{}", subgraphIndex, operatorIndex); - IConnectableLayer* layer = m_Network->AddAdditionLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Add, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); @@ -2506,7 +2506,7 @@ void TfLiteParserImpl::ParseMul(size_t subgraphIndex, size_t operatorIndex) armnn::TensorInfo input1TensorInfo = InputTensorInfo(subgraphIndex, operatorIndex, 1); auto layerName = fmt::format("Mul:{}:{}", subgraphIndex, operatorIndex); - IConnectableLayer* layer = m_Network->AddMultiplicationLayer(layerName.c_str()); + IConnectableLayer* layer = m_Network->AddElementwiseBinaryLayer(BinaryOperation::Mul, layerName.c_str()); ARMNN_ASSERT(layer != nullptr); TensorInfo outputTensorInfo = OutputTensorInfoFromInputs(subgraphIndex, operatorIndex, layer, 0, {0, 1}); diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp index c9d6c71f18..599d3538eb 100644 --- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp +++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2020,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -147,6 +147,27 @@ LayerType* FuseMultiplicationLayer(OptimizationViews& optimizationViews, } template<typename LayerType> +LayerType* FuseElementwiseBinaryLayer(OptimizationViews& optimizationViews, + LayerType* baseLayer, + ActivationLayer* activationLayer, + ActivationDescriptor& activationDesc, + BinaryOperation operation, + std::string name) +{ + IConnectableLayer* replacement = optimizationViews.GetINetwork()->AddElementwiseBinaryLayer(operation, + name.c_str()); + LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement); + + FuseLayer(optimizationViews, + baseLayer, + replacementLayer, + activationLayer, + activationDesc); + + return replacementLayer; +} + +template<typename LayerType> LayerType* FuseBatchNormalizationLayer(OptimizationViews& optimizationViews, LayerType* baseLayer, ActivationLayer* activationLayer, diff --git a/src/backends/backendsCommon/CMakeLists.txt b/src/backends/backendsCommon/CMakeLists.txt index 8d7e114fa5..28ff205d6e 100644 --- a/src/backends/backendsCommon/CMakeLists.txt +++ b/src/backends/backendsCommon/CMakeLists.txt @@ -1,5 +1,5 @@ # -# Copyright © 2017, 2023 Arm Ltd and Contributors. All rights reserved. +# Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT # diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp index 62dfc6a38b..6a5963ddcb 100644 --- a/src/backends/backendsCommon/WorkloadData.cpp +++ b/src/backends/backendsCommon/WorkloadData.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -35,11 +35,8 @@ DataType GetBiasDataType(DataType inputDataType) case DataType::Float32: return DataType::Float32; case DataType::QAsymmS8: - return DataType::Signed32; case DataType::QAsymmU8: - return DataType::Signed32; case DataType::QSymmS8: - return DataType::Signed32; case DataType::QSymmS16: return DataType::Signed32; default: @@ -3668,6 +3665,35 @@ void ComparisonQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const } } +void ElementwiseBinaryQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const +{ + const std::string descriptorName{"ElementwiseBinaryQueueDescriptor"}; + + ValidateNumInputs(workloadInfo, descriptorName, 2); + ValidateNumOutputs(workloadInfo, descriptorName, 1); + + const TensorInfo& inputTensorInfo0 = workloadInfo.m_InputTensorInfos[0]; + const TensorInfo& inputTensorInfo1 = workloadInfo.m_InputTensorInfos[1]; + const TensorInfo& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0]; + + std::vector<DataType> supportedTypes = + { + DataType::BFloat16, + DataType::Float16, + DataType::Float32, + DataType::QAsymmS8, + DataType::QAsymmU8, + DataType::QSymmS16, + DataType::Signed32 + }; + + ValidateDataTypes(inputTensorInfo0, supportedTypes, descriptorName); + ValidateDataTypes(inputTensorInfo1, supportedTypes, descriptorName); + + ValidateTensorDataTypesMatch(inputTensorInfo0, outputTensorInfo, descriptorName, "input", "output"); + ValidateTensorDataTypesMatch(inputTensorInfo1, outputTensorInfo, descriptorName, "input", "output"); +} + void ElementwiseUnaryQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const { const std::string descriptorName{"ElementwiseUnaryQueueDescriptor"}; diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp index 1283f67660..51bc3e60cb 100644 --- a/src/backends/backendsCommon/WorkloadFactory.cpp +++ b/src/backends/backendsCommon/WorkloadFactory.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -15,7 +15,6 @@ #include <armnn/utility/TransformIterator.hpp> #include <armnn/backends/WorkloadFactory.hpp> -#include <armnn/backends/TensorHandle.hpp> #include <sstream> @@ -91,7 +90,8 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, auto backendFactory = backendRegistry.GetFactory(backendId); auto backendObject = backendFactory(); - auto layerSupportObject = LayerSupportHandle(backendObject->GetLayerSupport(modelOptions), backendId); + auto layerSupport = backendObject->GetLayerSupport(modelOptions); + auto layerSupportObject = LayerSupportHandle(layerSupport, backendId); switch(layer.GetType()) { @@ -109,6 +109,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Addition: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -117,6 +118,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::ArgMinMax: @@ -392,6 +394,24 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, reason); break; } + case LayerType::ElementwiseBinary: + { + auto cLayer = PolymorphicDowncast<const ElementwiseBinaryLayer*>(&layer); + + const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); + const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); + const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); + std::vector<TensorInfo> infos = { OverrideDataType(input0, dataType), + OverrideDataType(input1, dataType), + OverrideDataType(output, dataType) }; + result = layerSupport->IsLayerSupported(LayerType::ElementwiseBinary, + infos, + cLayer->GetParameters(), + EmptyOptional(), + EmptyOptional(), + reason); + break; + } case LayerType::ElementwiseUnary: { auto cLayer = PolymorphicDowncast<const ElementwiseUnaryLayer*>(&layer); @@ -740,6 +760,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Maximum: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -748,6 +769,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::MemCopy: @@ -814,6 +836,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Multiplication: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -822,6 +845,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::Normalization: @@ -1052,6 +1076,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Division: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -1060,6 +1085,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::Rank: @@ -1254,6 +1280,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Subtraction: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -1262,6 +1289,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::Switch: @@ -1291,6 +1319,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, } case LayerType::Minimum: { + ARMNN_NO_DEPRECATE_WARN_BEGIN const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo(); const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); @@ -1298,6 +1327,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, OverrideDataType(input1, dataType), OverrideDataType(output, dataType), reason); + ARMNN_NO_DEPRECATE_WARN_END break; } case LayerType::Prelu: @@ -1670,6 +1700,11 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateWorkload(LayerType type, auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor); return CreateDivision(*divisionQueueDescriptor, info); } + case LayerType::ElementwiseBinary: + { + auto queueDescriptor = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor); + return CreateWorkload(LayerType::ElementwiseBinary, *queueDescriptor, info); + } case LayerType::ElementwiseUnary: { auto elementwiseUnaryQueueDescriptor diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk index 3545331c8f..986d2530c2 100644 --- a/src/backends/backendsCommon/common.mk +++ b/src/backends/backendsCommon/common.mk @@ -1,5 +1,5 @@ # -# Copyright © 2017 ARM Ltd. All rights reserved. +# Copyright © 2017-2023 ARM Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT # diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt index 509157a378..77335d550a 100644 --- a/src/backends/backendsCommon/test/CMakeLists.txt +++ b/src/backends/backendsCommon/test/CMakeLists.txt @@ -24,6 +24,7 @@ list(APPEND armnnBackendsCommonUnitTests_sources DetectionPostProcessEndToEndTestImpl.hpp DynamicBackendTests.cpp DynamicBackendTests.hpp + ElementwiseBinaryEndToEndTestImpl.hpp ElementwiseUnaryEndToEndTestImpl.hpp EndToEndTestImpl.hpp FillEndToEndTestImpl.hpp diff --git a/src/backends/backendsCommon/test/ElementwiseBinaryEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/ElementwiseBinaryEndToEndTestImpl.hpp new file mode 100644 index 0000000000..6546a6ae9e --- /dev/null +++ b/src/backends/backendsCommon/test/ElementwiseBinaryEndToEndTestImpl.hpp @@ -0,0 +1,107 @@ +// +// Copyright © 2023 Arm Ltd and contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "CommonTestUtils.hpp" + +#include <ResolveType.hpp> + +#include <armnn/INetwork.hpp> +#include <armnn/utility/NumericCast.hpp> + +#include <doctest/doctest.h> + +#include <vector> + +namespace +{ + +template<armnn::DataType ArmnnTypeInput> +INetworkPtr CreateElementwiseBinaryNetwork(const TensorShape& input1Shape, + const TensorShape& input2Shape, + const TensorShape& outputShape, + BinaryOperation operation, + const float qScale = 1.0f, + const int32_t qOffset = 0) +{ + using namespace armnn; + + INetworkPtr net(INetwork::Create()); + + TensorInfo input1TensorInfo(input1Shape, ArmnnTypeInput, qScale, qOffset, true); + TensorInfo input2TensorInfo(input2Shape, ArmnnTypeInput, qScale, qOffset, true); + TensorInfo outputTensorInfo(outputShape, ArmnnTypeInput, qScale, qOffset); + + IConnectableLayer* input1 = net->AddInputLayer(armnn::numeric_cast<LayerBindingId>(0)); + IConnectableLayer* input2 = net->AddInputLayer(armnn::numeric_cast<LayerBindingId>(1)); + IConnectableLayer* elementwiseBinaryLayer = net->AddElementwiseBinaryLayer(operation, "elementwiseUnary"); + IConnectableLayer* output = net->AddOutputLayer(0, "output"); + + Connect(input1, elementwiseBinaryLayer, input1TensorInfo, 0, 0); + Connect(input2, elementwiseBinaryLayer, input2TensorInfo, 0, 1); + Connect(elementwiseBinaryLayer, output, outputTensorInfo, 0, 0); + + return net; +} + +template<armnn::DataType ArmnnInType, + typename TInput = armnn::ResolveType<ArmnnInType>> +void ElementwiseBinarySimpleEndToEnd(const std::vector<BackendId>& backends, + BinaryOperation operation) +{ + using namespace armnn; + + const float qScale = IsQuantizedType<TInput>() ? 0.25f : 1.0f; + const int32_t qOffset = IsQuantizedType<TInput>() ? 50 : 0; + + const TensorShape& input1Shape = { 2, 2, 2, 2 }; + const TensorShape& input2Shape = { 1 }; + const TensorShape& outputShape = { 2, 2, 2, 2 }; + + // Builds up the structure of the network + INetworkPtr net = CreateElementwiseBinaryNetwork<ArmnnInType>(input1Shape, input2Shape, outputShape, + operation, qScale, qOffset); + + CHECK(net); + + const std::vector<float> input1({ 1, -1, 1, 1, 5, -5, 5, 5, -3, 3, 3, 3, 4, 4, -4, 4 }); + + const std::vector<float> input2({ 2 }); + std::vector<float> expectedOutput; + switch (operation) { + case armnn::BinaryOperation::Add: + expectedOutput = { 3, 1, 3, 3, 7, -3, 7, 7, -1, 5, 5, 5, 6, 6, -2, 6 }; + break; + case armnn::BinaryOperation::Div: + expectedOutput = {0.5f, -0.5f, 0.5f, 0.5f, 2.5f, -2.5f, 2.5f, 2.5f, -1.5f, 1.5f, 1.5f, 1.5f, 2, 2, -2, 2}; + break; + case armnn::BinaryOperation::Maximum: + expectedOutput = { 2, 2, 2, 2, 5, 2, 5, 5, 2, 3, 3, 3, 4, 4, 2, 4 }; + break; + case armnn::BinaryOperation::Minimum: + expectedOutput = { 1, -1, 1, 1, 2, -5, 2, 2, -3, 2, 2, 2, 2, 2, -4, 2 }; + break; + case armnn::BinaryOperation::Mul: + expectedOutput = { 2, -2, 2, 2, 10, -10, 10, 10, -6, 6, 6, 6, 8, 8, -8, 8 }; + break; + case armnn::BinaryOperation::Sub: + expectedOutput = { -1, -3, -1, -1, 3, -7, 3, 3, -5, 1, 1, 1, 2, 2, -6, 2 }; + break; + default: + throw("Invalid Elementwise Binary operation"); + } + const std::vector<float> expectedOutput_const = expectedOutput; + // quantize data + std::vector<TInput> qInput1Data = armnnUtils::QuantizedVector<TInput>(input1, qScale, qOffset); + std::vector<TInput> qInput2Data = armnnUtils::QuantizedVector<TInput>(input2, qScale, qOffset); + std::vector<TInput> qExpectedOutput = armnnUtils::QuantizedVector<TInput>(expectedOutput_const, qScale, qOffset); + + std::map<int, std::vector<TInput>> inputTensorData = {{ 0, qInput1Data }, { 1, qInput2Data }}; + std::map<int, std::vector<TInput>> expectedOutputData = {{ 0, qExpectedOutput }}; + + EndToEndLayerTestImpl<ArmnnInType, ArmnnInType>(std::move(net), inputTensorData, expectedOutputData, backends); +} + +} // anonymous namespace diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp index fb7a0271d4..5b95d3cd92 100644 --- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp +++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -664,6 +664,8 @@ DECLARE_LAYER_POLICY_1_PARAM(Dequantize) DECLARE_LAYER_POLICY_2_PARAM(DetectionPostProcess) +DECLARE_LAYER_POLICY_2_PARAM(ElementwiseBinary) + DECLARE_LAYER_POLICY_2_PARAM(ElementwiseUnary) DECLARE_LAYER_POLICY_2_PARAM(FakeQuantization) diff --git a/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp b/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp index cd865def71..5e619df8dd 100644 --- a/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp +++ b/src/backends/backendsCommon/test/OptimizedNetworkTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -21,7 +21,7 @@ TEST_CASE("SerializeToDot") //Defines layers. auto input = net->AddInputLayer(0); - auto add = net->AddAdditionLayer(); + auto add = net->AddElementwiseBinaryLayer(armnn::BinaryOperation::Add); auto output = net->AddOutputLayer(0); // Connects layers. @@ -54,7 +54,7 @@ TEST_CASE("SerializeToDot") " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n" " " << inputId << " [label=\"{Input|Guid : " << inputId << "\\lLayerType : Input\\l" "BackendID : CpuRef\\l}\"];\n" - " " << addId << " [label=\"{Addition|Guid : " << addId << "\\lLayerType : Addition\\l" + " " << addId << " [label=\"{ElementwiseBinary|Guid : " << addId << "\\lLayerType : ElementwiseBinary\\l" "BackendID : CpuRef\\l}\"];\n" " " << outputId << " [label=\"{Output|Guid : " << outputId << "\\lLayerType : Output\\l" "BackendID : CpuRef\\l}\"];\n" @@ -187,7 +187,7 @@ TEST_CASE("OptimizeValidateWorkloadsUndefinedComputeDevice") layer->GetOutputSlot(0).SetTensorInfo(desc); armnn::IConnectableLayer* prevLayer = layer; - layer = net->AddMultiplicationLayer("ml"); + layer = net->AddElementwiseBinaryLayer(armnn::BinaryOperation::Mul, "ml"); prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); @@ -258,7 +258,7 @@ TEST_CASE("OptimizeValidateWorkloadsUndefinedComputeDeviceWithFallback") layer->GetOutputSlot(0).SetTensorInfo(desc); armnn::IConnectableLayer* prevLayer = layer; - layer = net->AddMultiplicationLayer("ml"); + layer = net->AddElementwiseBinaryLayer(armnn::BinaryOperation::Mul, "ml"); prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); diff --git a/src/backends/backendsCommon/test/mockBackend/MockImportLayerSupport.hpp b/src/backends/backendsCommon/test/mockBackend/MockImportLayerSupport.hpp index 380ce4a3f5..da4b7ab7d0 100644 --- a/src/backends/backendsCommon/test/mockBackend/MockImportLayerSupport.hpp +++ b/src/backends/backendsCommon/test/mockBackend/MockImportLayerSupport.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2021,2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once @@ -16,7 +16,7 @@ class MockImportLayerSupport : public LayerSupportBase public: bool IsLayerSupported(const LayerType& type, const std::vector<TensorInfo>& infos, - const BaseDescriptor& /*descriptor*/, + const BaseDescriptor& descriptor, const Optional<LstmInputParamsInfo>& /*lstmParamsInfo*/, const Optional<QuantizedLstmInputParamsInfo>& /*quantizedLstmParamsInfo*/, Optional<std::string&> reasonIfUnsupported) const override @@ -25,6 +25,11 @@ public: { case LayerType::Addition: return IsAdditionSupported(infos[0], infos[1], infos[2], reasonIfUnsupported); + case LayerType::ElementwiseBinary: + { + auto elementwiseDesc = *(PolymorphicDowncast<const ElementwiseBinaryDescriptor*>(&descriptor)); + return (elementwiseDesc.m_Operation == BinaryOperation::Add); + } case LayerType::Input: return IsInputSupported(infos[0], reasonIfUnsupported); case LayerType::Output: diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp index d2e8fbfe32..a10b6fbb43 100644 --- a/src/backends/cl/ClBackend.cpp +++ b/src/backends/cl/ClBackend.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2022 Arm Ltd. All rights reserved. +// Copyright © 2017-2023 Arm Ltd. All rights reserved. // SPDX-License-Identifier: MIT // @@ -311,7 +311,8 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph, if ((base.GetType() == LayerType::DepthwiseConvolution2d || base.GetType() == LayerType::Convolution2d || base.GetType() == LayerType::BatchNormalization || base.GetType() == LayerType::FullyConnected || base.GetType() == LayerType::Addition || base.GetType() == LayerType::Multiplication - || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division) + || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division + || base.GetType() == LayerType::ElementwiseBinary) && (base.GetAdditionalInformation<ActivationDescriptor>() == nullptr)) { for (auto output = base.BeginOutputSlots(); output != base.EndOutputSlots(); ++output) @@ -542,6 +543,90 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph, untouched.erase(activationLayer->GetGuid()); } } + else if (base.GetType() == LayerType::ElementwiseBinary) + { + ElementwiseBinaryLayer* baseLayer = PolymorphicDowncast<ElementwiseBinaryLayer*>(&base); + + if (baseLayer->GetParameters().m_Operation == BinaryOperation::Add) + { + arm_compute::Status status = ClAdditionValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Add, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Div) + { + arm_compute::Status status = ClDivisionWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Div, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Mul) + { + arm_compute::Status status = ClMultiplicationWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Mul, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Sub) + { + arm_compute::Status status = ClSubtractionValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Sub, + name); + } + } + // No fusion available for other BinaryOperations + } } } } diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp index cb2d756037..89bcf9bc01 100644 --- a/src/backends/cl/ClLayerSupport.cpp +++ b/src/backends/cl/ClLayerSupport.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -347,6 +347,56 @@ bool ClLayerSupport::IsLayerSupported(const LayerType& type, return IsDequantizeSupported(infos[0], infos[1], reasonIfUnsupported); case LayerType::Division: return IsDivisionSupported(infos[0], infos[1], infos[2], reasonIfUnsupported); + case LayerType::ElementwiseBinary: + { + auto desc = *(PolymorphicDowncast<const ElementwiseBinaryDescriptor *>(&descriptor)); + + switch (desc.m_Operation) + { + case BinaryOperation::Add: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClAdditionValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Div: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClDivisionWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Minimum: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClMinimumWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2]); + case BinaryOperation::Maximum: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClMaximumWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2]); + case BinaryOperation::Mul: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClMultiplicationWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Sub: + FORWARD_WORKLOAD_VALIDATE_FUNC(ClSubtractionValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + default: + return false; + } + } case LayerType::ElementwiseUnary: return IsElementwiseUnarySupported(infos[0], infos[1], diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp index 6bf510a2ef..022867710c 100644 --- a/src/backends/cl/ClWorkloadFactory.cpp +++ b/src/backends/cl/ClWorkloadFactory.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include "ClWorkloadFactory.hpp" @@ -405,6 +405,75 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateWorkload(LayerType type, auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor); return std::make_unique<ClDivisionWorkload>(*divisionQueueDescriptor, info, m_CLCompileContext); } + case LayerType::ElementwiseBinary : + { + auto elementwiseBinaryQueueDescriptor + = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor); + + switch (elementwiseBinaryQueueDescriptor->m_Parameters.m_Operation) + { + case BinaryOperation::Add: + { + AdditionQueueDescriptor additionQueueDescriptor; + additionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + additionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + additionQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClAdditionWorkload>(additionQueueDescriptor, info, m_CLCompileContext); + } + case BinaryOperation::Div: + { + DivisionQueueDescriptor divisionQueueDescriptor; + divisionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + divisionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + divisionQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClDivisionWorkload>(divisionQueueDescriptor, info, m_CLCompileContext); + } + case BinaryOperation::Maximum: + { + MaximumQueueDescriptor maximumQueueDescriptor; + maximumQueueDescriptor.m_Inputs = descriptor.m_Inputs; + maximumQueueDescriptor.m_Outputs = descriptor.m_Outputs; + maximumQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClMaximumWorkload>(maximumQueueDescriptor, info, m_CLCompileContext); + } + case BinaryOperation::Minimum: + { + MinimumQueueDescriptor minimumQueueDescriptor; + minimumQueueDescriptor.m_Inputs = descriptor.m_Inputs; + minimumQueueDescriptor.m_Outputs = descriptor.m_Outputs; + minimumQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClMinimumWorkload>(minimumQueueDescriptor, info, m_CLCompileContext); + } + case BinaryOperation::Mul: + { + MultiplicationQueueDescriptor multiplicationQueueDescriptor; + multiplicationQueueDescriptor.m_Inputs = descriptor.m_Inputs; + multiplicationQueueDescriptor.m_Outputs = descriptor.m_Outputs; + multiplicationQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClMultiplicationWorkload>(multiplicationQueueDescriptor, + info, + m_CLCompileContext); + } + case BinaryOperation::Sub: + { + SubtractionQueueDescriptor subtractionQueueDescriptor; + subtractionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + subtractionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + subtractionQueueDescriptor.m_AdditionalInfoObject = + elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject; + return std::make_unique<ClSubtractionWorkload>(subtractionQueueDescriptor, + info, + m_CLCompileContext); + } + default: + return nullptr; + } + } case LayerType::ElementwiseUnary : { auto elementwiseUnaryQueueDescriptor diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp index adea733582..c49ca23266 100644 --- a/src/backends/cl/test/ClCreateWorkloadTests.cpp +++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -66,19 +66,17 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "CreateActivationFloat16Workload") } template <typename WorkloadType, - typename DescriptorType, - typename LayerType, armnn::DataType DataType> -static void ClCreateElementwiseWorkloadTest() +static void ClCreateElementwiseWorkloadTest(BinaryOperation binaryOperator) { Graph graph; ClWorkloadFactory factory = ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager()); - auto workload = CreateElementwiseWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph); + auto workload = CreateElementwiseBinaryWorkloadTest<WorkloadType, DataType>(factory, graph, binaryOperator); // Checks that inputs/outputs are as we expect them (see definition of CreateElementwiseWorkloadTest). - DescriptorType queueDescriptor = workload->GetData(); + auto queueDescriptor = workload->GetData(); auto inputHandle1 = PolymorphicDowncast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]); auto inputHandle2 = PolymorphicDowncast<IClTensorHandle*>(queueDescriptor.m_Inputs[1]); auto outputHandle = PolymorphicDowncast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]); @@ -93,73 +91,55 @@ static void ClCreateElementwiseWorkloadTest() TEST_CASE_FIXTURE(ClContextControlFixture, "CreateAdditionFloatWorkload") { ClCreateElementwiseWorkloadTest<ClAdditionWorkload, - AdditionQueueDescriptor, - AdditionLayer, - armnn::DataType::Float32>(); + armnn::DataType::Float32>(BinaryOperation::Add); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateAdditionFloat16Workload") { ClCreateElementwiseWorkloadTest<ClAdditionWorkload, - AdditionQueueDescriptor, - AdditionLayer, - armnn::DataType::Float16>(); + armnn::DataType::Float16>(BinaryOperation::Add); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateSubtractionFloatWorkload") { ClCreateElementwiseWorkloadTest<ClSubtractionWorkload, - SubtractionQueueDescriptor, - SubtractionLayer, - armnn::DataType::Float32>(); + armnn::DataType::Float32>(BinaryOperation::Sub); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateSubtractionFloat16Workload") { ClCreateElementwiseWorkloadTest<ClSubtractionWorkload, - SubtractionQueueDescriptor, - SubtractionLayer, - armnn::DataType::Float16>(); + armnn::DataType::Float16>(BinaryOperation::Sub); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateMultiplicationFloatWorkloadTest") { ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload, - MultiplicationQueueDescriptor, - MultiplicationLayer, - armnn::DataType::Float32>(); + armnn::DataType::Float32>(BinaryOperation::Mul); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateMultiplicationFloat16WorkloadTest") { ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload, - MultiplicationQueueDescriptor, - MultiplicationLayer, - armnn::DataType::Float16>(); + armnn::DataType::Float16>(BinaryOperation::Mul); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateMultiplicationUint8WorkloadTest") { ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload, - MultiplicationQueueDescriptor, - MultiplicationLayer, - armnn::DataType::QAsymmU8>(); + armnn::DataType::QAsymmU8>(BinaryOperation::Mul); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateDivisionFloatWorkloadTest") { ClCreateElementwiseWorkloadTest<ClDivisionWorkload, - DivisionQueueDescriptor, - DivisionLayer, - armnn::DataType::Float32>(); + armnn::DataType::Float32>(BinaryOperation::Div); } TEST_CASE_FIXTURE(ClContextControlFixture, "CreateDivisionFloat16WorkloadTest") { ClCreateElementwiseWorkloadTest<ClDivisionWorkload, - DivisionQueueDescriptor, - DivisionLayer, - armnn::DataType::Float16>(); + armnn::DataType::Float16>(BinaryOperation::Div); } template <typename WorkloadType, diff --git a/src/backends/cl/test/ClFallbackTests.cpp b/src/backends/cl/test/ClFallbackTests.cpp index 51a983a681..9443116c92 100644 --- a/src/backends/cl/test/ClFallbackTests.cpp +++ b/src/backends/cl/test/ClFallbackTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -24,8 +24,8 @@ TEST_CASE("ClImportEnabledFallbackToNeon") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -172,8 +172,8 @@ TEST_CASE("ClImportDisabledFallbackToNeon") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -301,8 +301,8 @@ TEST_CASE("ClImportEnabledFallbackSubgraphToNeon") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); @@ -460,8 +460,8 @@ TEST_CASE("ClImportDisableFallbackSubgraphToNeon") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); diff --git a/src/backends/cl/test/Fp16SupportTest.cpp b/src/backends/cl/test/Fp16SupportTest.cpp index da6ea10926..28ae4795ab 100644 --- a/src/backends/cl/test/Fp16SupportTest.cpp +++ b/src/backends/cl/test/Fp16SupportTest.cpp @@ -28,7 +28,7 @@ TEST_CASE("Fp16DataTypeSupport") Layer* const inputLayer1 = graph.AddLayer<InputLayer>(1, "input1"); Layer* const inputLayer2 = graph.AddLayer<InputLayer>(2, "input2"); - Layer* const additionLayer = graph.AddLayer<AdditionLayer>("addition"); + Layer* const additionLayer = graph.AddLayer<ElementwiseBinaryLayer>(BinaryOperation::Add, "addition"); Layer* const outputLayer = graph.AddLayer<armnn::OutputLayer>(0, "output"); TensorInfo fp16TensorInfo({1, 2, 3, 5}, armnn::DataType::Float16); @@ -57,7 +57,7 @@ TEST_CASE("Fp16AdditionTest") IConnectableLayer* inputLayer1 = net->AddInputLayer(0); IConnectableLayer* inputLayer2 = net->AddInputLayer(1); - IConnectableLayer* additionLayer = net->AddAdditionLayer(); + IConnectableLayer* additionLayer = net->AddElementwiseBinaryLayer(BinaryOperation::Add); IConnectableLayer* outputLayer = net->AddOutputLayer(0); inputLayer1->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0)); diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp index 968bce48c8..cea2aa3eba 100644 --- a/src/backends/neon/NeonBackend.cpp +++ b/src/backends/neon/NeonBackend.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -395,6 +395,92 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph untouched.erase(activationLayer->GetGuid()); } } + else if (base.GetType() == LayerType::ElementwiseBinary) + { + ElementwiseBinaryLayer* baseLayer = PolymorphicDowncast<ElementwiseBinaryLayer*>(&base); + + if (baseLayer->GetParameters().m_Operation == BinaryOperation::Add) + { + arm_compute::Status status = NeonAdditionWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Add, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Div) + { + arm_compute::Status status = NeonDivisionWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Div, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Mul) + { + arm_compute::Status status = NeonMultiplicationWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Mul, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + else if (baseLayer->GetParameters().m_Operation == BinaryOperation::Sub) + { + arm_compute::Status status = NeonSubtractionWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), + activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + &activationDesc); + + if (status) + { + FuseElementwiseBinaryLayer<ElementwiseBinaryLayer>(optimizationViews, + baseLayer, + activationLayer, + activationDesc, + BinaryOperation::Sub, + name); + untouched.erase(baseLayer->GetGuid()); + untouched.erase(activationLayer->GetGuid()); + } + } + // No fusion available for other BinaryOperations + } } } } diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp index ee155a2c64..672b2f377f 100644 --- a/src/backends/neon/NeonLayerSupport.cpp +++ b/src/backends/neon/NeonLayerSupport.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -346,6 +346,56 @@ bool IsLayerTypeSupported(const LayerType& type, } case LayerType::Division: return support.IsDivisionSupported(infos[0], infos[1], infos[2], reasonIfUnsupported); + case LayerType::ElementwiseBinary: + { + auto desc = *(PolymorphicDowncast<const ElementwiseBinaryDescriptor *>(&descriptor)); + + switch (desc.m_Operation) + { + case BinaryOperation::Add: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonAdditionWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Div: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonDivisionWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Maximum: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMaximumWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2]); + case BinaryOperation::Minimum: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMinimumWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2]); + case BinaryOperation::Mul: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMultiplicationWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + case BinaryOperation::Sub: + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonSubtractionWorkloadValidate, + reasonIfUnsupported, + infos[0], + infos[1], + infos[2], + nullptr); + default: + return false; + } + } case LayerType::ElementwiseUnary: return support.IsElementwiseUnarySupported(infos[0], infos[1], diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp index dccd4a3a36..08168eca2f 100644 --- a/src/backends/neon/NeonWorkloadFactory.cpp +++ b/src/backends/neon/NeonWorkloadFactory.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -280,6 +280,59 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateWorkload(LayerType type, auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor); return std::make_unique<NeonDivisionWorkload>(*divisionQueueDescriptor, info); } + case LayerType::ElementwiseBinary : + { + auto elementwiseBinaryQueueDescriptor + = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor); + + switch (elementwiseBinaryQueueDescriptor->m_Parameters.m_Operation) + { + case BinaryOperation::Add: + { + AdditionQueueDescriptor additionQueueDescriptor; + additionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + additionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonAdditionWorkload>(additionQueueDescriptor, info); + } + case BinaryOperation::Div: + { + DivisionQueueDescriptor divisionQueueDescriptor; + divisionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + divisionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonDivisionWorkload>(divisionQueueDescriptor, info); + } + case BinaryOperation::Maximum: + { + MaximumQueueDescriptor maximumQueueDescriptor; + maximumQueueDescriptor.m_Inputs = descriptor.m_Inputs; + maximumQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonMaximumWorkload>(maximumQueueDescriptor, info); + } + case BinaryOperation::Minimum: + { + MinimumQueueDescriptor minimumQueueDescriptor; + minimumQueueDescriptor.m_Inputs = descriptor.m_Inputs; + minimumQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonMinimumWorkload>(minimumQueueDescriptor, info); + } + case BinaryOperation::Mul: + { + MultiplicationQueueDescriptor multiplicationQueueDescriptor; + multiplicationQueueDescriptor.m_Inputs = descriptor.m_Inputs; + multiplicationQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonMultiplicationWorkload>(multiplicationQueueDescriptor, info); + } + case BinaryOperation::Sub: + { + SubtractionQueueDescriptor subtractionQueueDescriptor; + subtractionQueueDescriptor.m_Inputs = descriptor.m_Inputs; + subtractionQueueDescriptor.m_Outputs = descriptor.m_Outputs; + return std::make_unique<NeonSubtractionWorkload>(subtractionQueueDescriptor, info); + } + default: + return nullptr; + } + } case LayerType::ElementwiseUnary : { auto elementwiseUnaryQueueDescriptor diff --git a/src/backends/neon/test/NeonFallbackTests.cpp b/src/backends/neon/test/NeonFallbackTests.cpp index 8e0e0ab99b..40df2dc315 100644 --- a/src/backends/neon/test/NeonFallbackTests.cpp +++ b/src/backends/neon/test/NeonFallbackTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2020-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -38,8 +38,8 @@ TEST_CASE("FallbackImportToCpuAcc") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -183,7 +183,7 @@ TEST_CASE("FallbackPaddingCopyToCpuAcc") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); - IConnectableLayer* add = net->AddAdditionLayer("add"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); @@ -318,8 +318,8 @@ TEST_CASE("FallbackImportFromCpuAcc") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); - IConnectableLayer* add = net->AddAdditionLayer("add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); @@ -465,7 +465,7 @@ TEST_CASE("FallbackPaddingCopyFromCpuAcc") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); - IConnectableLayer* add = net->AddAdditionLayer("add"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(pooling->GetInputSlot(0)); @@ -599,8 +599,8 @@ TEST_CASE("FallbackDisableImportFromCpuAcc") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); - IConnectableLayer* add = net->AddAdditionLayer("add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(sub->GetInputSlot(0)); @@ -725,8 +725,8 @@ TEST_CASE("NeonImportEnabledFallbackToCl") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -878,8 +878,8 @@ TEST_CASE("NeonImportDisabledFallbackToCl") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); input0->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -1013,8 +1013,8 @@ TEST_CASE("NeonImportEnabledFallbackSubgraphToCl") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); @@ -1177,8 +1177,8 @@ TEST_CASE("NeonImportDisableFallbackSubgraphToCl") IConnectableLayer* input0 = net->AddInputLayer(0, "input0"); IConnectableLayer* input1 = net->AddInputLayer(1, "input1"); IConnectableLayer* input2 = net->AddInputLayer(2, "input2"); - IConnectableLayer* add = net->AddAdditionLayer("add"); - IConnectableLayer* sub = net->AddSubtractionLayer("sub"); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(BinaryOperation::Add, "add"); + IConnectableLayer* sub = net->AddElementwiseBinaryLayer(BinaryOperation::Sub, "sub"); IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling"); IConnectableLayer* output = net->AddOutputLayer(0, "output"); diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp index a5015a7376..cbc6723dbc 100644 --- a/src/backends/reference/RefLayerSupport.cpp +++ b/src/backends/reference/RefLayerSupport.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -189,6 +189,36 @@ bool RefLayerSupport::IsLayerSupported(const LayerType& type, return IsDequantizeSupported(infos[0], infos[1], reasonIfUnsupported); case LayerType::Division: return IsDivisionSupported(infos[0], infos[1], infos[2], reasonIfUnsupported); + case LayerType::ElementwiseBinary: + { + std::array<DataType, 7> supportedTypes = + { + DataType::Float32, + DataType::Float16, + DataType::QAsymmS8, + DataType::QAsymmU8, + DataType::QSymmS16, + DataType::Signed32 + }; + + bool supported = true; + supported &= CheckSupportRule(TypeAnyOf(infos[0], supportedTypes), reasonIfUnsupported, + "Reference elementwise unary: input type not supported"); + + supported &= CheckSupportRule(TypeAnyOf(infos[1], supportedTypes), reasonIfUnsupported, + "Reference elementwise unary: input type not supported"); + + supported &= CheckSupportRule(TypeAnyOf(infos[2], supportedTypes), reasonIfUnsupported, + "Reference elementwise unary: output type not supported"); + + supported &= CheckSupportRule(TypesAreEqual(infos[0], infos[1]), reasonIfUnsupported, + "Reference elementwise unary: input types not matching"); + + supported &= CheckSupportRule(TypesAreEqual(infos[0], infos[2]), reasonIfUnsupported, + "Reference elementwise unary: input and output types not matching"); + + return supported; + } case LayerType::ElementwiseUnary: return IsElementwiseUnarySupported(infos[0], infos[1], diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp index f0e9e35978..8e1f68ebfc 100644 --- a/src/backends/reference/RefLayerSupport.hpp +++ b/src/backends/reference/RefLayerSupport.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp index bfe37d7bf5..10f623eef3 100644 --- a/src/backends/reference/RefWorkloadFactory.cpp +++ b/src/backends/reference/RefWorkloadFactory.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #include <Layer.hpp> @@ -302,6 +302,12 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateWorkload(LayerType type, return std::make_unique<RefDivisionWorkload<float>>(*divisionQueueDescriptor, info); } } + case LayerType::ElementwiseBinary: + { + auto elementwiseBinaryQueueDescriptor + = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor); + return std::make_unique<RefElementwiseBinaryWorkload>(*elementwiseBinaryQueueDescriptor, info); + } case LayerType::ElementwiseUnary: { auto elementwiseUnaryQueueDescriptor diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk index eb2ec2df44..c23984c3e9 100644 --- a/src/backends/reference/backend.mk +++ b/src/backends/reference/backend.mk @@ -67,6 +67,7 @@ BACKEND_SOURCES := \ workloads/RefDepthwiseConvolution2dWorkload.cpp \ workloads/RefDequantizeWorkload.cpp \ workloads/RefDetectionPostProcessWorkload.cpp \ + workloads/RefElementwiseBinaryWorkload.cpp \ workloads/RefElementwiseWorkload.cpp \ workloads/RefElementwiseUnaryWorkload.cpp \ workloads/RefFakeQuantizationFloat32Workload.cpp \ diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp index 6ff57716d0..8bf414fdb0 100644 --- a/src/backends/reference/test/RefEndToEndTests.cpp +++ b/src/backends/reference/test/RefEndToEndTests.cpp @@ -18,6 +18,7 @@ #include <backendsCommon/test/DepthToSpaceEndToEndTestImpl.hpp> #include <backendsCommon/test/DequantizeEndToEndTestImpl.hpp> #include <backendsCommon/test/DetectionPostProcessEndToEndTestImpl.hpp> +#include <backendsCommon/test/ElementwiseBinaryEndToEndTestImpl.hpp> #include <backendsCommon/test/ElementwiseUnaryEndToEndTestImpl.hpp> #include <backendsCommon/test/FillEndToEndTestImpl.hpp> #include <backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp> @@ -185,7 +186,7 @@ TEST_CASE("TrivialAdd") IConnectableLayer* input1 = net->AddInputLayer(0); IConnectableLayer* input2 = net->AddInputLayer(1); - IConnectableLayer* add = net->AddAdditionLayer(); + IConnectableLayer* add = net->AddElementwiseBinaryLayer(ElementwiseBinaryDescriptor(BinaryOperation::Add)); IConnectableLayer* output = net->AddOutputLayer(0); input1->GetOutputSlot(0).Connect(add->GetInputSlot(0)); @@ -347,7 +348,7 @@ TEST_CASE("TrivialMin") IConnectableLayer* input1 = net->AddInputLayer(0); IConnectableLayer* input2 = net->AddInputLayer(1); - IConnectableLayer* min = net->AddMinimumLayer(); + IConnectableLayer* min = net->AddElementwiseBinaryLayer(ElementwiseBinaryDescriptor(BinaryOperation::Minimum)); IConnectableLayer* output = net->AddOutputLayer(0); input1->GetOutputSlot(0).Connect(min->GetInputSlot(0)); @@ -1547,6 +1548,55 @@ TEST_CASE("RefAsyncFP32StridedSlicedScheduledMultiThreadedEndToEndTest") { armnn::experimental::StridedSlicedEndToEndTest<armnn::DataType::Float32>(defaultBackends, 3); } + +TEST_CASE("RefAddEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Add); +} +TEST_CASE("RefAddEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Add); +} +TEST_CASE("RefDivEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Div); +} +TEST_CASE("RefDivEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Div); +} +TEST_CASE("RefMulEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Mul); +} +TEST_CASE("RefMulEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Mul); +} +TEST_CASE("RefSubEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Sub); +} +TEST_CASE("RefSubEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Sub); +} +TEST_CASE("RefMaximumEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Maximum); +} +TEST_CASE("RefMaximumEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Maximum); +} +TEST_CASE("RefMinimumEndToEndTestFloat32") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::Float32>(defaultBackends, BinaryOperation::Minimum); +} +TEST_CASE("RefMinimumEndToEndTestUint8") +{ + ElementwiseBinarySimpleEndToEnd<armnn::DataType::QAsymmU8>(defaultBackends, BinaryOperation::Minimum); +} #endif } diff --git a/src/backends/reference/test/RefOptimizedNetworkTests.cpp b/src/backends/reference/test/RefOptimizedNetworkTests.cpp index 7ca1e0e505..7e8064fc76 100644 --- a/src/backends/reference/test/RefOptimizedNetworkTests.cpp +++ b/src/backends/reference/test/RefOptimizedNetworkTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017,2022 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -53,7 +53,7 @@ TEST_CASE("OptimizeValidateCpuRefWorkloads") layer->GetOutputSlot(0).SetTensorInfo(desc); armnn::IConnectableLayer* prevLayer = layer; - layer = net->AddMultiplicationLayer("ml"); + layer = net->AddElementwiseBinaryLayer(armnn::BinaryOperation::Mul, "ml"); prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt index de6c042959..3592f2293d 100644 --- a/src/backends/reference/workloads/CMakeLists.txt +++ b/src/backends/reference/workloads/CMakeLists.txt @@ -1,5 +1,5 @@ # -# Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +# Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. # SPDX-License-Identifier: MIT # @@ -108,6 +108,8 @@ list(APPEND armnnRefBackendWorkloads_sources RefDequantizeWorkload.hpp RefDetectionPostProcessWorkload.cpp RefDetectionPostProcessWorkload.hpp + RefElementwiseBinaryWorkload.cpp + RefElementwiseBinaryWorkload.hpp RefElementwiseUnaryWorkload.cpp RefElementwiseUnaryWorkload.hpp RefFakeQuantizationFloat32Workload.cpp diff --git a/src/backends/reference/workloads/RefElementwiseBinaryWorkload.cpp b/src/backends/reference/workloads/RefElementwiseBinaryWorkload.cpp new file mode 100644 index 0000000000..5dc77f8496 --- /dev/null +++ b/src/backends/reference/workloads/RefElementwiseBinaryWorkload.cpp @@ -0,0 +1,120 @@ +// +// Copyright © 2023 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "RefElementwiseBinaryWorkload.hpp" + +#include "Decoders.hpp" +#include "ElementwiseFunction.hpp" +#include "Encoders.hpp" +#include "RefWorkloadUtils.hpp" +#include "Maximum.hpp" +#include "Minimum.hpp" + +#include <Profiling.hpp> + +#include <armnn/TypesUtils.hpp> + +#include <functional> + +namespace armnn +{ + +template<typename DataType> +void ExecuteFunction(std::vector<ITensorHandle*> inputs, + std::vector<ITensorHandle*> outputs, + BinaryOperation operation) +{ + const TensorInfo& inputInfo0 = GetTensorInfo(inputs[0]); + const TensorInfo& inputInfo1 = GetTensorInfo(inputs[1]); + const TensorInfo& outputInfo = GetTensorInfo(outputs[0]); + + const TensorShape& inShape0 = inputInfo0.GetShape(); + const TensorShape& inShape1 = inputInfo1.GetShape(); + const TensorShape& outShape = outputInfo.GetShape(); + + std::unique_ptr<Decoder<DataType>> input0 = MakeDecoder<DataType>(inputInfo0, inputs[0]->Map()); + std::unique_ptr<Decoder<DataType>> input1 = MakeDecoder<DataType>(inputInfo1, inputs[1]->Map()); + std::unique_ptr<Encoder<DataType>> output = MakeEncoder<DataType>(outputInfo, outputs[0]->Map()); + + using AddFunction = ElementwiseBinaryFunction<std::plus<DataType>>; + using DivFunction = ElementwiseBinaryFunction<std::divides<DataType>>; + using MaximumFunction = ElementwiseBinaryFunction<armnn::maximum<DataType>>; + using MinimumFunction = ElementwiseBinaryFunction<armnn::minimum<DataType>>; + using MulFunction = ElementwiseBinaryFunction<std::multiplies<DataType>>; + using SubFunction = ElementwiseBinaryFunction<std::minus<DataType>>; + + switch (operation) + { + case BinaryOperation::Add: + { + AddFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + case BinaryOperation::Div: + { + DivFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + case BinaryOperation::Maximum: + { + MaximumFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + case BinaryOperation::Minimum: + { + MinimumFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + case BinaryOperation::Mul: + { + MulFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + case BinaryOperation::Sub: + { + SubFunction(inShape0, inShape1, outShape, *input0, *input1, *output); + break; + } + default: + { + throw InvalidArgumentException(std::string("Unsupported binary operation ") + + GetBinaryOperationAsCString(operation), CHECK_LOCATION()); + } + } +} + +RefElementwiseBinaryWorkload::RefElementwiseBinaryWorkload(const ElementwiseBinaryQueueDescriptor& desc, + const WorkloadInfo& info) + : RefBaseWorkload<ElementwiseBinaryQueueDescriptor>(desc, info) +{} + +void RefElementwiseBinaryWorkload::Execute() const +{ + Execute(m_Data.m_Inputs, m_Data.m_Outputs); +} + +void RefElementwiseBinaryWorkload::ExecuteAsync(ExecutionData& executionData) +{ + + WorkingMemDescriptor* workingMemDescriptor = static_cast<WorkingMemDescriptor*>(executionData.m_Data); + Execute(workingMemDescriptor->m_Inputs, workingMemDescriptor->m_Outputs); +} + +void RefElementwiseBinaryWorkload::Execute(std::vector<ITensorHandle*> inputs, + std::vector<ITensorHandle*> outputs) const +{ + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefElementwiseBinaryWorkload_Execute"); + + if (GetTensorInfo(inputs[0]).GetDataType() == DataType::Signed32) + { + ExecuteFunction<int32_t>(inputs, outputs, m_Data.m_Parameters.m_Operation); + } + else + { + ExecuteFunction<float>(inputs, outputs, m_Data.m_Parameters.m_Operation); + } +} + +} // namespace armnn diff --git a/src/backends/reference/workloads/RefElementwiseBinaryWorkload.hpp b/src/backends/reference/workloads/RefElementwiseBinaryWorkload.hpp new file mode 100644 index 0000000000..37458a1705 --- /dev/null +++ b/src/backends/reference/workloads/RefElementwiseBinaryWorkload.hpp @@ -0,0 +1,29 @@ +// +// Copyright © 2023 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "BaseIterator.hpp" + +#include "RefBaseWorkload.hpp" +#include <armnn/backends/WorkloadData.hpp> + +namespace armnn +{ + +class RefElementwiseBinaryWorkload : public RefBaseWorkload<ElementwiseBinaryQueueDescriptor> +{ +public: + using RefBaseWorkload<ElementwiseBinaryQueueDescriptor>::m_Data; + + RefElementwiseBinaryWorkload(const ElementwiseBinaryQueueDescriptor& descriptor, const WorkloadInfo& info); + void Execute() const override; + void ExecuteAsync(ExecutionData& executionData) override; + +private: + void Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const; +}; + +} // namespace armnn diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp index afed71bfff..dba880bafc 100644 --- a/src/backends/reference/workloads/RefWorkloads.hpp +++ b/src/backends/reference/workloads/RefWorkloads.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -26,6 +26,7 @@ #include "RefDetectionPostProcessWorkload.hpp" #include "RefDequantizeWorkload.hpp" #include "RefElementwiseWorkload.hpp" +#include "RefElementwiseBinaryWorkload.hpp" #include "RefElementwiseUnaryWorkload.hpp" #include "RefFakeQuantizationFloat32Workload.hpp" #include "RefFillWorkload.hpp" |