diff options
author | Derek Lamberti <derek.lamberti@arm.com> | 2019-03-25 15:41:58 +0000 |
---|---|---|
committer | derek.lamberti <derek.lamberti@arm.com> | 2019-03-27 16:06:46 +0000 |
commit | a9cca6aa935b3e290181a05fdb2c5f5557a49c09 (patch) | |
tree | 60493eeada1357fa2334f7f952879561480e192f | |
parent | 045ea78ab73b1c88a38d3ecb501583a38161ca0e (diff) | |
download | armnn-a9cca6aa935b3e290181a05fdb2c5f5557a49c09.tar.gz |
IVGCVSW-2870 Support QuantizeLayer on frontend
Change-Id: I2014a8d801f1f222d27a80dddf4f188ddcb3a5c9
Signed-off-by: Derek Lamberti <derek.lamberti@arm.com>
22 files changed, 548 insertions, 220 deletions
diff --git a/Android.mk b/Android.mk index 1b0ffea024..c65fc5ef93 100644 --- a/Android.mk +++ b/Android.mk @@ -116,6 +116,7 @@ LOCAL_SRC_FILES := \ src/armnn/layers/PermuteLayer.cpp \ src/armnn/layers/Pooling2dLayer.cpp \ src/armnn/layers/PreCompiledLayer.cpp \ + src/armnn/layers/QuantizeLayer.cpp \ src/armnn/layers/ReshapeLayer.cpp \ src/armnn/layers/ResizeBilinearLayer.cpp \ src/armnn/layers/RsqrtLayer.cpp \ diff --git a/CMakeLists.txt b/CMakeLists.txt index fe1c1a71c9..dcdc4bd9de 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -253,6 +253,8 @@ list(APPEND armnn_sources src/armnn/layers/PermuteLayer.cpp src/armnn/layers/Pooling2dLayer.hpp src/armnn/layers/Pooling2dLayer.cpp + src/armnn/layers/QuantizeLayer.cpp + src/armnn/layers/QuantizeLayer.hpp src/armnn/layers/DivisionLayer.cpp src/armnn/layers/DivisionLayer.hpp src/armnn/layers/PreCompiledLayer.hpp diff --git a/include/armnn/ILayerSupport.hpp b/include/armnn/ILayerSupport.hpp index 7896a80a8f..76d9c10d7f 100644 --- a/include/armnn/ILayerSupport.hpp +++ b/include/armnn/ILayerSupport.hpp @@ -209,6 +209,10 @@ public: const PreCompiledDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0; + virtual bool IsQuantizeSupported(const TensorInfo& input, + const TensorInfo& output, + Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0; + virtual bool IsReshapeSupported(const TensorInfo& input, const ReshapeDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0; diff --git a/include/armnn/ILayerVisitor.hpp b/include/armnn/ILayerVisitor.hpp index ce09246d07..76027f6eed 100644 --- a/include/armnn/ILayerVisitor.hpp +++ b/include/armnn/ILayerVisitor.hpp @@ -281,6 +281,12 @@ public: const PadDescriptor& padDescriptor, const char* name = nullptr) = 0; + /// Function a quantize layer should call back to when its Accept(ILayerVisitor&) function is invoked. + /// @param layer - pointer to the layer which is calling back to this visit function. + /// @param name - Optional name for the layer. + virtual void VisitQuantizeLayer(const IConnectableLayer* layer, + const char* name = nullptr) = 0; + /// Function a strided slice layer should call back to when its Accept(ILayerVisitor&) function is invoked. /// @param layer - pointer to the layer which is calling back to this visit function. /// @param StridedSliceDescriptor - Parameters for the strided slice operation. diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp index b61206fae1..7c1fbdbd49 100644 --- a/include/armnn/INetwork.hpp +++ b/include/armnn/INetwork.hpp @@ -353,6 +353,11 @@ public: virtual IConnectableLayer* AddPadLayer(const PadDescriptor& padDescriptor, const char* name = nullptr) = 0; + /// Add a quantize layer to the network + ///@param name - Optional name for the layer. + /// @return - Interface for configuring the layer. + virtual IConnectableLayer* AddQuantizeLayer(const char* name = nullptr) = 0; + /// Adds a strided slice layer to the network. /// @param StridedSliceDescriptor - Parameters for the strided slice operation. /// @param name - Optional name for the layer. diff --git a/include/armnn/LayerVisitorBase.hpp b/include/armnn/LayerVisitorBase.hpp index 2c37a21786..d30005f0ed 100644 --- a/include/armnn/LayerVisitorBase.hpp +++ b/include/armnn/LayerVisitorBase.hpp @@ -151,6 +151,9 @@ public: const PadDescriptor&, const char*) override { DefaultPolicy::Apply(); } + void VisitQuantizeLayer(const IConnectableLayer*, + const char*) override { DefaultPolicy::Apply(); } + void VisitStridedSliceLayer(const IConnectableLayer*, const StridedSliceDescriptor&, const char*) override { DefaultPolicy::Apply(); } diff --git a/src/armnn/InternalTypes.hpp b/src/armnn/InternalTypes.hpp index a61c7b8147..d6d66031a7 100644 --- a/src/armnn/InternalTypes.hpp +++ b/src/armnn/InternalTypes.hpp @@ -47,6 +47,7 @@ enum class LayerType Permute, Pooling2d, PreCompiled, + Quantize, Reshape, ResizeBilinear, Rsqrt, diff --git a/src/armnn/LayerSupport.cpp b/src/armnn/LayerSupport.cpp index 5916488fb5..0e3d2522e4 100644 --- a/src/armnn/LayerSupport.cpp +++ b/src/armnn/LayerSupport.cpp @@ -406,6 +406,15 @@ bool IsPadSupported(const BackendId& backend, FORWARD_LAYER_SUPPORT_FUNC(backend, IsPadSupported, input, output, descriptor); } +bool IsQuantizeSupported(const BackendId& backend, + const TensorInfo& input, + const TensorInfo& output, + char* reasonIfUnsupported, + size_t reasonIfUnsupportedMaxLength) +{ + FORWARD_LAYER_SUPPORT_FUNC(backend, IsQuantizeSupported, input, output); +} + bool IsPermuteSupported(const BackendId& backend, const TensorInfo& input, const TensorInfo& output, diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp index 497b517f28..9300a75a07 100644 --- a/src/armnn/LayersFwd.hpp +++ b/src/armnn/LayersFwd.hpp @@ -39,6 +39,7 @@ #include "layers/PermuteLayer.hpp" #include "layers/Pooling2dLayer.hpp" #include "layers/PreCompiledLayer.hpp" +#include "layers/QuantizeLayer.hpp" #include "layers/ReshapeLayer.hpp" #include "layers/ResizeBilinearLayer.hpp" #include "layers/RsqrtLayer.hpp" @@ -108,6 +109,7 @@ DECLARE_LAYER(Pad) DECLARE_LAYER(Permute) DECLARE_LAYER(Pooling2d) DECLARE_LAYER(PreCompiled) +DECLARE_LAYER(Quantize) DECLARE_LAYER(Reshape) DECLARE_LAYER(ResizeBilinear) DECLARE_LAYER(Rsqrt) diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp index bd5adccf86..c5dfbd75ec 100644 --- a/src/armnn/Network.cpp +++ b/src/armnn/Network.cpp @@ -930,6 +930,11 @@ IConnectableLayer* Network::AddPadLayer(const PadDescriptor& padDescriptor, cons return m_Graph->AddLayer<PadLayer>(padDescriptor,name); } +IConnectableLayer *Network::AddQuantizeLayer(const char *name) +{ + return m_Graph->AddLayer<QuantizeLayer>(name); +} + IConnectableLayer* Network::AddStridedSliceLayer(const StridedSliceDescriptor& stridedSliceDescriptor, const char* name) { diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp index e50ce79052..5ed8cca2f2 100644 --- a/src/armnn/Network.hpp +++ b/src/armnn/Network.hpp @@ -159,6 +159,8 @@ public: IConnectableLayer* AddPadLayer(const PadDescriptor& padDescriptor, const char* name = nullptr) override; + IConnectableLayer* AddQuantizeLayer(const char* name = nullptr) override; + IConnectableLayer* AddStridedSliceLayer(const StridedSliceDescriptor& stridedSliceDescriptor, const char* name = nullptr) override; diff --git a/src/armnn/layers/QuantizeLayer.cpp b/src/armnn/layers/QuantizeLayer.cpp new file mode 100644 index 0000000000..fbf8b322ab --- /dev/null +++ b/src/armnn/layers/QuantizeLayer.cpp @@ -0,0 +1,48 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "QuantizeLayer.hpp" + +#include "LayerCloneBase.hpp" + +#include <armnn/ILayerVisitor.hpp> + +namespace armnn +{ + +QuantizeLayer::QuantizeLayer(const char* name) +: Layer(1, 1, LayerType::Quantize, name) +{} + +std::unique_ptr<IWorkload> QuantizeLayer::CreateWorkload(const Graph& graph, + const IWorkloadFactory& factory) const +{ + return nullptr; +} + +Layer* QuantizeLayer::Clone(Graph& graph) const +{ + QuantizeLayer* clone = CloneBase<QuantizeLayer>(graph, GetName()); + return clone; +} + +void QuantizeLayer::ValidateTensorShapesFromInputs() +{ + VerifyLayerConnections(1, CHECK_LOCATION()); + + auto inferredShapes = InferOutputShapes({ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape() }); + + ConditionalThrowIfNotEqual<LayerValidationException>( + "QuantizeLayer: TensorShape set on OutputSlot[0] does not match the inferred shape.", + GetOutputSlot(0).GetTensorInfo().GetShape(), + inferredShapes[0]); +} + +void QuantizeLayer::Accept(ILayerVisitor& visitor) const +{ + visitor.VisitQuantizeLayer(this, GetName()); +} + +} //namespace armnn
\ No newline at end of file diff --git a/src/armnn/layers/QuantizeLayer.hpp b/src/armnn/layers/QuantizeLayer.hpp new file mode 100644 index 0000000000..fabb4492c5 --- /dev/null +++ b/src/armnn/layers/QuantizeLayer.hpp @@ -0,0 +1,35 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include <Layer.hpp> + +namespace armnn { + +//Forward +class IWorkload; +class IWorkloadFactory; +class ILayerVisitor; + +class QuantizeLayer : public Layer +{ +public: + virtual std::unique_ptr<IWorkload> CreateWorkload(const Graph& graph, + const IWorkloadFactory& factory) const override; + + Layer* Clone(Graph& graph) const override; + + void ValidateTensorShapesFromInputs() override; + + void Accept(ILayerVisitor& visitor) const override; + +protected: + QuantizeLayer(const char* name); + ~QuantizeLayer() = default; + +}; + +} //namespace armnn diff --git a/src/armnn/test/NetworkTests.cpp b/src/armnn/test/NetworkTests.cpp index 0963df6222..4de09a2804 100644 --- a/src/armnn/test/NetworkTests.cpp +++ b/src/armnn/test/NetworkTests.cpp @@ -6,6 +6,7 @@ #include "GraphUtils.hpp" #include <armnn/ArmNN.hpp> +#include <armnn/LayerVisitorBase.hpp> #include <Network.hpp> #include <boost/test/unit_test.hpp> @@ -366,4 +367,54 @@ BOOST_AUTO_TEST_CASE(NetworkModification_SplitterMultiplication) prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); } +BOOST_AUTO_TEST_CASE(Network_AddQuantize) +{ + struct Test : public armnn::LayerVisitorBase<armnn::VisitorNoThrowPolicy> + { + void VisitQuantizeLayer(const armnn::IConnectableLayer* layer, const char* name) override + { + m_Visited = true; + + BOOST_TEST(layer); + + std::string expectedName = std::string("quantize"); + BOOST_TEST(std::string(layer->GetName()) == expectedName); + BOOST_TEST(std::string(name) == expectedName); + + BOOST_TEST(layer->GetNumInputSlots() == 1); + BOOST_TEST(layer->GetNumOutputSlots() == 1); + + const armnn::TensorInfo& infoIn = layer->GetInputSlot(0).GetConnection()->GetTensorInfo(); + BOOST_TEST((infoIn.GetDataType() == armnn::DataType::Float32)); + + const armnn::TensorInfo& infoOut = layer->GetOutputSlot(0).GetTensorInfo(); + BOOST_TEST((infoOut.GetDataType() == armnn::DataType::QuantisedAsymm8)); + } + + bool m_Visited = false; + }; + + + auto graph = armnn::INetwork::Create(); + + auto input = graph->AddInputLayer(0, "input"); + auto quantize = graph->AddQuantizeLayer("quantize"); + auto output = graph->AddOutputLayer(1, "output"); + + input->GetOutputSlot(0).Connect(quantize->GetInputSlot(0)); + quantize->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + armnn::TensorInfo infoIn({3,1}, armnn::DataType::Float32); + input->GetOutputSlot(0).SetTensorInfo(infoIn); + + armnn::TensorInfo infoOut({3,1}, armnn::DataType::QuantisedAsymm8); + quantize->GetOutputSlot(0).SetTensorInfo(infoOut); + + Test testQuantize; + graph->Accept(testQuantize); + + BOOST_TEST(testQuantize.m_Visited == true); + +} + BOOST_AUTO_TEST_SUITE_END() diff --git a/src/armnn/test/TestLayerVisitor.hpp b/src/armnn/test/TestLayerVisitor.hpp index 8c0da50be3..e43227f520 100644 --- a/src/armnn/test/TestLayerVisitor.hpp +++ b/src/armnn/test/TestLayerVisitor.hpp @@ -4,13 +4,13 @@ // #pragma once -#include <armnn/ILayerVisitor.hpp> +#include <armnn/LayerVisitorBase.hpp> #include <armnn/Descriptors.hpp> namespace armnn { // Abstract base class with do nothing implementations for all layer visit methods -class TestLayerVisitor : public ILayerVisitor +class TestLayerVisitor : public LayerVisitorBase<VisitorNoThrowPolicy> { protected: virtual ~TestLayerVisitor() {} @@ -34,147 +34,6 @@ public: m_LayerName = ""; } } - - void VisitInputLayer(const IConnectableLayer* layer, - LayerBindingId id, - const char* name = nullptr) override {} - - void VisitConvolution2dLayer(const IConnectableLayer* layer, - const Convolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional<ConstTensor>& biases, - const char* name = nullptr) override {} - - void VisitDepthwiseConvolution2dLayer(const IConnectableLayer* layer, - const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional<ConstTensor>& biases, - const char* name = nullptr) override {} - - void VisitDetectionPostProcessLayer(const IConnectableLayer* layer, - const DetectionPostProcessDescriptor& descriptor, - const ConstTensor& anchors, - const char* name = nullptr) override {} - - void VisitFullyConnectedLayer(const IConnectableLayer* layer, - const FullyConnectedDescriptor& fullyConnectedDescriptor, - const ConstTensor& weights, - const Optional<ConstTensor>& biases, - const char* name = nullptr) override {} - - void VisitPermuteLayer(const IConnectableLayer* layer, - const PermuteDescriptor& permuteDescriptor, - const char* name = nullptr) override {} - - void VisitBatchToSpaceNdLayer(const IConnectableLayer* layer, - const BatchToSpaceNdDescriptor& batchToSpaceNdDescriptor, - const char* name = nullptr) override {} - - void VisitPooling2dLayer(const IConnectableLayer* layer, - const Pooling2dDescriptor& pooling2dDescriptor, - const char* name = nullptr) override {} - - void VisitActivationLayer(const IConnectableLayer* layer, - const ActivationDescriptor& activationDescriptor, - const char* name = nullptr) override {} - - void VisitNormalizationLayer(const IConnectableLayer* layer, - const NormalizationDescriptor& normalizationDescriptor, - const char* name = nullptr) override {} - - void VisitSoftmaxLayer(const IConnectableLayer* layer, - const SoftmaxDescriptor& softmaxDescriptor, - const char* name = nullptr) override {} - - void VisitSplitterLayer(const IConnectableLayer* layer, - const ViewsDescriptor& splitterDescriptor, - const char* name = nullptr) override {} - - void VisitMergerLayer(const IConnectableLayer* layer, - const OriginsDescriptor& mergerDescriptor, - const char* name = nullptr) override {} - - void VisitAdditionLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitMultiplicationLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitBatchNormalizationLayer(const IConnectableLayer* layer, - const BatchNormalizationDescriptor& desc, - const ConstTensor& mean, - const ConstTensor& variance, - const ConstTensor& beta, - const ConstTensor& gamma, - const char* name = nullptr) override {} - - void VisitResizeBilinearLayer(const IConnectableLayer* layer, - const ResizeBilinearDescriptor& resizeDesc, - const char* name = nullptr) override {} - - void VisitL2NormalizationLayer(const IConnectableLayer* layer, - const L2NormalizationDescriptor& desc, - const char* name = nullptr) override {} - - void VisitConstantLayer(const IConnectableLayer* layer, - const ConstTensor& input, - const char* name = nullptr) override {} - - void VisitReshapeLayer(const IConnectableLayer* layer, - const ReshapeDescriptor& reshapeDescriptor, - const char* name = nullptr) override {} - - void VisitSpaceToBatchNdLayer(const IConnectableLayer* layer, - const SpaceToBatchNdDescriptor& spaceToBatchNdDescriptor, - const char* name = nullptr) override {} - - void VisitFloorLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitOutputLayer(const IConnectableLayer* layer, - LayerBindingId id, - const char* name = nullptr) override {} - - void VisitLstmLayer(const IConnectableLayer* layer, - const LstmDescriptor& descriptor, - const LstmInputParams& params, - const char* name = nullptr) override {} - - void VisitDivisionLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitSubtractionLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitMaximumLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitMeanLayer(const IConnectableLayer* layer, - const MeanDescriptor& meanDescriptor, - const char* name = nullptr) override {} - - void VisitPadLayer(const IConnectableLayer* layer, - const PadDescriptor& padDescriptor, - const char* name = nullptr) override {} - - void VisitStridedSliceLayer(const IConnectableLayer* layer, - const StridedSliceDescriptor& stridedSliceDescriptor, - const char* name = nullptr) override {} - - void VisitMinimumLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitGreaterLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitEqualLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitRsqrtLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} - - void VisitGatherLayer(const IConnectableLayer* layer, - const char* name = nullptr) override {} }; } //namespace armnn diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp index c170ac05e2..137e77eebe 100644 --- a/src/backends/backendsCommon/LayerSupportBase.cpp +++ b/src/backends/backendsCommon/LayerSupportBase.cpp @@ -315,6 +315,13 @@ bool LayerSupportBase::IsPreCompiledSupported(const TensorInfo& input, return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported); } +bool LayerSupportBase::IsQuantizeSupported(const armnn::TensorInfo& input, + const armnn::TensorInfo& output, + armnn::Optional<std::string&> reasonIfUnsupported) const +{ + return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported); +} + bool LayerSupportBase::IsReshapeSupported(const TensorInfo& input, const ReshapeDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported) const diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp index 75c366cf3b..ceb3b2768e 100644 --- a/src/backends/backendsCommon/LayerSupportBase.hpp +++ b/src/backends/backendsCommon/LayerSupportBase.hpp @@ -198,6 +198,10 @@ public: const PreCompiledDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override; + bool IsQuantizeSupported(const TensorInfo& input, + const TensorInfo& output, + Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override; + bool IsReshapeSupported(const TensorInfo& input, const ReshapeDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override; diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp index 7474b9bc9a..e30a3f36b7 100644 --- a/src/backends/backendsCommon/WorkloadData.cpp +++ b/src/backends/backendsCommon/WorkloadData.cpp @@ -946,6 +946,24 @@ void PadQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const } } +void QuantizeQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const +{ + ValidateSingleInput(workloadInfo, "QuantizeQueueDescriptor"); + ValidateSingleOutput(workloadInfo, "QuantizeQueueDescriptor"); + + + if (workloadInfo.m_InputTensorInfos[0].GetDataType() != DataType::Float32) + { + throw InvalidArgumentException("Quantize only accepts Float32 inputs."); + } + + if (workloadInfo.m_OutputTensorInfos[0].GetDataType() != DataType::QuantisedAsymm8 && + workloadInfo.m_OutputTensorInfos[0].GetDataType() != DataType::QuantisedSymm16) + { + throw InvalidArgumentException("Output of quantized layer must be quantized type."); + } +} + void BatchToSpaceNdQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const { ValidateSingleInput(workloadInfo, "BatchToSpaceNdQueueDescriptor"); diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp index 18bd921d5e..9250ceac43 100644 --- a/src/backends/backendsCommon/WorkloadData.hpp +++ b/src/backends/backendsCommon/WorkloadData.hpp @@ -229,6 +229,11 @@ struct PadQueueDescriptor : QueueDescriptorWithParameters<PadDescriptor> void Validate(const WorkloadInfo& workloadInfo) const; }; +struct QuantizeQueueDescriptor : QueueDescriptor +{ + void Validate(const WorkloadInfo& workloadInfo) const; +}; + // Equal layer workload data struct EqualQueueDescriptor : QueueDescriptor { diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp index 0996a8aaee..833f3b894e 100644 --- a/src/backends/backendsCommon/WorkloadFactory.cpp +++ b/src/backends/backendsCommon/WorkloadFactory.cpp @@ -4,6 +4,8 @@ // #include "CpuTensorHandle.hpp" +#include "WorkloadFactory.hpp" + #include <Layer.hpp> #include <LayersFwd.hpp> @@ -605,6 +607,13 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId, reason); break; } + case LayerType::Quantize: + { + const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); + const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo(); + result = layerSupportObject->IsQuantizeSupported(input, output, reason); + break; + } case LayerType::Division: { const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo(); @@ -751,4 +760,251 @@ bool IWorkloadFactory::IsLayerSupported(const IConnectableLayer& connectableLaye return IsLayerSupported(layer->GetBackendId(), connectableLayer, dataType, outReasonIfUnsupported); } +// Default Implementations +std::unique_ptr<IWorkload> IWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateBatchNormalization( + const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateDebug(const DebugQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateDepthwiseConvolution2d( + const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateDetectionPostProcess( + const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateDivision(const DivisionQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateEqual(const EqualQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateFakeQuantization(const FakeQuantizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateFullyConnected(const FullyConnectedQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateGather(const GatherQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateGreater(const GreaterQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMerger(const MergerQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMinimum(const MinimumQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateMultiplication(const MultiplicationQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateNormalization(const NormalizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateQuantize(const QuantizeQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateResizeBilinear(const ResizeBilinearQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateRsqrt(const RsqrtQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor, + const WorkloadInfo& Info) const +{ + return std::unique_ptr<IWorkload>(); +} + +std::unique_ptr<IWorkload> IWorkloadFactory::CreateSubtraction(const SubtractionQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::unique_ptr<IWorkload>(); +} + } diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp index 88833491b2..2aa3854c4a 100644 --- a/src/backends/backendsCommon/WorkloadFactory.hpp +++ b/src/backends/backendsCommon/WorkloadFactory.hpp @@ -49,125 +49,128 @@ public: virtual std::unique_ptr<ITensorHandle> CreateTensorHandle(const TensorInfo& tensorInfo, DataLayout dataLayout) const = 0; - virtual std::unique_ptr<IWorkload> CreateOutput(const OutputQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; - virtual std::unique_ptr<IWorkload> CreateActivation(const ActivationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateAddition(const AdditionQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateSplitter(const SplitterQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateBatchNormalization(const BatchNormalizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateMerger(const MergerQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; - virtual std::unique_ptr<IWorkload> CreateFullyConnected(const FullyConnectedQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateConstant(const ConstantQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreatePermute(const PermuteQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreatePooling2d(const Pooling2dQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor, + const WorkloadInfo& info) const; virtual std::unique_ptr<IWorkload> CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + const WorkloadInfo& info) const; + + virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor, + const WorkloadInfo& info) const; virtual std::unique_ptr<IWorkload> CreateDepthwiseConvolution2d( - const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const = 0; + const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const; virtual std::unique_ptr<IWorkload> CreateDetectionPostProcess( - const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info) const = 0; + const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateNormalization(const NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; - - virtual std::unique_ptr<IWorkload> CreateAddition(const AdditionQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateDivision(const DivisionQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateMultiplication(const MultiplicationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateEqual(const EqualQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; - virtual std::unique_ptr<IWorkload> CreateBatchNormalization(const BatchNormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateFakeQuantization(const FakeQuantizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor, - const WorkloadInfo& Info) const = 0; + virtual std::unique_ptr<IWorkload> CreateFloor(const FloorQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateMemCopy(const MemCopyQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateFullyConnected(const FullyConnectedQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateResizeBilinear(const ResizeBilinearQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateGather(const GatherQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateFakeQuantization(const FakeQuantizationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateGreater(const GreaterQueueDescriptor& descriptor, + const WorkloadInfo& info) const; virtual std::unique_ptr<IWorkload> CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; - - virtual std::unique_ptr<IWorkload> CreateConstant(const ConstantQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateReshape(const ReshapeQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateLstm(const LstmQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMaximum(const MaximumQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateFloor(const FloorQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMean(const MeanQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; - virtual std::unique_ptr<IWorkload> CreateLstm(const LstmQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMemCopy(const MemCopyQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMerger(const MergerQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMinimum(const MinimumQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateDivision(const DivisionQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateMultiplication(const MultiplicationQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateSubtraction(const SubtractionQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateNormalization(const NormalizationQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateMaximum(const MaximumQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateOutput(const OutputQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateMean(const MeanQueueDescriptor& descriptor, - const WorkloadInfo& Info) const = 0; + virtual std::unique_ptr<IWorkload> CreatePad(const PadQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; - virtual std::unique_ptr<IWorkload> CreateMinimum(const MinimumQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreatePermute(const PermuteQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreatePad(const PadQueueDescriptor& descriptor, - const WorkloadInfo& Info) const = 0; + virtual std::unique_ptr<IWorkload> CreatePooling2d(const Pooling2dQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateEqual(const EqualQueueDescriptor& descriptor, - const WorkloadInfo& Info) const = 0; + virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor, - const WorkloadInfo& Info) const = 0; + virtual std::unique_ptr<IWorkload> CreateQuantize(const QuantizeQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; - virtual std::unique_ptr<IWorkload> CreateGreater(const GreaterQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateReshape(const ReshapeQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateResizeBilinear(const ResizeBilinearQueueDescriptor& descriptor, + const WorkloadInfo& info) const; virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, + const WorkloadInfo& info) const; - virtual std::unique_ptr<IWorkload> CreateGather(const GatherQueueDescriptor& descriptor, - const WorkloadInfo& info) const = 0; + virtual std::unique_ptr<IWorkload> CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, + const WorkloadInfo& info) const; + + virtual std::unique_ptr<IWorkload> CreateSubtraction(const SubtractionQueueDescriptor& descriptor, + const WorkloadInfo& info) const; + + virtual std::unique_ptr<IWorkload> CreateSplitter(const SplitterQueueDescriptor& descriptor, + const WorkloadInfo& info) const; + + virtual std::unique_ptr<IWorkload> CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor, + const WorkloadInfo& Info) const; }; } //namespace armnn diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp index 79213c18e0..8f86132274 100644 --- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp +++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp @@ -372,6 +372,8 @@ DECLARE_LAYER_POLICY_CUSTOM_PARAM(Output, armnn::LayerBindingId) DECLARE_LAYER_POLICY_2_PARAM(Pad) +DECLARE_LAYER_POLICY_1_PARAM(Quantize) + DECLARE_LAYER_POLICY_2_PARAM(Permute) DECLARE_LAYER_POLICY_2_PARAM(Pooling2d) |