From a9cca6aa935b3e290181a05fdb2c5f5557a49c09 Mon Sep 17 00:00:00 2001 From: Derek Lamberti Date: Mon, 25 Mar 2019 15:41:58 +0000 Subject: IVGCVSW-2870 Support QuantizeLayer on frontend Change-Id: I2014a8d801f1f222d27a80dddf4f188ddcb3a5c9 Signed-off-by: Derek Lamberti --- src/armnn/InternalTypes.hpp | 1 + src/armnn/LayerSupport.cpp | 9 +++ src/armnn/LayersFwd.hpp | 2 + src/armnn/Network.cpp | 5 ++ src/armnn/Network.hpp | 2 + src/armnn/layers/QuantizeLayer.cpp | 48 ++++++++++++ src/armnn/layers/QuantizeLayer.hpp | 35 +++++++++ src/armnn/test/NetworkTests.cpp | 51 +++++++++++++ src/armnn/test/TestLayerVisitor.hpp | 145 +----------------------------------- 9 files changed, 155 insertions(+), 143 deletions(-) create mode 100644 src/armnn/layers/QuantizeLayer.cpp create mode 100644 src/armnn/layers/QuantizeLayer.hpp (limited to 'src/armnn') 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(padDescriptor,name); } +IConnectableLayer *Network::AddQuantizeLayer(const char *name) +{ + return m_Graph->AddLayer(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 + +namespace armnn +{ + +QuantizeLayer::QuantizeLayer(const char* name) +: Layer(1, 1, LayerType::Quantize, name) +{} + +std::unique_ptr QuantizeLayer::CreateWorkload(const Graph& graph, + const IWorkloadFactory& factory) const +{ + return nullptr; +} + +Layer* QuantizeLayer::Clone(Graph& graph) const +{ + QuantizeLayer* clone = CloneBase(graph, GetName()); + return clone; +} + +void QuantizeLayer::ValidateTensorShapesFromInputs() +{ + VerifyLayerConnections(1, CHECK_LOCATION()); + + auto inferredShapes = InferOutputShapes({ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape() }); + + ConditionalThrowIfNotEqual( + "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 + +namespace armnn { + +//Forward +class IWorkload; +class IWorkloadFactory; +class ILayerVisitor; + +class QuantizeLayer : public Layer +{ +public: + virtual std::unique_ptr 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 +#include #include #include @@ -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 + { + 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 +#include #include namespace armnn { // Abstract base class with do nothing implementations for all layer visit methods -class TestLayerVisitor : public ILayerVisitor +class TestLayerVisitor : public LayerVisitorBase { 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& biases, - const char* name = nullptr) override {} - - void VisitDepthwiseConvolution2dLayer(const IConnectableLayer* layer, - const DepthwiseConvolution2dDescriptor& convolution2dDescriptor, - const ConstTensor& weights, - const Optional& 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& 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 -- cgit v1.2.1