From c5fe6e71cd39096af7c2523ec2afe96008c51b0c Mon Sep 17 00:00:00 2001 From: Matthew Sloyan Date: Fri, 25 Nov 2022 16:10:00 +0000 Subject: IVGCVSW-7168 Add Conv2d and Constant support to TOSA Reference Backend * Added TOSA Conv2d and Constant mappings. * Added unique naming to mappings based on previous and following layers, so they are connected correctly. * Updated existing mappings with new naming convention. * Added all mappings to one main block in OptimizeSubgraphView. * Removed isMain from mapping functions. * Added Conv2d EndToEnd test. Signed-off-by: Matthew Sloyan Change-Id: I27c3e238407c32379ce25a1f01dad11523ef5d2b --- src/backends/backendsCommon/test/CMakeLists.txt | 1 + .../test/Convolution2dEndToEndTestImpl.hpp | 134 +++++++++++++++++ src/backends/reference/test/RefEndToEndTests.cpp | 16 ++ src/backends/tosaCommon/TosaLayerSupportRules.hpp | 28 +++- src/backends/tosaCommon/TosaMappings.cpp | 50 +++---- src/backends/tosaCommon/TosaMappings.hpp | 16 +- .../operatorMappings/AdditionOperator.cpp | 50 ++++--- .../operatorMappings/AdditionOperator.hpp | 9 +- .../AvgPool2DIgnoreValueOperator.cpp | 55 +++---- .../AvgPool2DIgnoreValueOperator.hpp | 7 +- .../tosaCommon/operatorMappings/CMakeLists.txt | 4 + .../operatorMappings/ConstantOperator.cpp | 44 ++++++ .../operatorMappings/ConstantOperator.hpp | 19 +++ .../tosaCommon/operatorMappings/Conv2dOperator.cpp | 123 ++++++++++++++++ .../tosaCommon/operatorMappings/Conv2dOperator.hpp | 20 +++ .../operatorMappings/Pooling2DOperator.cpp | 39 ++--- .../operatorMappings/Pooling2DOperator.hpp | 7 +- .../operatorMappings/TosaCommonOperators.hpp | 2 + .../operatorMappings/TosaOperatorUtils.hpp | 101 ++++++++++++- .../test/AvgPool2DIgnoreValueChecker.hpp | 12 +- .../tosaCommon/test/OneToManyMappingTests.cpp | 4 +- .../tosaCommon/test/OneToOneMappingTests.cpp | 136 ++++++++++++++++- src/backends/tosaCommon/test/TosaTestUtils.hpp | 51 ++++++- src/backends/tosaReference/TosaRefBackend.cpp | 47 +++++- src/backends/tosaReference/TosaRefLayerSupport.cpp | 162 +++++++++++++++------ .../tosaReference/test/TosaRefEndToEndTests.cpp | 12 ++ .../test/TosaRefLayerSupportTests.cpp | 101 ++++++++++++- .../workloads/TosaRefPreCompiledWorkload.cpp | 31 ++-- 28 files changed, 1066 insertions(+), 215 deletions(-) create mode 100644 src/backends/backendsCommon/test/Convolution2dEndToEndTestImpl.hpp create mode 100644 src/backends/tosaCommon/operatorMappings/ConstantOperator.cpp create mode 100644 src/backends/tosaCommon/operatorMappings/ConstantOperator.hpp create mode 100644 src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp create mode 100644 src/backends/tosaCommon/operatorMappings/Conv2dOperator.hpp diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt index c9668a22f2..d833caa3fe 100644 --- a/src/backends/backendsCommon/test/CMakeLists.txt +++ b/src/backends/backendsCommon/test/CMakeLists.txt @@ -14,6 +14,7 @@ list(APPEND armnnBackendsCommonUnitTests_sources ComparisonEndToEndTestImpl.hpp CompatibilityTests.cpp ConcatEndToEndTestImpl.hpp + Convolution2dEndToEndTestImpl.hpp Convolution3dEndToEndTestImpl.hpp CustomMemoryOptimizerStrategyTests.cpp DefaultAsyncExecuteTest.cpp diff --git a/src/backends/backendsCommon/test/Convolution2dEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/Convolution2dEndToEndTestImpl.hpp new file mode 100644 index 0000000000..bc9a94289b --- /dev/null +++ b/src/backends/backendsCommon/test/Convolution2dEndToEndTestImpl.hpp @@ -0,0 +1,134 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "EndToEndTestImpl.hpp" +#include + +#include + +#include +#include + +#include +#include + +namespace +{ + +armnn::INetworkPtr CreateConstConvolution2dNetwork(const armnn::Convolution2dDescriptor& descriptor, + const armnn::TensorInfo& inputInfo, + const armnn::TensorInfo& weightsInfo, + const armnn::TensorInfo& biasInfo, + const armnn::TensorInfo& outputInfo, + const armnn::ConstTensor& weights, + const armnn::ConstTensor& biases, + bool biasEnabled) +{ + using namespace armnn; + + INetworkPtr network(INetwork::Create()); + IConnectableLayer* input = network->AddInputLayer(0, "input"); + IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights"); + IConnectableLayer* convolution2d = network->AddConvolution2dLayer(descriptor, "convolution2d"); + IConnectableLayer* output = network->AddOutputLayer(0, "output"); + + Connect(input, convolution2d, inputInfo, 0, 0); + Connect(weightsLayer, convolution2d, weightsInfo, 0, 1); + + if(biasEnabled) + { + armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biases, "Bias"); + Connect(biasLayer, convolution2d, biasInfo, 0, 2); + } + + Connect(convolution2d, output, outputInfo, 0, 0); + + return network; +} + +template> +void Convolution2dEndToEnd(const std::vector& backends, + armnn::DataLayout dataLayout, + bool biasEnabled = true) +{ + using namespace armnn; + + const float qScale = IsQuantizedType() ? 0.25f : 1.0f; + const int32_t qOffset = IsQuantizedType() ? 50 : 0; + + TensorInfo inputInfo({ 1, 5, 5, 1 }, ArmnnType, qScale, qOffset, true); + TensorInfo outputInfo({ 1, 3, 3, 1 }, ArmnnType, qScale, qOffset); + TensorInfo weightsInfo({ 1, 3, 3, 1 }, ArmnnType, qScale, qOffset, true); + TensorInfo biasesInfo({ 1 }, ArmnnType, qScale * qScale, 0, true); + + std::vector inputData = + { + 1.0f, 5.0f, 2.0f, 3.0f, 5.0f, + 8.0f, 7.0f, 3.0f, 6.0f, 3.0f, + 3.0f, 3.0f, 9.0f, 1.0f, 9.0f, + 4.0f, 1.0f, 8.0f, 1.0f, 3.0f, + 6.0f, 8.0f, 1.0f, 9.0f, 2.0f + }; + + std::vector weightsData = + { + 4.0f, 5.0f, 6.0f, + 0.0f, 0.0f, 0.0f, + 3.0f, 2.0f, 1.0f + }; + + std::vector biasesData = { 1.0f }; + + float bias = biasEnabled ? biasesData[0] : 0.0f; + std::vector expectedOutputData = + { + 65.0f + bias, 76.0f + bias, 91.0f + bias, + 107.0f + bias, 99.0f + bias, 89.0f + bias, + 116.0f + bias, 98.0f + bias, 118.0f + bias, + }; + + Convolution2dDescriptor descriptor; + descriptor.m_PadLeft = 0; + descriptor.m_PadRight = 0; + descriptor.m_PadTop = 0; + descriptor.m_PadBottom = 0; + descriptor.m_StrideX = 1; + descriptor.m_StrideY = 1; + descriptor.m_BiasEnabled = biasEnabled; + descriptor.m_DataLayout = dataLayout; + + if (dataLayout == DataLayout::NCHW) + { + PermuteTensorNhwcToNchw(inputInfo, inputData); + PermuteTensorNhwcToNchw(weightsInfo, weightsData); + PermuteTensorNhwcToNchw(outputInfo, expectedOutputData); + } + + // Quantize data + std::vector qInputData = armnnUtils::QuantizedVector(inputData, qScale, qOffset); + std::vector qWeightsData = armnnUtils::QuantizedVector(weightsData, qScale, qOffset); + std::vector qExpectedOutputData = armnnUtils::QuantizedVector(expectedOutputData, qScale, qOffset); + std::vector qBiasesData = armnnUtils::QuantizedVector(biasesData, qScale * qScale, 0); + + ConstTensor weights(weightsInfo, qWeightsData); + ConstTensor biases(biasesInfo, qBiasesData); + + INetworkPtr network = CreateConstConvolution2dNetwork(descriptor, + inputInfo, + weightsInfo, + biasesInfo, + outputInfo, + weights, + biases, + biasEnabled); + + EndToEndLayerTestImpl(std::move(network), + {{ 0, qInputData }}, + {{ 0, qExpectedOutputData }}, + backends); +} + +} // anonymous namespace diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp index 218f6dd695..017330ed53 100644 --- a/src/backends/reference/test/RefEndToEndTests.cpp +++ b/src/backends/reference/test/RefEndToEndTests.cpp @@ -13,6 +13,7 @@ #include #include #include +#include #include #include #include @@ -595,6 +596,21 @@ TEST_CASE("RefConcatEndToEndDim3Uint8Test") ConcatDim3EndToEnd(defaultBackends); } +TEST_CASE("RefConvolution2dFloat32Test") +{ + Convolution2dEndToEnd(defaultBackends, armnn::DataLayout::NHWC); +} + +TEST_CASE("RefConvolution2dNchwFloat32Test") +{ + Convolution2dEndToEnd(defaultBackends, armnn::DataLayout::NCHW); +} + +TEST_CASE("RefConvolution2dFloat16Test") +{ + Convolution2dEndToEnd(defaultBackends, armnn::DataLayout::NHWC); +} + TEST_CASE("RefConvolution3dFloat32Test") { Convolution3dEndToEnd(defaultBackends, diff --git a/src/backends/tosaCommon/TosaLayerSupportRules.hpp b/src/backends/tosaCommon/TosaLayerSupportRules.hpp index 792908c619..8855dd612d 100644 --- a/src/backends/tosaCommon/TosaLayerSupportRules.hpp +++ b/src/backends/tosaCommon/TosaLayerSupportRules.hpp @@ -48,14 +48,34 @@ struct TosaAssertSize : public Rule } }; -struct TosaContainerContains : public Rule +struct TosaContainerContainsTwoTypes : public Rule { - explicit TosaContainerContains(std::tuple& check, const std::vector>& c) + explicit TosaContainerContainsTwoTypes(std::tuple& check, + const std::vector>& c) { for (auto item: c) { - if (std::get<0>(check) == std::get<0>(item) - && std::get<1>(check) == std::get<1>(item)) + if (std::get<0>(check) == std::get<0>(item) && + std::get<1>(check) == std::get<1>(item)) + { + m_Res = true; + return; + } + } + m_Res = false; + } +}; + +struct TosaContainerContainsThreeTypes : public Rule +{ + explicit TosaContainerContainsThreeTypes(std::tuple& check, + const std::vector>& c) + { + for (auto item: c) + { + if (std::get<0>(check) == std::get<0>(item) && + std::get<1>(check) == std::get<1>(item) && + std::get<2>(check) == std::get<2>(item)) { m_Res = true; return; diff --git a/src/backends/tosaCommon/TosaMappings.cpp b/src/backends/tosaCommon/TosaMappings.cpp index a37eaf29b3..00ba429555 100644 --- a/src/backends/tosaCommon/TosaMappings.cpp +++ b/src/backends/tosaCommon/TosaMappings.cpp @@ -8,40 +8,33 @@ using namespace armnn; using namespace tosa; -void SetBasicBlockConstantTensorData(Layer* layer, TosaSerializationBasicBlock* /*basicBlock*/) -{ - switch (layer->GetType()) - { - case LayerType::Convolution2d: - { - // ToDo: using Convolution2d as an example as it has constant tensors for weights and bias. - // ToDo: manually set TosaOperator data of basicBlock where constant tensors exist. - } - default: - // If no switch statement for layer, no constant tensors exist in that layer, return - return; - } -} - TosaSerializationBasicBlock* CreateEmptyTosaSerializationBasicBlock() { - // empty basic block when no tosa mapping implemented/exists - TosaSerializationOperator* op = - new TosaSerializationOperator(Op_UNKNOWN, Attribute_NONE, nullptr, {}, {}); + // Empty basic block when no TOSA mapping implemented/exists + auto* op = new TosaSerializationOperator(Op_UNKNOWN, Attribute_NONE, nullptr, {}, {}); return new TosaSerializationBasicBlock("", {op}, {}, {}, {}); } -TosaSerializationBasicBlock* GetTosaMapping(const LayerType type, +TosaSerializationBasicBlock* GetTosaMapping(const Layer* layer, + const LayerType type, const std::vector& inputs, const std::vector& outputs, - const BaseDescriptor& descriptor, - bool isMain = false) + const BaseDescriptor& descriptor) { switch (type) { case LayerType::Addition: { - return ConvertAdditionToTosaOperator(inputs, outputs, isMain); + return ConvertAdditionToTosaOperator(layer, inputs, outputs); + } + case LayerType::Constant: + { + return ConvertConstantToTosaOperator(layer, outputs); + } + case LayerType::Convolution2d: + { + auto conv2dDesc = PolymorphicDowncast(&descriptor); + return ConvertConv2dToTosaOperator(layer, inputs, outputs, conv2dDesc); } case LayerType::Pooling2d: { @@ -57,11 +50,11 @@ TosaSerializationBasicBlock* GetTosaMapping(const LayerType type, } else if (avgPoolIgnoreValue) { - return ConvertAvgPool2DIgnoreValueToTosaOperator(inputs, outputs, isMain, poolDesc); + return ConvertAvgPool2DIgnoreValueToTosaOperator(layer, inputs, outputs, poolDesc); } else { - return ConvertPooling2DToTosaOperator(inputs, outputs, isMain, poolDesc); + return ConvertPooling2DToTosaOperator(layer, inputs, outputs, poolDesc); } } default: @@ -71,7 +64,7 @@ TosaSerializationBasicBlock* GetTosaMapping(const LayerType type, } } -TosaSerializationBasicBlock* GetTosaMappingFromLayer(Layer* layer, bool isMain = false) +TosaSerializationBasicBlock* GetTosaMappingFromLayer(Layer* layer) { std::vector inputs; for (auto inputSlot : layer->GetInputSlots()) @@ -85,11 +78,10 @@ TosaSerializationBasicBlock* GetTosaMappingFromLayer(Layer* layer, bool isMain = outputs.push_back(&outputSlot.GetTensorInfo()); } - TosaSerializationBasicBlock* basicBlock = GetTosaMapping(layer->GetType(), + TosaSerializationBasicBlock* basicBlock = GetTosaMapping(layer, + layer->GetType(), inputs, outputs, - layer->GetParameters(), - isMain); - SetBasicBlockConstantTensorData(layer, basicBlock); + layer->GetParameters()); return basicBlock; } diff --git a/src/backends/tosaCommon/TosaMappings.hpp b/src/backends/tosaCommon/TosaMappings.hpp index 8df2493d6e..cc41f1b7c8 100644 --- a/src/backends/tosaCommon/TosaMappings.hpp +++ b/src/backends/tosaCommon/TosaMappings.hpp @@ -13,22 +13,18 @@ using namespace armnn; using namespace tosa; -// From the input armnn::Layer, set the corresponding data field in the -// tosa::TosaSerializationTensor where constant tensor data exists in the armnn::Layer. -void SetBasicBlockConstantTensorData(Layer* layer, TosaSerializationBasicBlock* /*basicBlock*/); - // Populates a tosa::TosaSerializationBasicBlock from constructing // tosa::TosaSerializationOperator(s) and tosa::TosaSerializationTensor(s) // based on the input armnn::LayerType and associated armnn::TensorInfos and armnn::Descriptor. // -// If an armnn::LayerType does not have a tosa mapping or the mapping is not implemented in ArmNN, +// If an armnn::LayerType does not have a TOSA mapping or the mapping is not implemented in ArmNN, // an empty tosa::TosaSerializationBasicBlock() is returned with operator tosa::Op_UNKNOWN. -TosaSerializationBasicBlock* GetTosaMapping(const LayerType type, +TosaSerializationBasicBlock* GetTosaMapping(const Layer* layer, + const LayerType type, const std::vector& inputs, const std::vector& outputs, - const BaseDescriptor& /*descriptor*/, - bool isMain); + const BaseDescriptor& /*descriptor*/); // Function called in armnn::OptimizeSubgraphView() when access to armnn::Layer is available -// and there is an option to set tosa basic block data from constant layer tenors available from the input layer. -TosaSerializationBasicBlock* GetTosaMappingFromLayer(Layer* layer, bool isMain); +// and there is an option to set TOSA basic block data from constant layer tensors available from the input layer. +TosaSerializationBasicBlock* GetTosaMappingFromLayer(Layer* layer); diff --git a/src/backends/tosaCommon/operatorMappings/AdditionOperator.cpp b/src/backends/tosaCommon/operatorMappings/AdditionOperator.cpp index 796797728e..66ca869ac4 100644 --- a/src/backends/tosaCommon/operatorMappings/AdditionOperator.cpp +++ b/src/backends/tosaCommon/operatorMappings/AdditionOperator.cpp @@ -5,28 +5,36 @@ #include "AdditionOperator.hpp" -TosaSerializationBasicBlock* ConvertAdditionToTosaOperator(const std::vector& inputs, - const std::vector& outputs, - bool isMain) +TosaSerializationBasicBlock* ConvertAdditionToTosaOperator(const Layer* layer, + const std::vector& inputs, + const std::vector& outputs) { - // A helper function with static global variables ensures uniqueness - // for dynamically generating input, output and block names - std::string input0Name = std::string("Op_ADD_input0_") + GetUniqueTosaMappingID(); - std::string input1Name = std::string("Op_ADD_input1_") + GetUniqueTosaMappingID(); - std::string outputName = std::string("Op_ADD_output0_") + GetUniqueTosaMappingID(); - std::string blockName = std::string("Op_ADD_block_") + GetUniqueTosaMappingID(); - - // If it's the first block, overwrite block name with main. - if (isMain) + std::string input0Name = std::string("input0_"); + std::string input1Name = std::string("input1_"); + std::string outputName = std::string("output0_"); + std::string blockName = std::string("Op_ADD_block_") + GetUniqueTosaMappingID(); + + // If a layer is present then the block will be used for execution, so input and output names need to be determined + // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter. + if(layer != nullptr) { - blockName = std::string("main"); + // Get the layers connected to the input slots and determine unique layer names. + Layer& connectedLayer0 = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer(); + input0Name = GenerateUniqueName(connectedLayer0, 0); + + Layer& connectedLayer1 = layer->GetInputSlot(1).GetConnectedOutputSlot()->GetOwningLayer(); + input1Name = GenerateUniqueName(connectedLayer1, 1); + + // Get the layer connected to the output slot and determine unique layer name. + Layer& connectedOutputLayer = layer->GetOutputSlot().GetConnection(0)->GetOwningLayer(); + outputName = GenerateUniqueName(connectedOutputLayer, 0); } - TosaSerializationOperator* op = new TosaSerializationOperator(Op_ADD, - Attribute_NONE, - nullptr, - {input0Name, input1Name}, - {outputName}); + auto* op = new TosaSerializationOperator(Op_ADD, + Attribute_NONE, + nullptr, + {input0Name, input1Name}, + {outputName}); std::vector inputShape0 = GetTosaTensorShape(inputs[0]->GetShape()); DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType()); @@ -37,9 +45,9 @@ TosaSerializationBasicBlock* ConvertAdditionToTosaOperator(const std::vector outputShape0 = GetTosaTensorShape(outputs[0]->GetShape()); DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType()); - TosaSerializationTensor* inputTensor0 = new TosaSerializationTensor(input0Name, inputShape0, inputDType0, {}); - TosaSerializationTensor* inputTensor1 = new TosaSerializationTensor(input1Name, inputShape1, inputDType1, {}); - TosaSerializationTensor* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}); + auto* inputTensor0 = new TosaSerializationTensor(input0Name, inputShape0, inputDType0, {}); + auto* inputTensor1 = new TosaSerializationTensor(input1Name, inputShape1, inputDType1, {}); + auto* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}); // operatorInputNames/operatorOutputNames ends up being the same as // blockInputNames/blockOutputNames for one-to-one ArmNN to Tosa mappings diff --git a/src/backends/tosaCommon/operatorMappings/AdditionOperator.hpp b/src/backends/tosaCommon/operatorMappings/AdditionOperator.hpp index f467bb6d10..5eb7441531 100644 --- a/src/backends/tosaCommon/operatorMappings/AdditionOperator.hpp +++ b/src/backends/tosaCommon/operatorMappings/AdditionOperator.hpp @@ -5,15 +5,16 @@ #pragma once +#include "TosaOperatorUtils.hpp" + #include #include -#include "TosaOperatorUtils.hpp" using namespace armnn; using namespace tosa; -TosaSerializationBasicBlock* ConvertAdditionToTosaOperator(const std::vector& inputs, - const std::vector& outputs, - bool isMain); +TosaSerializationBasicBlock* ConvertAdditionToTosaOperator(const Layer* layer, + const std::vector& inputs, + const std::vector& outputs); diff --git a/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.cpp b/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.cpp index b3d2687c30..2601a6243d 100644 --- a/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.cpp +++ b/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.cpp @@ -5,23 +5,27 @@ #include "Pooling2DOperator.hpp" -TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const std::vector& inputs, +TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const Layer* layer, + const std::vector& inputs, const std::vector& outputs, - bool isMain, const Pooling2dDescriptor* poolDescriptor) { + std::string padInputName = std::string("input0_"); + std::string padOutputName = std::string("intermediate0_") + GetUniqueTosaMappingID(); + std::string poolOutputName = std::string("output0_"); + std::string blockName = std::string("Op_AVG_POOL2D_block_") + GetUniqueTosaMappingID(); - // A helper function with static global variables ensures uniqueness - // for dynamically generating input, output and block names - std::string padInputName = std::string("Op_PAD_input0_") + GetUniqueTosaMappingID(); - std::string padOutputName = std::string("Op_PAD_intermediate0_") + GetUniqueTosaMappingID(); - std::string poolOutputName = std::string("Op_AVG_POOL2D_output0_") + GetUniqueTosaMappingID(); - std::string blockName = std::string("Op_AVG_POOL2D_block_") + GetUniqueTosaMappingID(); - - // If it's the first block, overwrite block name with main. - if (isMain) + // If a layer is present then the block will be used for execution, so input and output names need to be determined + // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter. + if(layer != nullptr) { - blockName = std::string("main"); + // Get the layers connected to the input slots and determine unique layer names. + Layer& connectedInputLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer(); + padInputName = GenerateUniqueName(connectedInputLayer, 0); + + // Get the layer connected to the output slot and determine unique layer name. + Layer& connectedOutputLayer = layer->GetOutputSlot().GetConnection(0)->GetOwningLayer(); + poolOutputName = GenerateUniqueName(connectedOutputLayer, 0); } std::vector paddings; @@ -51,11 +55,11 @@ TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const std } TosaPadAttribute padAttribute(paddings, 0, 0.0f); - TosaSerializationOperator* opPad = new TosaSerializationOperator(Op_PAD, - Attribute_PadAttribute, - &padAttribute, - {padInputName}, - {padOutputName}); + auto* opPad = new TosaSerializationOperator(Op_PAD, + Attribute_PadAttribute, + &padAttribute, + {padInputName}, + {padOutputName}); std::vector pad = {0, 0, 0, 0}; std::vector kernel = {static_cast(poolDescriptor->m_PoolHeight), @@ -64,11 +68,11 @@ TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const std static_cast(poolDescriptor->m_StrideX)}; TosaPoolAttribute poolAttribute(pad, kernel, stride, 0, 0, ArmNNToDType(inputs[0]->GetDataType())); - TosaSerializationOperator* opPool = new TosaSerializationOperator(Op_AVG_POOL2D, - Attribute_PoolAttribute, - &poolAttribute, - {padOutputName}, - {poolOutputName}); + auto* opPool = new TosaSerializationOperator(Op_AVG_POOL2D, + Attribute_PoolAttribute, + &poolAttribute, + {padOutputName}, + {poolOutputName}); std::vector inputShape = GetTosaTensorShape(inputs[0]->GetShape()); DType inputDType = ArmNNToDType(inputs[0]->GetDataType()); @@ -92,10 +96,9 @@ TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const std inputShape[3] + paddings[6] + paddings[7]}; } - TosaSerializationTensor* inputTensor = new TosaSerializationTensor(padInputName, inputShape, inputDType, {}); - TosaSerializationTensor* intermediateTensor = new TosaSerializationTensor( - padOutputName, intermediateShape, inputDType, {}); - TosaSerializationTensor* outputTensor = new TosaSerializationTensor(poolOutputName, outputShape, outputDType, {}); + auto* inputTensor = new TosaSerializationTensor(padInputName, inputShape, inputDType, {}); + auto* intermediateTensor = new TosaSerializationTensor(padOutputName, intermediateShape, inputDType, {}); + auto* outputTensor = new TosaSerializationTensor(poolOutputName, outputShape, outputDType, {}); // operatorInputNames/operatorOutputNames ends up being the same as // blockInputNames/blockOutputNames for one-to-one ArmNN to Tosa mappings diff --git a/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.hpp b/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.hpp index 63ae190cc9..f9d09754b0 100644 --- a/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.hpp +++ b/src/backends/tosaCommon/operatorMappings/AvgPool2DIgnoreValueOperator.hpp @@ -5,15 +5,16 @@ #pragma once +#include "TosaOperatorUtils.hpp" + #include #include -#include "TosaOperatorUtils.hpp" using namespace armnn; using namespace tosa; -TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const std::vector& inputs, +TosaSerializationBasicBlock* ConvertAvgPool2DIgnoreValueToTosaOperator(const Layer* layer, + const std::vector& inputs, const std::vector& outputs, - bool isMain, const Pooling2dDescriptor* poolDescriptor); diff --git a/src/backends/tosaCommon/operatorMappings/CMakeLists.txt b/src/backends/tosaCommon/operatorMappings/CMakeLists.txt index 7fdc9226af..b256eddda1 100644 --- a/src/backends/tosaCommon/operatorMappings/CMakeLists.txt +++ b/src/backends/tosaCommon/operatorMappings/CMakeLists.txt @@ -8,6 +8,10 @@ list(APPEND armnnTosaBackendOperators_sources AdditionOperator.cpp AvgPool2DIgnoreValueOperator.hpp AvgPool2DIgnoreValueOperator.cpp + ConstantOperator.hpp + ConstantOperator.cpp + Conv2dOperator.hpp + Conv2dOperator.cpp Pooling2DOperator.hpp Pooling2DOperator.cpp TosaOperatorUtils.hpp diff --git a/src/backends/tosaCommon/operatorMappings/ConstantOperator.cpp b/src/backends/tosaCommon/operatorMappings/ConstantOperator.cpp new file mode 100644 index 0000000000..5e3973f8ec --- /dev/null +++ b/src/backends/tosaCommon/operatorMappings/ConstantOperator.cpp @@ -0,0 +1,44 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "ConstantOperator.hpp" + +#include + +TosaSerializationBasicBlock* ConvertConstantToTosaOperator(const Layer* layer, + const std::vector& outputs) +{ + std::string outputName = std::string("constant_"); + std::string blockName = std::string("Op_CONST_block_") + GetUniqueTosaMappingID(); + + std::vector uint8Data; + + // If a layer is present then the block will be used for execution, so names need to be unique. + // Also, set constant tensor data. + if(layer != nullptr) + { + outputName.append(std::to_string(layer->GetGuid())); + blockName.append(std::to_string(layer->GetGuid())); + + auto constantLayer = PolymorphicDowncast(layer); + auto tensorInfo = constantLayer->GetOutputSlot().GetTensorInfo(); + + uint8Data = ConvertConstantTensorDataToBuffer(constantLayer->m_LayerOutput); + } + + auto* op = new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {outputName}); + + std::vector outputShape0 = GetTosaTensorShape(outputs[0]->GetShape()); + DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType()); + + // Setup output tensor with constant tensor data if available. + auto* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, uint8Data); + + return new TosaSerializationBasicBlock(blockName, // name + {op}, // operators + {outputTensor0}, // tensors + {}, // inputs + {outputName}); // outputs +} \ No newline at end of file diff --git a/src/backends/tosaCommon/operatorMappings/ConstantOperator.hpp b/src/backends/tosaCommon/operatorMappings/ConstantOperator.hpp new file mode 100644 index 0000000000..df158aca3d --- /dev/null +++ b/src/backends/tosaCommon/operatorMappings/ConstantOperator.hpp @@ -0,0 +1,19 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "TosaOperatorUtils.hpp" + +#include + +#include + +using namespace armnn; +using namespace tosa; + +TosaSerializationBasicBlock* ConvertConstantToTosaOperator(const Layer* layer, + const std::vector& outputs); + diff --git a/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp new file mode 100644 index 0000000000..9c095d627f --- /dev/null +++ b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp @@ -0,0 +1,123 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "Conv2dOperator.hpp" + +TosaSerializationBasicBlock* ConvertConv2dToTosaOperator(const Layer* layer, + const std::vector& inputs, + const std::vector& outputs, + const Convolution2dDescriptor* conv2dDescriptor) +{ + std::vector inputNames; + std::string outputName = std::string("output0_"); + std::string blockName = std::string("Op_CONV2D_block_") + GetUniqueTosaMappingID(); + + // Set input names for validation purposes only. + if(layer == nullptr) + { + inputNames.emplace_back("input0_"); + inputNames.emplace_back("input1_"); + if(conv2dDescriptor->m_BiasEnabled) + { + inputNames.emplace_back("input2_"); + } + } + else + { + // If a layer is present then the block will be used for execution, so input and output names need to be + // determined using the previous and following layers so the graph is connected correctly. + // For validation this doesn't matter. + for (uint32_t i = 0; i < inputs.size(); ++i) + { + // Get the layer connected to the input slot and determine unique layer name. + Layer& connectedLayer = layer->GetInputSlot(i).GetConnectedOutputSlot()->GetOwningLayer(); + + std::string inputName = GenerateUniqueName(connectedLayer, i); + inputNames.push_back(inputName); + } + + // Get the layer connected to the output slot and determine unique layer name. + Layer& connectedLayer = layer->GetOutputSlot().GetConnection(0)->GetOwningLayer(); + + outputName = GenerateUniqueName(connectedLayer, 0); + } + + std::vector tensors; + std::vector operators; + + // Setup input Tensor + std::vector inputShape0 = GetTosaTensorShape(inputs[0]->GetShape()); + DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType()); + + tensors.push_back(new TosaSerializationTensor(inputNames[0], inputShape0, inputDType0, {})); + + // Only add input tensors if weights and bias are not constant or if running validation. + // Constant tensors will be created in the ConvertConstantToTosaOperator function. + if(!inputs[1]->IsConstant() || layer == nullptr) + { + std::vector inputShape1 = GetTosaTensorShape(inputs[1]->GetShape()); + DType inputDType1 = ArmNNToDType(inputs[1]->GetDataType()); + + tensors.push_back(new TosaSerializationTensor(inputNames[1], inputShape1, inputDType1, {})); + } + + if(conv2dDescriptor->m_BiasEnabled) + { + if(!inputs[2]->IsConstant() || layer == nullptr) + { + std::vector inputShape2 = GetTosaTensorShape(inputs[2]->GetShape()); + DType inputDType2 = ArmNNToDType(inputs[2]->GetDataType()); + + tensors.push_back(new TosaSerializationTensor(inputNames[2], inputShape2, inputDType2, {})); + } + } + else + { + // If bias is disabled, create a constant bias of 0 as three inputs are required. + std::string constantName = std::string("constant_") + GetUniqueTosaMappingID(); + + operators.push_back(new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {constantName})); + + std::vector uint8Data; + std::vector data = { 0.0 }; + + TosaSerializationHandler::ConvertF32toU8(data, uint8Data); + + tensors.push_back(new TosaSerializationTensor(constantName, {1}, DType_FP32, uint8Data)); + inputNames.emplace_back(constantName); + } + + // Setup Output Tensor + std::vector outputShape0 = GetTosaTensorShape(outputs[0]->GetShape()); + DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType()); + + tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {})); + + // Set up CONV2D operator + std::vector pad = {static_cast(conv2dDescriptor->m_PadTop), + static_cast(conv2dDescriptor->m_PadBottom), + static_cast(conv2dDescriptor->m_PadLeft), + static_cast(conv2dDescriptor->m_PadRight)}; + std::vector stride = {static_cast(conv2dDescriptor->m_StrideY), + static_cast(conv2dDescriptor->m_StrideX)}; + std::vector dilation = {static_cast(conv2dDescriptor->m_DilationY), + static_cast(conv2dDescriptor->m_DilationX)}; + TosaConvAttribute attribute(pad, dilation, stride, 0, 0, ArmNNToDType(inputs[0]->GetDataType())); + + auto* op = new TosaSerializationOperator(Op_CONV2D, + Attribute_ConvAttribute, + &attribute, + inputNames, + {outputName}); + operators.push_back(op); + + // operatorInputNames/operatorOutputNames ends up being the same as + // blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings + return new TosaSerializationBasicBlock(blockName, // name + operators, // operators + tensors, // tensors + inputNames, // inputs + {outputName}); // outputs +} \ No newline at end of file diff --git a/src/backends/tosaCommon/operatorMappings/Conv2dOperator.hpp b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.hpp new file mode 100644 index 0000000000..909151b9ac --- /dev/null +++ b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.hpp @@ -0,0 +1,20 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "TosaOperatorUtils.hpp" + +#include + +#include + +using namespace armnn; +using namespace tosa; + +TosaSerializationBasicBlock* ConvertConv2dToTosaOperator(const Layer* layer, + const std::vector& inputs, + const std::vector& outputs, + const Convolution2dDescriptor* conv2dDescriptor); diff --git a/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.cpp b/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.cpp index cd707edb3a..eaeb8a4cde 100644 --- a/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.cpp +++ b/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.cpp @@ -5,24 +5,29 @@ #include "Pooling2DOperator.hpp" -TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const std::vector& inputs, +TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const Layer* layer, + const std::vector& inputs, const std::vector& outputs, - bool isMain, const Pooling2dDescriptor* poolDescriptor) { std::string poolType = (poolDescriptor->m_PoolType == PoolingAlgorithm::Max) ? "Op_MAX" : "Op_AVG"; Op opcode = (poolDescriptor->m_PoolType == PoolingAlgorithm::Max) ? Op_MAX_POOL2D : Op_AVG_POOL2D; - // A helper function with static global variables ensures uniqueness - // for dynamically generating input, output and block names - std::string input0Name = poolType + std::string("_POOL2D_input0_") + GetUniqueTosaMappingID(); - std::string outputName = poolType + std::string("_POOL2D_output0_") + GetUniqueTosaMappingID(); - std::string blockName = poolType + std::string("_POOL2D_block_") + GetUniqueTosaMappingID(); + std::string input0Name = std::string("input0_"); + std::string outputName = std::string("output0_"); + std::string blockName = std::string("Op_") + poolType + std::string("_POOL2D_block_") + GetUniqueTosaMappingID(); - // If it's the first block, overwrite block name with main. - if (isMain) + // If a layer is present then the block will be used for execution, so input and output names need to be determined + // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter. + if(layer != nullptr) { - blockName = std::string("main"); + // Get the layers connected to the input slots and determine unique layer names. + Layer& connectedInputLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer(); + input0Name = GenerateUniqueName(connectedInputLayer, 0); + + // Get the layer connected to the output slot and determine unique layer name. + Layer& connectedOutputLayer = layer->GetOutputSlot().GetConnection(0)->GetOwningLayer(); + outputName = GenerateUniqueName(connectedOutputLayer, 0); } std::vector pad = {static_cast(poolDescriptor->m_PadTop), @@ -35,11 +40,11 @@ TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const std::vector(poolDescriptor->m_StrideX)}; TosaPoolAttribute attribute(pad, kernel, stride, 0, 0, ArmNNToDType(inputs[0]->GetDataType())); - TosaSerializationOperator* op = new TosaSerializationOperator(opcode, - Attribute_PoolAttribute, - &attribute, - {input0Name}, - {outputName}); + auto* op = new TosaSerializationOperator(opcode, + Attribute_PoolAttribute, + &attribute, + {input0Name}, + {outputName}); std::vector inputShape0 = GetTosaTensorShape(inputs[0]->GetShape()); DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType()); @@ -47,8 +52,8 @@ TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const std::vector outputShape0 = GetTosaTensorShape(outputs[0]->GetShape()); DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType()); - TosaSerializationTensor* inputTensor0 = new TosaSerializationTensor(input0Name, inputShape0, inputDType0, {}); - TosaSerializationTensor* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}); + auto* inputTensor0 = new TosaSerializationTensor(input0Name, inputShape0, inputDType0, {}); + auto* outputTensor0 = new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}); // operatorInputNames/operatorOutputNames ends up being the same as // blockInputNames/blockOutputNames for one-to-one ArmNN to Tosa mappings diff --git a/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.hpp b/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.hpp index 22d2a3ae29..cc9ec097f9 100644 --- a/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.hpp +++ b/src/backends/tosaCommon/operatorMappings/Pooling2DOperator.hpp @@ -5,15 +5,16 @@ #pragma once +#include "TosaOperatorUtils.hpp" + #include #include -#include "TosaOperatorUtils.hpp" using namespace armnn; using namespace tosa; -TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const std::vector& inputs, +TosaSerializationBasicBlock* ConvertPooling2DToTosaOperator(const Layer* layer, + const std::vector& inputs, const std::vector& outputs, - bool isMain, const Pooling2dDescriptor* poolDescriptor); diff --git a/src/backends/tosaCommon/operatorMappings/TosaCommonOperators.hpp b/src/backends/tosaCommon/operatorMappings/TosaCommonOperators.hpp index 00b5f0fa68..513db0c039 100644 --- a/src/backends/tosaCommon/operatorMappings/TosaCommonOperators.hpp +++ b/src/backends/tosaCommon/operatorMappings/TosaCommonOperators.hpp @@ -6,5 +6,7 @@ #pragma once #include "AdditionOperator.hpp" +#include "ConstantOperator.hpp" +#include "Conv2dOperator.hpp" #include "AvgPool2DIgnoreValueOperator.hpp" #include "Pooling2DOperator.hpp" \ No newline at end of file diff --git a/src/backends/tosaCommon/operatorMappings/TosaOperatorUtils.hpp b/src/backends/tosaCommon/operatorMappings/TosaOperatorUtils.hpp index f51b2109b4..176e4e1cfb 100644 --- a/src/backends/tosaCommon/operatorMappings/TosaOperatorUtils.hpp +++ b/src/backends/tosaCommon/operatorMappings/TosaOperatorUtils.hpp @@ -5,10 +5,13 @@ #pragma once +#include #include #include -#include +#include "common/include/ProfilingGuid.hpp" + +#include using namespace armnn; using namespace tosa; @@ -53,6 +56,33 @@ inline std::vector GetTosaTensorShape(const TensorShape& shape) return returnShape; } +// Function that generates unique name using the layer type, input slot and layer guid. +inline std::string GenerateUniqueName(const Layer& layer, uint32_t layerSlot) +{ + std::string name; + std::string guid = std::to_string(layer.GetGuid()); + std::string slotAndGuid = std::to_string(layerSlot) + "_" + guid; + LayerType layerType = layer.GetType(); + + if (layerType == LayerType::Input) + { + name = "input" + slotAndGuid; + } + else if (layerType == LayerType::Output) + { + name = "output" + slotAndGuid; + } + else if (layerType == LayerType::Constant) + { + name = "constant_" + guid; + } + else + { + name = "intermediate" + slotAndGuid; + } + return name; +} + // Function to return unique int as a string to ensure uniqueness between all input, output and block names. static int uniqueTosaMappingID = 0; inline std::string GetUniqueTosaMappingID() @@ -206,3 +236,72 @@ inline std::string TosaOpToString(Op tosaOp) } return ""; } + +inline std::vector ConvertConstantTensorDataToBuffer(const std::shared_ptr& tensorHandle) +{ + tosa_err_t error; + std::vector uint8Data; + auto tensorInfo = tensorHandle->GetTensorInfo(); + + switch (tensorInfo.GetDataType()) + { + case DataType::Float32: + { + std::vector data(tensorInfo.GetNumElements()); + memcpy(data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + + error = TosaSerializationHandler::ConvertF32toU8(data, uint8Data); + break; + } + case DataType::Float16: + { + std::vector data(tensorInfo.GetNumElements()); + memcpy(data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + + error = TosaSerializationHandler::ConvertF16toU8(data, uint8Data); + break; + } + case DataType::QSymmS8: + case DataType::QAsymmS8: + { + std::vector data(tensorInfo.GetNumElements()); + memcpy(data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + + error = TosaSerializationHandler::ConvertI8toU8(data, uint8Data); + break; + } + case DataType::QAsymmU8: + { + memcpy(uint8Data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + break; + } + case DataType::QSymmS16: + { + std::vector data(tensorInfo.GetNumElements()); + memcpy(data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + + error = TosaSerializationHandler::ConvertI16toU8(data, uint8Data); + break; + } + case DataType::Signed32: + { + std::vector data(tensorInfo.GetNumElements()); + memcpy(data.data(), tensorHandle->Map(true), tensorInfo.GetNumBytes()); + + error = TosaSerializationHandler::ConvertI32toU8(data, uint8Data); + break; + } + default: + { + throw armnn::Exception("SetConstantTensorData: An unsupported data type was encountered."); + } + } + + if(error != tosa_err_t::TOSA_OK) + { + throw armnn::Exception("SetConstantTensorData: An error occurred when converting constant data"); + } + + tensorHandle->Unmap(); + return uint8Data; +} diff --git a/src/backends/tosaCommon/test/AvgPool2DIgnoreValueChecker.hpp b/src/backends/tosaCommon/test/AvgPool2DIgnoreValueChecker.hpp index 8869b3a8ff..a38f66b466 100644 --- a/src/backends/tosaCommon/test/AvgPool2DIgnoreValueChecker.hpp +++ b/src/backends/tosaCommon/test/AvgPool2DIgnoreValueChecker.hpp @@ -17,10 +17,8 @@ void VerifyAvgPool2DIgnoreValue(TosaSerializationBasicBlock* basicBlock, { uint32_t numInputs = static_cast(inputShape.size()); uint32_t numOutputs = static_cast(outputShape.size()); - std::string operatorString0 = TosaOpToString(Op_PAD); - std::string operatorString1 = TosaOpToString(Op_AVG_POOL2D); - std::string blockStr = operatorString1 + "_block_"; + std::string blockStr = TosaOpToString(Op_AVG_POOL2D) + "_block_"; CHECK(basicBlock->GetName().find(blockStr) != std::string::npos); CHECK(basicBlock->GetInputs().size() == numInputs); CHECK(basicBlock->GetOutputs().size() == numOutputs); @@ -41,7 +39,7 @@ void VerifyAvgPool2DIgnoreValue(TosaSerializationBasicBlock* basicBlock, std::basic_string blockInputName = basicBlock->GetInputs()[i]; std::basic_string operatorInputName = padOp->GetInputTensorNames()[i]; - std::string opStr = operatorString0 + "_input" + std::to_string(i) + "_"; + std::string opStr = "input" + std::to_string(i) + "_"; CHECK(blockInputName == operatorInputName); CHECK(basicBlock->GetTensorByName(blockInputName)); @@ -56,7 +54,7 @@ void VerifyAvgPool2DIgnoreValue(TosaSerializationBasicBlock* basicBlock, for (uint32_t i = 0; i < padOpOutputs; i++) { std::basic_string operatorOutputName = padOp->GetOutputTensorNames()[i]; - std::string opStr = operatorString0 + "_intermediate" + std::to_string(i) + "_"; + std::string opStr = "intermediate" + std::to_string(i) + "_"; CHECK(basicBlock->GetTensorByName(operatorOutputName)); CHECK(operatorOutputName.find(opStr) != std::string::npos); @@ -86,7 +84,7 @@ void VerifyAvgPool2DIgnoreValue(TosaSerializationBasicBlock* basicBlock, for (uint32_t i = 0; i < poolOpInputs; i++) { std::basic_string operatorInputName = poolOp->GetInputTensorNames()[i]; - std::string opStr = operatorString0 + "_intermediate" + std::to_string(i) + "_"; + std::string opStr = "intermediate" + std::to_string(i) + "_"; CHECK(basicBlock->GetTensorByName(operatorInputName)); CHECK(operatorInputName.find(opStr) != std::string::npos); @@ -102,7 +100,7 @@ void VerifyAvgPool2DIgnoreValue(TosaSerializationBasicBlock* basicBlock, std::basic_string blockOutputName = basicBlock->GetOutputs()[i]; std::basic_string operatorOutputName = poolOp->GetOutputTensorNames()[i]; - std::string opStr = operatorString1 + "_output" + std::to_string(i) + "_"; + std::string opStr = "output" + std::to_string(i) + "_"; CHECK(blockOutputName == operatorOutputName); CHECK(basicBlock->GetTensorByName(blockOutputName)); diff --git a/src/backends/tosaCommon/test/OneToManyMappingTests.cpp b/src/backends/tosaCommon/test/OneToManyMappingTests.cpp index 98fd563da1..dd61ba8191 100644 --- a/src/backends/tosaCommon/test/OneToManyMappingTests.cpp +++ b/src/backends/tosaCommon/test/OneToManyMappingTests.cpp @@ -30,7 +30,7 @@ TEST_CASE("GetTosaMapping_AvgPool2DIgnoreValueLayer") std::vector> outputShape = {{ 1, 1, 3, 3 }}; TosaSerializationBasicBlock* basicBlock = - GetTosaMapping(LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor, false); + GetTosaMapping(nullptr, LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor); VerifyAvgPool2DIgnoreValue(basicBlock, inputShape, outputShape, @@ -74,7 +74,7 @@ TEST_CASE("GetTosaMappingFromLayer_AvgPool2DIgnoreValueLayer") pool->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); TosaSerializationBasicBlock* basicBlock = - GetTosaMappingFromLayer(PolymorphicDowncast(pool), false); + GetTosaMappingFromLayer(PolymorphicDowncast(pool)); VerifyAvgPool2DIgnoreValue(basicBlock, inputShape, outputShape, diff --git a/src/backends/tosaCommon/test/OneToOneMappingTests.cpp b/src/backends/tosaCommon/test/OneToOneMappingTests.cpp index 04d1eb46aa..af9f9e26df 100644 --- a/src/backends/tosaCommon/test/OneToOneMappingTests.cpp +++ b/src/backends/tosaCommon/test/OneToOneMappingTests.cpp @@ -4,6 +4,7 @@ // #include "TosaTestUtils.hpp" +#include "CommonTestUtils.hpp" using namespace armnn; using namespace tosa; @@ -18,7 +19,7 @@ TEST_CASE("GetTosaMapping_AdditionLayer") std::vector> outputShape = {{ 1, 2, 4, 2 }}; TosaSerializationBasicBlock* basicBlock = - GetTosaMapping(LayerType::Addition, {&info, &info}, {&info}, BaseDescriptor(), false); + GetTosaMapping(nullptr, LayerType::Addition, {&info, &info}, {&info}, BaseDescriptor()); AssertTosaOneToOneMappingBasicBlock( basicBlock, inputShape, outputShape, Op_ADD, Attribute_NONE, BaseDescriptor(), LayerType::Addition); } @@ -50,11 +51,132 @@ TEST_CASE("GetTosaMappingFromLayer_AdditionLayer") std::vector> outputShape = {{ 1, 2, 4, 2 }}; TosaSerializationBasicBlock* basicBlock = - GetTosaMappingFromLayer(PolymorphicDowncast(add), false); + GetTosaMappingFromLayer(PolymorphicDowncast(add)); AssertTosaOneToOneMappingBasicBlock( basicBlock, inputShape, outputShape, Op_ADD, Attribute_NONE, BaseDescriptor(), LayerType::Addition); } +TEST_CASE("GetTosaMapping_ConstantLayer") +{ + TensorInfo outputInfo = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32, 0.0f, 0, true); + std::vector> outputShape = {{ 1, 2, 4, 2 }}; + + TosaSerializationBasicBlock* basicBlock = + GetTosaMapping(nullptr, LayerType::Constant, {}, {&outputInfo}, BaseDescriptor()); + AssertTosaOneToOneMappingBasicBlock( + basicBlock, {}, outputShape, Op_CONST, Attribute_NONE, BaseDescriptor(), LayerType::Constant); +} + +TEST_CASE("GetTosaMappingFromLayer_ConstantLayer") +{ + IRuntime::CreationOptions options; + IRuntimePtr runtime(IRuntime::Create(options)); + + // Builds up the structure of the network. + INetworkPtr net(INetwork::Create()); + + TensorInfo info = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32, 0.0f, 0, true); + std::vector> outputShape = {{ 1, 2, 4, 2 }}; + + std::vector data = GenerateRandomData(info.GetNumElements()); + armnn::ConstTensor constTensor(info, data); + + IConnectableLayer* constant = net->AddConstantLayer(constTensor, "constant"); + IConnectableLayer* output = net->AddOutputLayer(0, "output"); + + constant->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + constant->GetOutputSlot(0).SetTensorInfo(info); + + TosaSerializationBasicBlock* basicBlock = + GetTosaMappingFromLayer(PolymorphicDowncast(constant)); + AssertTosaOneToOneMappingBasicBlock( + basicBlock, {}, outputShape, Op_CONST, Attribute_NONE, BaseDescriptor(), LayerType::Constant); +} + +TEST_CASE("GetTosaMapping_Conv2dLayer") +{ + armnn::Convolution2dDescriptor descriptor; + descriptor.m_PadLeft = 1; + descriptor.m_PadRight = 1; + descriptor.m_PadTop = 1; + descriptor.m_PadBottom = 1; + descriptor.m_StrideX = 2; + descriptor.m_StrideY = 2; + descriptor.m_DilationX = 2; + descriptor.m_DilationY = 2; + descriptor.m_BiasEnabled = true; + + const armnn::TensorInfo inputInfo ({ 1, 5, 5, 1 }, armnn::DataType::Float32); + const armnn::TensorInfo outputInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32); + const armnn::TensorInfo weightsInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32, 0.0f, 0, true); + const armnn::TensorInfo biasesInfo ({ 1 }, armnn::DataType::Float32, 0.0f, 0, true); + + std::vector> inputShape = {{ 1, 5, 5, 1 }, { 1, 3, 3, 1 }, { 1 }}; + std::vector> outputShape = {{ 1, 3, 3, 1 }}; + + TosaSerializationBasicBlock* basicBlock = GetTosaMapping(nullptr, + LayerType::Convolution2d, + {&inputInfo, &weightsInfo, &biasesInfo}, + {&outputInfo}, + descriptor); + AssertTosaOneToOneMappingBasicBlock( + basicBlock, inputShape, outputShape, Op_CONV2D, Attribute_ConvAttribute, descriptor, LayerType::Convolution2d); +} + +TEST_CASE("GetTosaMappingFromLayer_Conv2dLayer") +{ + IRuntime::CreationOptions options; + IRuntimePtr runtime(IRuntime::Create(options)); + + // Builds up the structure of the network. + INetworkPtr net(INetwork::Create()); + + armnn::Convolution2dDescriptor descriptor; + descriptor.m_PadLeft = 1; + descriptor.m_PadRight = 1; + descriptor.m_PadTop = 1; + descriptor.m_PadBottom = 1; + descriptor.m_StrideX = 2; + descriptor.m_StrideY = 2; + descriptor.m_DilationX = 2; + descriptor.m_DilationY = 2; + descriptor.m_BiasEnabled = true; + + const armnn::TensorInfo inputInfo ({ 1, 5, 5, 1 }, armnn::DataType::Float32); + const armnn::TensorInfo outputInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32); + const armnn::TensorInfo weightsInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32, 0.0f, 0, true); + const armnn::TensorInfo biasesInfo ({ 1 }, armnn::DataType::Float32, 0.0f, 0, true); + + std::vector> inputShape = {{ 1, 5, 5, 1 }}; + std::vector> outputShape = {{ 1, 3, 3, 1 }}; + + std::vector weightsData = GenerateRandomData(weightsInfo.GetNumElements()); + armnn::ConstTensor weights(weightsInfo, weightsData); + + std::vector biasesData = GenerateRandomData(biasesInfo.GetNumElements()); + armnn::ConstTensor biases(biasesInfo, biasesData); + + armnn::IConnectableLayer* const inputLayer = net->AddInputLayer(0, "input0"); + armnn::IConnectableLayer* const weightsLayer = net->AddConstantLayer(weights, "weights"); + armnn::IConnectableLayer* const biasesLayer = net->AddConstantLayer(biases, "biases"); + armnn::IConnectableLayer* const convLayer = net->AddConvolution2dLayer(descriptor, "conv2d"); + armnn::IConnectableLayer* const outputLayer = net->AddOutputLayer(0); + + inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0)); + weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1)); + biasesLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(2)); + convLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0)); + + inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo); + weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo); + biasesLayer->GetOutputSlot(0).SetTensorInfo(biasesInfo); + convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); + + TosaSerializationBasicBlock* basicBlock = GetTosaMappingFromLayer(PolymorphicDowncast(convLayer)); + AssertTosaOneToOneMappingBasicBlock( + basicBlock, inputShape, outputShape, Op_CONV2D, Attribute_ConvAttribute, descriptor, LayerType::Convolution2d); +} + TEST_CASE("GetTosaMapping_MaxPool2DLayer") { armnn::Pooling2dDescriptor descriptor; @@ -74,7 +196,7 @@ TEST_CASE("GetTosaMapping_MaxPool2DLayer") std::vector> outputShape = {{ 1, 1, 3, 3 }}; TosaSerializationBasicBlock* basicBlock = - GetTosaMapping(LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor, false); + GetTosaMapping(nullptr, LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor); AssertTosaOneToOneMappingBasicBlock( basicBlock, inputShape, outputShape, Op_MAX_POOL2D, Attribute_PoolAttribute, descriptor, LayerType::Pooling2d); } @@ -114,7 +236,7 @@ TEST_CASE("GetTosaMappingFromLayer_MaxPool2DLayer") pool->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); TosaSerializationBasicBlock* basicBlock = - GetTosaMappingFromLayer(PolymorphicDowncast(pool), false); + GetTosaMappingFromLayer(PolymorphicDowncast(pool)); AssertTosaOneToOneMappingBasicBlock( basicBlock, inputShape, outputShape, Op_MAX_POOL2D, Attribute_PoolAttribute, descriptor, LayerType::Pooling2d); } @@ -138,7 +260,7 @@ TEST_CASE("GetTosaMapping_AvgPool2DLayer") std::vector> outputShape = {{ 1, 1, 3, 3 }}; TosaSerializationBasicBlock* basicBlock = - GetTosaMapping(LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor, false); + GetTosaMapping(nullptr, LayerType::Pooling2d, {&inputTensorInfo}, {&outputTensorInfo}, descriptor); AssertTosaOneToOneMappingBasicBlock(basicBlock, inputShape, outputShape, @@ -183,7 +305,7 @@ TEST_CASE("GetTosaMappingFromLayer_AvgPool2DLayer") pool->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); TosaSerializationBasicBlock* basicBlock = - GetTosaMappingFromLayer(PolymorphicDowncast(pool), false); + GetTosaMappingFromLayer(PolymorphicDowncast(pool)); AssertTosaOneToOneMappingBasicBlock(basicBlock, inputShape, outputShape, @@ -196,7 +318,7 @@ TEST_CASE("GetTosaMappingFromLayer_AvgPool2DLayer") TEST_CASE("GetTosaMapping_Unimplemented") { TosaSerializationBasicBlock* basicBlock = - GetTosaMapping(LayerType::UnidirectionalSequenceLstm, {}, {}, BaseDescriptor(), false); + GetTosaMapping(nullptr, LayerType::UnidirectionalSequenceLstm, {}, {}, BaseDescriptor()); CHECK(basicBlock->GetName() == ""); CHECK(basicBlock->GetTensors().size() == 0); diff --git a/src/backends/tosaCommon/test/TosaTestUtils.hpp b/src/backends/tosaCommon/test/TosaTestUtils.hpp index a362bde10d..dd63c0efdf 100644 --- a/src/backends/tosaCommon/test/TosaTestUtils.hpp +++ b/src/backends/tosaCommon/test/TosaTestUtils.hpp @@ -21,6 +21,24 @@ inline void VerifyTosaAttributeFromDescriptor(const BaseDescriptor& descriptor, { switch (type) { + case LayerType::Convolution2d: + { + auto conv2dDesc = PolymorphicDowncast(&descriptor); + std::vector pad = {static_cast(conv2dDesc->m_PadTop), + static_cast(conv2dDesc->m_PadBottom), + static_cast(conv2dDesc->m_PadLeft), + static_cast(conv2dDesc->m_PadRight)}; + + std::vector dilation = {static_cast(conv2dDesc->m_DilationY), + static_cast(conv2dDesc->m_DilationX)}; + std::vector stride = {static_cast(conv2dDesc->m_StrideY), + static_cast(conv2dDesc->m_StrideX)}; + TosaConvAttribute convAttribute(attribute); + CHECK(pad == convAttribute.pad()); + CHECK(dilation == convAttribute.dilation()); + CHECK(stride == convAttribute.stride()); + break; + } case LayerType::Pooling2d: { auto poolDesc = PolymorphicDowncast(&descriptor); @@ -80,6 +98,7 @@ inline void VerifyTosaAttributeFromDescriptor(const BaseDescriptor& descriptor, CHECK(pad == poolAttribute.pad()); CHECK(kernel == poolAttribute.kernel()); CHECK(stride == poolAttribute.stride()); + break; } default: break; @@ -97,18 +116,30 @@ inline void AssertTosaOneToOneMappingBasicBlock(TosaSerializationBasicBlock* bas DType dataType = DType_FP32) { uint32_t numInputs = static_cast(inputShape.size()); + uint32_t numInputTensors = static_cast(inputShape.size()); uint32_t numOutputs = static_cast(outputShape.size()); std::string operatorString = TosaOpToString(tosaOp); + // The number of tensors in the block can be different if there are constant layers, as they are created separately. + if(type == LayerType::Convolution2d) + { + numInputTensors = 2; + auto conv2dDesc = PolymorphicDowncast(&descriptor); + if(conv2dDesc->m_BiasEnabled) + { + numInputTensors = 3; + } + } + std::string blockStr = operatorString + "_block_"; CHECK(basicBlock->GetName().find(blockStr) != std::string::npos); - CHECK(basicBlock->GetInputs().size() == numInputs); + CHECK(basicBlock->GetInputs().size() == numInputTensors); CHECK(basicBlock->GetOutputs().size() == numOutputs); CHECK(basicBlock->GetOperators().size() == 1); CHECK(basicBlock->GetTensors().size() == (numInputs + numOutputs)); TosaSerializationOperator* op = basicBlock->GetOperators().at(0); - CHECK(op->GetInputTensorNames().size() == numInputs); + CHECK(op->GetInputTensorNames().size() == numInputTensors); CHECK(op->GetOutputTensorNames().size() == numOutputs); for (uint32_t i = 0; i < numInputs; i++) @@ -117,11 +148,11 @@ inline void AssertTosaOneToOneMappingBasicBlock(TosaSerializationBasicBlock* bas std::basic_string operatorInputName = op->GetInputTensorNames()[i]; std::basic_string tensorName = basicBlock->GetTensors()[i]->GetName(); - std::string opStr = operatorString + "_input" + std::to_string(i) + "_"; + std::string opStr = "input" + std::to_string(i) + "_"; CHECK(blockInputName == operatorInputName); CHECK(tensorName == operatorInputName); - CHECK(blockInputName.find(opStr) != std::string::npos); + CHECK(blockInputName.find(opStr) != std::string::npos); } for (uint32_t i = 0; i < numOutputs; i++) @@ -130,7 +161,11 @@ inline void AssertTosaOneToOneMappingBasicBlock(TosaSerializationBasicBlock* bas std::basic_string operatorOutputName = op->GetOutputTensorNames()[i]; std::basic_string tensorName = basicBlock->GetTensors()[numInputs + i]->GetName(); - std::string opStr = operatorString + "_output" + std::to_string(i) + "_"; + std::string opStr = "output" + std::to_string(i) + "_"; + if (tosaOp == Op_CONST) + { + opStr = "constant_"; + } CHECK(blockOutputName == operatorOutputName); CHECK(tensorName == operatorOutputName); @@ -152,8 +187,12 @@ inline void AssertTosaOneToOneMappingBasicBlock(TosaSerializationBasicBlock* bas { TosaSerializationTensor* tensor = basicBlock->GetTensors()[i + inputShape.size()]; CHECK(tensor->GetDtype() == dataType); - CHECK(tensor->GetData().size() == 0); CHECK(tensor->GetShape() == outputShape[static_cast(i)]); + if (tosaOp != Op_CONST) + { + // Const tensors contain data. + CHECK(tensor->GetData().size() == 0); + } } VerifyTosaAttributeFromDescriptor(descriptor, diff --git a/src/backends/tosaReference/TosaRefBackend.cpp b/src/backends/tosaReference/TosaRefBackend.cpp index e3a516a5f9..554bb10f0f 100644 --- a/src/backends/tosaReference/TosaRefBackend.cpp +++ b/src/backends/tosaReference/TosaRefBackend.cpp @@ -83,16 +83,20 @@ OptimizationViews TosaRefBackend::OptimizeSubgraphView(const SubgraphView& subgr const ModelOptions& modelOptions) const { OptimizationViews optimizationViews(modelOptions); + auto handler = std::make_unique(); - // A main block should only be added once. - bool isMain = true; + std::vector graphInputs; + std::vector graphOutputs; + + std::vector operators; + std::vector tensors; auto it = subgraph.endIConnectable(); while (it != subgraph.beginIConnectable()) { --it; - Layer &base = *(PolymorphicDowncast(*it)); + Layer& base = *(PolymorphicDowncast(*it)); if(base.GetType() == armnn::LayerType::Input || base.GetType() == armnn::LayerType::Output) @@ -100,15 +104,44 @@ OptimizationViews TosaRefBackend::OptimizeSubgraphView(const SubgraphView& subgr continue; } - tosa::TosaSerializationBasicBlock* mappings = GetTosaMappingFromLayer(&base, isMain); - handler.get()->GetBlocks().push_back(mappings); + tosa::TosaSerializationBasicBlock* mappings = GetTosaMappingFromLayer(&base); - if(isMain) + // Loop through inputs to see if there are any graph inputs, if so save them. + // If it's an input to the graph "input" can be found in the string. + for (uint32_t i = 0; i < mappings->GetInputs().size(); i++) { - isMain = false; + std::basic_string blockInputName = mappings->GetInputs()[i]; + + if (blockInputName.find("input") != std::string::npos) + { + graphInputs.push_back(blockInputName); + } } + + // Loop through outputs to see if there are any graph outputs, if so save them. + // If it's an output to the graph "output" can be found in the string. + for (uint32_t i = 0; i < mappings->GetOutputs().size(); i++) + { + std::basic_string blockOutputName = mappings->GetOutputs()[i]; + + if (blockOutputName.find("output") != std::string::npos) + { + graphOutputs.push_back(blockOutputName); + } + } + + auto blockOperators = mappings->GetOperators(); + operators.insert(operators.end(), blockOperators.begin(), blockOperators.end()); + + auto blockTensors = mappings->GetTensors(); + tensors.insert(tensors.end(), blockTensors.begin(), blockTensors.end()); } + // Add all mappings to main block, the TOSA Reference Model requires the full graph to be in one block called main. + auto* block = new TosaSerializationBasicBlock("main", operators, tensors, graphInputs, graphOutputs); + + handler.get()->GetBlocks().push_back(block); + auto compiledBlob = std::make_unique(handler.release(), DeleteAsType); diff --git a/src/backends/tosaReference/TosaRefLayerSupport.cpp b/src/backends/tosaReference/TosaRefLayerSupport.cpp index ce4abbf921..848b7efdce 100644 --- a/src/backends/tosaReference/TosaRefLayerSupport.cpp +++ b/src/backends/tosaReference/TosaRefLayerSupport.cpp @@ -102,7 +102,7 @@ static bool RunTosaLayerChecksInputOutputDataType(TosaSerializationOperator* op, std::tuple mappingType(input->GetDtype(), output->GetDtype()); // Check Dtype from tensor (GetDtype) - supported &= CheckSupportRule(TosaContainerContains(mappingType, supportedMappingTypes), + supported &= CheckSupportRule(TosaContainerContainsTwoTypes(mappingType, supportedMappingTypes), reasonIfUnsupported, std::string("TOSA Reference Operator: " + opString + " for input: " + input->GetName() + " and output: " + output->GetName() + @@ -125,6 +125,58 @@ static bool RunTosaLayerChecksInputOutputDataType(TosaSerializationOperator* op, return supported; } +static bool RunTosaLayerChecksInputWeightsOutputDataType( + TosaSerializationOperator* op, + const std::vector& inputs, + const std::vector& outputs, + const std::vector& supportedAttributes, + const std::vector>& supportedMappingTypes, + Optional reasonIfUnsupported) +{ + bool supported = true; + + std::string opString = TosaOpToString(op->GetOp()); + + // Check Attribute from operator (GetAttribute) + supported &= CheckSupportRule(TosaOperatorAttributeOfAny(op, supportedAttributes), reasonIfUnsupported, + std::string("TOSA Reference Operator: " + opString + + " has an unsupported attribute.").c_str()); + + // Check combination of input, weights and output types. + // Bias is the same as output type, so it is covered. + std::tuple mappingTypes(inputs[0]->GetDtype(), inputs[1]->GetDtype(), outputs[0]->GetDtype()); + + // Check Dtype from tensor (GetDtype) + supported &= CheckSupportRule(TosaContainerContainsThreeTypes(mappingTypes, supportedMappingTypes), + reasonIfUnsupported, + std::string("TOSA Reference Operator: " + opString + " for input 0: " + + inputs[0]->GetName() + ", input 1: " + inputs[1]->GetName() + + " and output: " + outputs[0]->GetName() + + " has an unsupported input data type combination.").c_str()); + + for (auto input : inputs) + { + // Check Shape from tensor (GetShape) + supported &= CheckSupportRule(TosaTensorNumDimensionsWithinBounds(input), + reasonIfUnsupported, + std::string("Tosa Reference Operator: " + opString + " for input: " + + input->GetName() + " exceeds MaxNumOfTensorDimensions.").c_str()); + } + + for (auto output : outputs) + { + // Check Shape from tensor (GetShape) + supported &= CheckSupportRule(TosaTensorNumDimensionsWithinBounds(output), + reasonIfUnsupported, + std::string("Tosa Reference Operator: " + opString + " for output: " + + output->GetName() + " exceeds MaxNumOfTensorDimensions.").c_str()); + } + + return supported; +} + + + static bool IsTosaLayerSupported(TosaSerializationOperator* op, const std::vector& inputs, const std::vector& outputs, @@ -134,10 +186,7 @@ static bool IsTosaLayerSupported(TosaSerializationOperator* op, { case tosa::Op_ADD: { - std::vector supportedAttributes = - { - Attribute_NONE - }; + std::vector supportedAttributes = { Attribute_NONE }; // Only Int32, Fp32 and Fp16 are currently supported by the TOSA Reference Model. std::vector supportedTypes = @@ -148,20 +197,47 @@ static bool IsTosaLayerSupported(TosaSerializationOperator* op, }; // Check the attribute, data types and bounds for inputs and outputs. - return RunTosaLayerChecksSingleDataType(op, - inputs, - outputs, - supportedAttributes, - supportedTypes, - reasonIfUnsupported); + return RunTosaLayerChecksSingleDataType( + op, inputs, outputs, supportedAttributes, supportedTypes, reasonIfUnsupported); } - case tosa::Op_AVG_POOL2D: + case tosa::Op_CONST: + { + std::vector supportedAttributes = { Attribute_NONE }; + + std::vector supportedTypes = + { + DType_FP16, + DType_FP32, + DType_UINT8, + DType_INT8, + DType_INT16, + DType_INT32, + DType_BOOL + }; + + // Check the attribute, data types and bounds for inputs and outputs. + return RunTosaLayerChecksSingleDataType( + op, inputs, outputs, supportedAttributes, supportedTypes, reasonIfUnsupported); + } + case tosa::Op_CONV2D: { - std::vector supportedAttributes = + std::vector supportedAttributes = { Attribute_ConvAttribute }; + + std::vector> supportedTypesMapping = { - Attribute_PoolAttribute + std::tuple(DType_FP16, DType_FP16, DType_FP16), + std::tuple(DType_FP16, DType_FP16, DType_FP32), + std::tuple(DType_FP32, DType_FP32, DType_FP32), + std::tuple(DType_INT8, DType_INT8, DType_INT32) }; + return RunTosaLayerChecksInputWeightsOutputDataType( + op, inputs, outputs, supportedAttributes, supportedTypesMapping, reasonIfUnsupported); + } + case tosa::Op_AVG_POOL2D: + { + std::vector supportedAttributes = { Attribute_PoolAttribute }; + std::vector> supportedTypesMapping = { std::tuple(DType_FP16, DType_FP16), @@ -172,19 +248,12 @@ static bool IsTosaLayerSupported(TosaSerializationOperator* op, }; // Check the attribute, data types and bounds for inputs and outputs. - return RunTosaLayerChecksInputOutputDataType(op, - inputs, - outputs, - supportedAttributes, - supportedTypesMapping, - reasonIfUnsupported); + return RunTosaLayerChecksInputOutputDataType( + op, inputs, outputs, supportedAttributes, supportedTypesMapping, reasonIfUnsupported); } case tosa::Op_MAX_POOL2D: { - std::vector supportedAttributes = - { - Attribute_PoolAttribute - }; + std::vector supportedAttributes = { Attribute_PoolAttribute }; std::vector supportedTypes = { @@ -195,19 +264,12 @@ static bool IsTosaLayerSupported(TosaSerializationOperator* op, }; // Check the attribute, data types and bounds for inputs and outputs. - return RunTosaLayerChecksSingleDataType(op, - inputs, - outputs, - supportedAttributes, - supportedTypes, - reasonIfUnsupported); + return RunTosaLayerChecksSingleDataType( + op, inputs, outputs, supportedAttributes, supportedTypes, reasonIfUnsupported); } case tosa::Op_PAD: { - std::vector supportedAttributes = - { - Attribute_PadAttribute - }; + std::vector supportedAttributes = { Attribute_PadAttribute }; std::vector supportedTypes = { @@ -220,12 +282,8 @@ static bool IsTosaLayerSupported(TosaSerializationOperator* op, }; // Check the attribute, data types and bounds for inputs and outputs. - return RunTosaLayerChecksSingleDataType(op, - inputs, - outputs, - supportedAttributes, - supportedTypes, - reasonIfUnsupported); + return RunTosaLayerChecksSingleDataType( + op, inputs, outputs, supportedAttributes, supportedTypes, reasonIfUnsupported); } default: SetValueChecked(reasonIfUnsupported, "Operation is currently unsupported by the TOSA Reference Backend."); @@ -248,15 +306,31 @@ bool TosaRefLayerSupport::IsLayerSupported(const LayerType& type, switch (type) { + case LayerType::Input: + case LayerType::Output: + return true; case LayerType::Addition: // Setup inputs and outputs inputInfos.push_back(&infos[0]); inputInfos.push_back(&infos[1]); outputInfos.push_back(&infos[2]); break; - case LayerType::Input: - case LayerType::Output: - return true; + case LayerType::Constant: + outputInfos.push_back(&infos[0]); + break; + case LayerType::Convolution2d: + { + inputInfos.push_back(&infos[0]); // input + outputInfos.push_back(&infos[1]); // output + inputInfos.push_back(&infos[2]); // weights + + auto conv2dDesc = PolymorphicDowncast(&descriptor); + if(conv2dDesc->m_BiasEnabled) + { + inputInfos.push_back(&infos[3]); // bias + } + break; + } case LayerType::Pooling2d: // Setup inputs and outputs inputInfos.push_back(&infos[0]); @@ -266,7 +340,7 @@ bool TosaRefLayerSupport::IsLayerSupported(const LayerType& type, break; } - auto mappings = GetTosaMapping(type, inputInfos, outputInfos, descriptor, false); + auto mappings = GetTosaMapping(nullptr, type, inputInfos, outputInfos, descriptor); if (mappings->GetName() == "") { // There currently isn't a TOSA mapping for this layer, as the default was returned. diff --git a/src/backends/tosaReference/test/TosaRefEndToEndTests.cpp b/src/backends/tosaReference/test/TosaRefEndToEndTests.cpp index fbe1265fe3..4245f0d4c4 100644 --- a/src/backends/tosaReference/test/TosaRefEndToEndTests.cpp +++ b/src/backends/tosaReference/test/TosaRefEndToEndTests.cpp @@ -6,6 +6,7 @@ #include "backendsCommon/test/EndToEndTestImpl.hpp" #include "backendsCommon/test/AdditionEndToEndTestImpl.hpp" +#include "backendsCommon/test/Convolution2dEndToEndTestImpl.hpp" #include "backendsCommon/test/Pooling2dEndToEndTestImpl.hpp" #include @@ -30,6 +31,17 @@ TEST_CASE("TosaRefAdditionEndtoEndTestFloat16") AdditionEndToEndFloat16(tosaDefaultBackends); } +// Conv2d +TEST_CASE("TosaRefConv2dEndtoEndTestFloat32") +{ + Convolution2dEndToEnd(tosaDefaultBackends, armnn::DataLayout::NHWC); +} + +TEST_CASE("TosaRefConv2dWithoutBiasEndtoEndTestFloat32") +{ + Convolution2dEndToEnd(tosaDefaultBackends, armnn::DataLayout::NHWC, false); +} + // Max Pool 2D TEST_CASE("TosaRefMaxPool2DEndtoEndTestFloat32") { diff --git a/src/backends/tosaReference/test/TosaRefLayerSupportTests.cpp b/src/backends/tosaReference/test/TosaRefLayerSupportTests.cpp index 48eca344bc..e6fbbf9688 100644 --- a/src/backends/tosaReference/test/TosaRefLayerSupportTests.cpp +++ b/src/backends/tosaReference/test/TosaRefLayerSupportTests.cpp @@ -58,11 +58,98 @@ TEST_CASE("IsLayerSupportedTosaReferenceAdditionUnsupported") CHECK(!supported); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_ADD for input: Op_ADD_input0_") != std::string::npos); + "TOSA Reference Operator: Op_ADD for input: input0_") != std::string::npos); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_ADD for input: Op_ADD_input1_") != std::string::npos); + "TOSA Reference Operator: Op_ADD for input: input1_") != std::string::npos); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_ADD for output: Op_ADD_output0_") != std::string::npos); + "TOSA Reference Operator: Op_ADD for output: output0_") != std::string::npos); +} + +TEST_CASE("IsLayerSupportedTosaReferenceConstant") +{ + armnn::TensorInfo outputInfo({1,1,3,4}, armnn::DataType::Float32); + + armnn::TosaRefLayerSupport supportChecker; + std::string reasonIfNotSupported; + auto supported = supportChecker.IsLayerSupported(armnn::LayerType::Constant, + {outputInfo}, + armnn::BaseDescriptor(), + armnn::EmptyOptional(), + armnn::EmptyOptional(), + reasonIfNotSupported); + + CHECK(supported); +} + +TEST_CASE("IsLayerSupportedTosaReferenceConstantUnsupported") +{ + armnn::TensorInfo outputInfo({1,1,3,4}, armnn::DataType::Signed64); + + armnn::TosaRefLayerSupport supportChecker; + std::string reasonIfNotSupported; + auto supported = supportChecker.IsLayerSupported(armnn::LayerType::Constant, + {outputInfo}, + armnn::BaseDescriptor(), + armnn::EmptyOptional(), + armnn::EmptyOptional(), + reasonIfNotSupported); + + CHECK(!supported); + REQUIRE(reasonIfNotSupported.find( + "TOSA Reference Operator: Op_CONST for output: constant_") != std::string::npos); +} + +TEST_CASE("IsLayerSupportedTosaReferenceConv2d") +{ + armnn::TensorInfo inputInfo ({ 1, 5, 5, 1 }, armnn::DataType::Float32); + armnn::TensorInfo outputInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32); + armnn::TensorInfo weightsInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32); + armnn::TensorInfo biasesInfo ({ 1 }, armnn::DataType::Float32); + + armnn::Convolution2dDescriptor desc; + desc.m_BiasEnabled = true; + + armnn::TosaRefLayerSupport supportChecker; + std::string reasonIfNotSupported; + auto supported = supportChecker.IsLayerSupported(armnn::LayerType::Convolution2d, + {inputInfo, outputInfo, weightsInfo, biasesInfo}, + desc, + armnn::EmptyOptional(), + armnn::EmptyOptional(), + reasonIfNotSupported); + + CHECK(supported); +} + +TEST_CASE("IsLayerSupportedTosaReferenceConv2dUnsupported") +{ + // If inputs and weights are Fp32, output must match. + armnn::TensorInfo inputInfo ({ 1, 5, 5, 1 }, armnn::DataType::Float32); + armnn::TensorInfo outputInfo({ 1, 3, 3, 1 }, armnn::DataType::Signed64); + armnn::TensorInfo weightsInfo({ 1, 3, 3, 1 }, armnn::DataType::Float32, 0.0f, 0, true); + armnn::TensorInfo biasesInfo ({ 1 }, armnn::DataType::Float32, 0.0f, 0, true); + + armnn::Convolution2dDescriptor desc; + desc.m_BiasEnabled = true; + + armnn::TosaRefLayerSupport supportChecker; + std::string reasonIfNotSupported; + auto supported = supportChecker.IsLayerSupported(armnn::LayerType::Convolution2d, + {inputInfo, outputInfo, weightsInfo, biasesInfo}, + desc, + armnn::EmptyOptional(), + armnn::EmptyOptional(), + reasonIfNotSupported); + + CHECK(!supported); + REQUIRE(reasonIfNotSupported.find( + "TOSA Reference Operator: Op_CONV2D for input 0: input0_") != std::string::npos); + REQUIRE(reasonIfNotSupported.find( + "input 1: input1_") != std::string::npos); + REQUIRE(reasonIfNotSupported.find( + "and output: output0_") != std::string::npos); + REQUIRE(reasonIfNotSupported.find( + "has an unsupported input data type combination.") != std::string::npos); } TEST_CASE("IsLayerSupportedTosaReferenceMaxPooling2d") @@ -150,9 +237,9 @@ TEST_CASE("IsLayerSupportedTosaReferenceMaxPooling2dUnsupported") CHECK(!supported); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_MAX_POOL2D for input: Op_MAX_POOL2D_input0_") != std::string::npos); + "TOSA Reference Operator: Op_MAX_POOL2D for input: input0_") != std::string::npos); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_MAX_POOL2D for output: Op_MAX_POOL2D_output0_") != std::string::npos); + "TOSA Reference Operator: Op_MAX_POOL2D for output: output0_") != std::string::npos); } TEST_CASE("IsLayerSupportedTosaReferenceAvgPooling2dUnsupported_InputOutputDatatypeDifferent") @@ -177,9 +264,9 @@ TEST_CASE("IsLayerSupportedTosaReferenceAvgPooling2dUnsupported_InputOutputDatat CHECK(!supported); REQUIRE(reasonIfNotSupported.find( - "TOSA Reference Operator: Op_AVG_POOL2D for input: Op_PAD_intermediate0_") != std::string::npos); + "TOSA Reference Operator: Op_AVG_POOL2D for input: intermediate0_") != std::string::npos); REQUIRE(reasonIfNotSupported.find( - " and output: Op_AVG_POOL2D_output0_") != std::string::npos); + " and output: output0_") != std::string::npos); REQUIRE(reasonIfNotSupported.find( " has an unsupported input data type: 8 to output data type: 10") != std::string::npos); } diff --git a/src/backends/tosaReference/workloads/TosaRefPreCompiledWorkload.cpp b/src/backends/tosaReference/workloads/TosaRefPreCompiledWorkload.cpp index ffdbf6f49b..ba353a32c7 100644 --- a/src/backends/tosaReference/workloads/TosaRefPreCompiledWorkload.cpp +++ b/src/backends/tosaReference/workloads/TosaRefPreCompiledWorkload.cpp @@ -23,13 +23,10 @@ TosaRefPreCompiledWorkload::TosaRefPreCompiledWorkload(const PreCompiledQueueDes void TosaRefPreCompiledWorkload::Execute() const { - uint32_t numInputBuffers = static_cast(m_Data.m_Inputs.size()); - uint32_t numOutputBuffers = static_cast(m_Data.m_Outputs.size()); - tosa::TosaSerializationHandler* handler = static_cast(m_Data.m_PreCompiledObject); - std::vector input_names = handler->GetInputs(); - std::vector output_names = handler->GetOutputs(); + std::vector inputNames = handler->GetInputs(); + std::vector outputNames = handler->GetOutputs(); TosaReference::IModelRunner runner; GraphStatus status; @@ -42,29 +39,29 @@ void TosaRefPreCompiledWorkload::Execute() const } // Set the inputs - for (uint32_t inputSlotIdx = 0; inputSlotIdx < numInputBuffers; ++inputSlotIdx) + for (uint32_t inputSlotIdx = 0; inputSlotIdx < inputNames.size(); ++inputSlotIdx) { DataType dataType = m_workloadInfo.m_InputTensorInfos[inputSlotIdx].GetDataType(); switch (dataType) { case DataType::Float16: - SetInput(runner, input_names[inputSlotIdx], inputSlotIdx); + SetInput(runner, inputNames[inputSlotIdx], inputSlotIdx); break; case DataType::Float32: - SetInput(runner, input_names[inputSlotIdx], inputSlotIdx); + SetInput(runner, inputNames[inputSlotIdx], inputSlotIdx); break; case DataType::QAsymmU8: case DataType::QAsymmS8: case DataType::QSymmS8: case DataType::QSymmS16: case DataType::Signed32: - SetInput(runner, input_names[inputSlotIdx], inputSlotIdx); + SetInput(runner, inputNames[inputSlotIdx], inputSlotIdx); break; case DataType::Signed64: - SetInput(runner, input_names[inputSlotIdx], inputSlotIdx); + SetInput(runner, inputNames[inputSlotIdx], inputSlotIdx); break; case DataType::Boolean: - SetInput(runner, input_names[inputSlotIdx], inputSlotIdx); + SetInput(runner, inputNames[inputSlotIdx], inputSlotIdx); break; default: throw armnn::Exception("Input data type is unsupported in TOSA Reference Backend."); @@ -79,29 +76,29 @@ void TosaRefPreCompiledWorkload::Execute() const } // Gets the outputs - for (uint32_t outputSlotIdx = 0; outputSlotIdx < numOutputBuffers; ++outputSlotIdx) + for (uint32_t outputSlotIdx = 0; outputSlotIdx < outputNames.size(); ++outputSlotIdx) { DataType dataType = m_workloadInfo.m_OutputTensorInfos[outputSlotIdx].GetDataType(); switch (dataType) { case DataType::Float16: - GetOutput(runner, output_names[outputSlotIdx], outputSlotIdx); + GetOutput(runner, outputNames[outputSlotIdx], outputSlotIdx); break; case DataType::Float32: - GetOutput(runner, output_names[outputSlotIdx], outputSlotIdx); + GetOutput(runner, outputNames[outputSlotIdx], outputSlotIdx); break; case DataType::QAsymmU8: case DataType::QAsymmS8: case DataType::QSymmS8: case DataType::QSymmS16: case DataType::Signed32: - GetOutput(runner, output_names[outputSlotIdx], outputSlotIdx); + GetOutput(runner, outputNames[outputSlotIdx], outputSlotIdx); break; case DataType::Signed64: - GetOutput(runner, output_names[outputSlotIdx], outputSlotIdx); + GetOutput(runner, outputNames[outputSlotIdx], outputSlotIdx); break; case DataType::Boolean: - GetOutput(runner, output_names[outputSlotIdx], outputSlotIdx); + GetOutput(runner, outputNames[outputSlotIdx], outputSlotIdx); break; default: throw armnn::Exception("Output data type is unsupported in TOSA Reference Backend."); -- cgit v1.2.1