From c9cc80455ff29fd2c8622c9487ec9c57ade6ea30 Mon Sep 17 00:00:00 2001 From: Aron Virginas-Tar Date: Thu, 1 Nov 2018 16:15:57 +0000 Subject: IVGCVSW-1946: Remove armnn/src from the include paths Change-Id: I663a0a0fccb43ee960ec070121a59df9db0bb04e --- .../backendsCommon/test/Conv2dTestImpl.hpp | 1240 ++++++++++++++++++++ 1 file changed, 1240 insertions(+) create mode 100755 src/backends/backendsCommon/test/Conv2dTestImpl.hpp (limited to 'src/backends/backendsCommon/test/Conv2dTestImpl.hpp') diff --git a/src/backends/backendsCommon/test/Conv2dTestImpl.hpp b/src/backends/backendsCommon/test/Conv2dTestImpl.hpp new file mode 100755 index 0000000000..aa3a44db5d --- /dev/null +++ b/src/backends/backendsCommon/test/Conv2dTestImpl.hpp @@ -0,0 +1,1240 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include +#include +#include +#include + +#include +#include "QuantizeHelper.hpp" + +#include +#include +#include "Permute.hpp" +#include + +// Mapping from input type to bias type for fully connected layers. +// float => float, uint8_t => int32_t +template +struct FullyConnectedBiasTypeForInputType; + +template<> +struct FullyConnectedBiasTypeForInputType +{ + using Type = float; +}; + +template<> +struct FullyConnectedBiasTypeForInputType +{ + using Type = int32_t; +}; + +// Modifies a std::vector in-place using a specified bias. +template +void ApplyBias(std::vector& v, float vScale, int32_t vOffset, + const std::vector& bias, float bScale, int32_t bOffset, uint32_t w, uint32_t h) +{ + BOOST_ASSERT_MSG((armnn::IsQuantizedType() && vScale != 0.0f) || (!armnn::IsQuantizedType()), + "Invalid type and parameter combination."); + BOOST_ASSERT_MSG((armnn::IsQuantizedType() && bScale != 0.0f) || (!armnn::IsQuantizedType()), + "Invalid type and parameter combination."); + + // Note we need to dequantize and re-quantize the image value and the bias. + for (uint32_t i = 0; i < bias.size(); ++i) + { + float dBias = SelectiveDequantize(bias[i], bScale, bOffset); + for (uint32_t y = 0; y < h; ++y) + { + for (uint32_t x = 0; x < w; ++x) + { + uint32_t offset = (i * h + y) * w + x; + BOOST_ASSERT(offset < v.size()); + T& outRef = v[offset]; + float dOutput = SelectiveDequantize(outRef, vScale, vOffset); + outRef = SelectiveQuantize(dOutput + dBias, vScale, vOffset); + } + } + } +} + +template +LayerTestResult SimpleConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, + const boost::multi_array& originalInput, + const boost::multi_array& originalKernel, + const boost::multi_array& bias, + const boost::multi_array& originalOutputExpected, + float qScale, + int32_t qOffset, + const armnn::DataLayoutIndexed& layout = armnn::DataLayout::NCHW, + uint32_t padLeft = 0, + uint32_t padTop = 0, + uint32_t padRight = 0, + uint32_t padBottom = 0) +{ + unsigned int inputHeight = boost::numeric_cast(originalInput.shape()[2]); + unsigned int inputWidth = boost::numeric_cast(originalInput.shape()[3]); + unsigned int inputChannels = boost::numeric_cast(originalInput.shape()[1]); + unsigned int inputNum = boost::numeric_cast(originalInput.shape()[0]); + + unsigned int outputHeight = boost::numeric_cast(originalOutputExpected.shape()[2]); + unsigned int outputWidth = boost::numeric_cast(originalOutputExpected.shape()[3]); + unsigned int outputChannels = boost::numeric_cast(originalOutputExpected.shape()[1]); + unsigned int outputNum = boost::numeric_cast(originalOutputExpected.shape()[0]); + + unsigned int kernelHeight = boost::numeric_cast(originalKernel.shape()[2]); + unsigned int kernelWidth = boost::numeric_cast(originalKernel.shape()[3]); + unsigned int kernelChannels = boost::numeric_cast(originalKernel.shape()[1]); + unsigned int kernelDepthMul = boost::numeric_cast(originalKernel.shape()[0]); + + bool biasEnabled = bias.size() > 0; + + // This function currently assumes 1 batch of input/output (and duplicates this into 2 batches). + BOOST_ASSERT(inputNum == 1); + BOOST_ASSERT(outputNum == 1); + + // If a bias is used, its size must equal the number of output channels. + BOOST_ASSERT(!biasEnabled || bias.size() == outputChannels); + + + // Note these tensors will use two (identical) batches. + armnn::TensorInfo inputTensorInfo = GetTensorInfo(2*inputNum, inputChannels, inputHeight, inputWidth, layout); + armnn::TensorInfo outputTensorInfo = GetTensorInfo( + 2*outputNum, outputChannels, outputHeight, outputWidth, layout); + armnn::TensorInfo kernelDesc = GetTensorInfo(kernelDepthMul, kernelChannels, kernelHeight, kernelWidth, layout); + armnn::TensorInfo biasDesc({static_cast(bias.size())}, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if(armnn::IsQuantizedType()) + { + inputTensorInfo.SetQuantizationScale(qScale); + inputTensorInfo.SetQuantizationOffset(qOffset); + outputTensorInfo.SetQuantizationScale(qScale); + outputTensorInfo.SetQuantizationOffset(qOffset); + kernelDesc.SetQuantizationScale(qScale); + kernelDesc.SetQuantizationOffset(qOffset); + biasDesc.SetQuantizationScale(qScale*qScale); + biasDesc.SetQuantizationOffset(0); + } + + LayerTestResult ret(outputTensorInfo); + + // Construct input data - two batches of the same input image. + std::vector inputImage; + inputImage.assign(originalInput.data(), originalInput.data() + 1*inputChannels*inputHeight*inputWidth); + std::vector inputData; + inputData.insert(inputData.end(), inputImage.begin(), inputImage.end()); + inputData.insert(inputData.end(), inputImage.begin(), inputImage.end()); + + // at this point if we require it permute the input data + const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(inputData.size()); + armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data()); + inputData = tmp; + } + + auto batchedInput = MakeTensor(inputTensorInfo, inputData); + + std::vector outputImage; + outputImage.assign(originalOutputExpected.data(), + originalOutputExpected.data() + outputChannels*outputHeight*outputWidth); + + // Apply bias to output image if it is enabled. + if(biasEnabled) + { + std::vector biasV; + biasV.assign(bias.data(), bias.data() + outputChannels); + ApplyBias(outputImage, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), + biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), + outputWidth, outputHeight); + } + + // Construct expected output data - two identical images. + std::vector outputData; + outputData.insert(outputData.end(), outputImage.begin(), outputImage.end()); + outputData.insert(outputData.end(), outputImage.begin(), outputImage.end()); + + // at this point if we require it permute the expected output + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(outputData.size()); + armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputData.data(), tmp.data()); + outputData = tmp; + } + ret.outputExpected = MakeTensor(outputTensorInfo, outputData); + + // Todo: nontrivial padding and strides. + uint32_t strideX = 1; + uint32_t strideY = 1; + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::Convolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + // Permute the kernel if necessary + boost::multi_array kernel = boost::multi_array(originalKernel); + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data()); + } + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + + if(biasEnabled) + { + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + } + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padLeft; + data.m_Parameters.m_PadRight = padRight; + data.m_Parameters.m_PadTop = padTop; + data.m_Parameters.m_PadBottom = padBottom; + data.m_Parameters.m_BiasEnabled = biasEnabled; + data.m_Parameters.m_DataLayout = layout.GetDataLayout(); + + std::unique_ptr workload = workloadFactory.CreateConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult SimpleConvolution2dNhwcTestImpl(armnn::IWorkloadFactory& workloadFactory, + const boost::multi_array& input, + const boost::multi_array& kernel, + const boost::multi_array& bias, + const boost::multi_array& outputExpected, + armnn::DataLayout dataLayout, + float qScale, + int32_t qOffset, + uint32_t padLeft = 1, + uint32_t padTop = 1, + uint32_t padRight = 1, + uint32_t padBottom = 1, + uint32_t strideX = 1, + uint32_t strideY = 1) +{ + unsigned int inputNum = boost::numeric_cast(input.shape()[0]); + unsigned int inputChannels = boost::numeric_cast(input.shape()[3]); + unsigned int inputHeight = boost::numeric_cast(input.shape()[1]); + unsigned int inputWidth = boost::numeric_cast(input.shape()[2]); + + unsigned int kernelChanMul = boost::numeric_cast(kernel.shape()[0]); + unsigned int kernelChannels = boost::numeric_cast(kernel.shape()[3]); + unsigned int kernelHeight = boost::numeric_cast(kernel.shape()[1]); + unsigned int kernelWidth = boost::numeric_cast(kernel.shape()[2]); + + unsigned int outputNum = boost::numeric_cast(outputExpected.shape()[0]); + unsigned int outputChannels = boost::numeric_cast(outputExpected.shape()[3]); + unsigned int outputHeight = boost::numeric_cast(outputExpected.shape()[1]); + unsigned int outputWidth = boost::numeric_cast(outputExpected.shape()[2]); + + bool biasEnabled = bias.size() > 0; + + // Creates the tensors. + armnn::TensorInfo inputTensorInfo({inputNum, inputHeight, inputWidth, inputChannels}, armnn::GetDataType()); + armnn::TensorInfo outputTensorInfo({outputNum, outputHeight, outputWidth, outputChannels}, + armnn::GetDataType()); + armnn::TensorInfo kernelDesc({kernelChanMul, kernelHeight, kernelWidth, kernelChannels}, armnn::GetDataType()); + armnn::TensorInfo biasDesc({static_cast(bias.size())}, armnn::GetDataType()); + + // Construct the input data. + std::vector inputData; + inputData.assign(input.data(), input.data() + inputHeight*inputWidth*inputChannels); + auto batchedInput = MakeTensor(inputTensorInfo, inputData); + + // Construct the output data, with bias applied, as appropriate. + std::vector outputData; + outputData.assign(outputExpected.data(), outputExpected.data() + outputHeight*outputWidth*outputChannels); + + LayerTestResult ret(outputTensorInfo); + ret.outputExpected = MakeTensor(outputTensorInfo, outputData); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + armnn::Convolution2dQueueDescriptor data; + + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padLeft; + data.m_Parameters.m_PadRight = padRight; + data.m_Parameters.m_PadTop = padTop; + data.m_Parameters.m_PadBottom = padBottom; + data.m_Parameters.m_BiasEnabled = biasEnabled; + data.m_Parameters.m_DataLayout = dataLayout; + + armnn::WorkloadInfo info; + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + std::unique_ptr workload = workloadFactory.CreateConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadFactory& workloadFactory, + const boost::multi_array& input, + const boost::multi_array& originalKernel, + const boost::multi_array& bias, + const boost::multi_array& outputExpected, + float qScale, + int32_t qOffset, + const armnn::DataLayoutIndexed& layout, + uint32_t padLeft = 0, + uint32_t padTop = 0, + uint32_t padRight = 0, + uint32_t padBottom = 0, + uint32_t strideX = 1, + uint32_t strideY = 1) +{ + unsigned int inputNum = boost::numeric_cast(input.shape()[0]); + unsigned int inputChannels = boost::numeric_cast(input.shape()[1]); + unsigned int inputHeight = boost::numeric_cast(input.shape()[2]); + unsigned int inputWidth = boost::numeric_cast(input.shape()[3]); + unsigned int kernelChanMul = boost::numeric_cast(originalKernel.shape()[0]); + unsigned int kernelChannels = boost::numeric_cast(originalKernel.shape()[1]); + unsigned int kernelHeight = boost::numeric_cast(originalKernel.shape()[2]); + unsigned int kernelWidth = boost::numeric_cast(originalKernel.shape()[3]); + unsigned int outputNum = boost::numeric_cast(outputExpected.shape()[0]); + unsigned int outputChannels = boost::numeric_cast(outputExpected.shape()[1]); + unsigned int outputHeight = boost::numeric_cast(outputExpected.shape()[2]); + unsigned int outputWidth = boost::numeric_cast(outputExpected.shape()[3]); + + // If a bias is used, its size must equal the number of output channels. + bool biasEnabled = bias.size() > 0; + BOOST_ASSERT(!biasEnabled || bias.size() == outputChannels); + + // Creates the tensors. + armnn::TensorInfo inputTensorInfo = GetTensorInfo(inputNum, inputChannels, inputHeight, inputWidth, layout); + armnn::TensorInfo outputTensorInfo = GetTensorInfo(outputNum, outputChannels, outputHeight, outputWidth, layout); + armnn::TensorInfo kernelDesc = GetTensorInfo(kernelChanMul, kernelChannels, kernelHeight, kernelWidth, layout); + armnn::TensorInfo biasDesc({static_cast(bias.size())}, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if (armnn::IsQuantizedType()) + { + inputTensorInfo.SetQuantizationScale(qScale); + inputTensorInfo.SetQuantizationOffset(qOffset); + outputTensorInfo.SetQuantizationScale(qScale); + outputTensorInfo.SetQuantizationOffset(qOffset); + kernelDesc.SetQuantizationScale(qScale); + kernelDesc.SetQuantizationOffset(qOffset); + biasDesc.SetQuantizationScale(qScale*qScale); + biasDesc.SetQuantizationOffset(0); + } + + // Construct the input data. + std::vector inputData; + inputData.assign(input.data(), input.data() + inputChannels*inputHeight*inputWidth); + + // At this point if we require it permute the input data + const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(inputData.size()); + armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data()); + inputData = tmp; + } + + auto batchedInput = MakeTensor(inputTensorInfo, inputData); + + // Construct the output data, with bias applied, as appropriate. + std::vector outputData; + outputData.assign(outputExpected.data(), outputExpected.data() + outputChannels*outputHeight*outputWidth); + if (biasEnabled) + { + std::vector biasV; + biasV.assign(bias.data(), bias.data() + outputChannels); + ApplyBias(outputData, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), + biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), + outputWidth, outputHeight); + } + + LayerTestResult ret(outputTensorInfo); + + // At this point if we require it permute the expected output + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(outputData.size()); + armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputData.data(), tmp.data()); + outputData = tmp; + } + + ret.outputExpected = MakeTensor(outputTensorInfo, outputData); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + + // Permute the kernel if necessary + boost::multi_array kernel = boost::multi_array(originalKernel); + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data()); + } + + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + if (biasEnabled) + { + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + } + + armnn::DepthwiseConvolution2dQueueDescriptor data; + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs. + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padLeft; + data.m_Parameters.m_PadRight = padRight; + data.m_Parameters.m_PadTop = padTop; + data.m_Parameters.m_PadBottom = padBottom; + data.m_Parameters.m_BiasEnabled = biasEnabled; + data.m_Parameters.m_DataLayout = layout.GetDataLayout(); + + armnn::WorkloadInfo info; + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + std::unique_ptr workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult DepthwiseConvolution2dDepthMul1TestImpl(armnn::IWorkloadFactory& workloadFactory, + float qScale, + int32_t qOffset, + bool biasEnabled, + const armnn::DataLayoutIndexed& layout) +{ + unsigned int inputHeight = 3; + unsigned int inputWidth = 3; + unsigned int inputChannels = 2; + unsigned int inputNum = 1; + + unsigned int kernelHeight = 3; + unsigned int kernelWidth = 3; + unsigned int kernelChannels = inputChannels; + + unsigned int outputHeight = 1; + unsigned int outputWidth = 1; + unsigned int outputChannels = kernelChannels; + unsigned int outputNum = inputNum; + + armnn::TensorInfo inputTensorInfo = GetTensorInfo(inputNum, inputChannels, inputHeight, inputWidth, layout); + armnn::TensorInfo outputTensorInfo = GetTensorInfo(outputNum, outputChannels, outputHeight, outputWidth, layout); + armnn::TensorInfo kernelDesc = GetTensorInfo(1, outputChannels, kernelHeight, kernelWidth, layout); + armnn::TensorInfo biasDesc({ outputChannels }, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if(armnn::IsQuantizedType()) + { + inputTensorInfo.SetQuantizationScale(qScale); + inputTensorInfo.SetQuantizationOffset(qOffset); + outputTensorInfo.SetQuantizationScale(qScale); + outputTensorInfo.SetQuantizationOffset(qOffset); + kernelDesc.SetQuantizationScale(qScale); + kernelDesc.SetQuantizationOffset(qOffset); + biasDesc.SetQuantizationScale(qScale*qScale); + biasDesc.SetQuantizationOffset(0); + } + std::vector inputData = std::vector( + QuantizedVector(inputTensorInfo.GetQuantizationScale(), inputTensorInfo.GetQuantizationOffset(), { + 1.f, 2.f, 1.f, + 2.f, 1.f, 2.f, + 1.f, 2.f, 1.f, + + 1.f, 2.f, 1.f, + 2.f, 1.f, 2.f, + 1.f, 2.f, 1.f, + })); + // at this point if we require it permute the input data + const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(inputData.size()); + armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data()); + inputData = tmp; + } + auto input = MakeTensor(inputTensorInfo, inputData); + + std::vector biasV(QuantizedVector(biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), + {0, 2})); + auto bias = MakeTensor(biasDesc, biasV); + + std::vector kernelData = std::vector( + QuantizedVector(kernelDesc.GetQuantizationScale(), kernelDesc.GetQuantizationOffset(), { + 1.f, 0.f, 1.f, + 0.f, 0.f, 0.f, + -1.f, 0.f, -1.f, + + 1.f, 0.f, 1.f, + 0.f, 0.f, 0.f, + -1.f, 0.f, -1.f, + })); + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(kernelData.size()); + armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, kernelData.data(), tmp.data()); + kernelData = tmp; + } + auto kernel = MakeTensor(kernelDesc, kernelData); + + // Manually calculated. + std::vector outputImage( + QuantizedVector(outputTensorInfo.GetQuantizationScale(), + outputTensorInfo.GetQuantizationOffset(), + {0.f, 0.f}) + ); + + // Optionally apply bias to output image. + if(biasEnabled) + { + ApplyBias(outputImage, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), + biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), + outputWidth, outputHeight); + } + + LayerTestResult ret(outputTensorInfo); + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + std::vector tmp(outputImage.size()); + armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputImage.data(), tmp.data()); + outputImage = tmp; + } + + ret.outputExpected = MakeTensor(outputTensorInfo, outputImage); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::DepthwiseConvolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. + data.m_Parameters.m_StrideX = 1; + data.m_Parameters.m_StrideY = 1; + data.m_Parameters.m_PadLeft = 0; + data.m_Parameters.m_PadRight = 0; + data.m_Parameters.m_PadTop = 0; + data.m_Parameters.m_PadBottom = 0; + data.m_Parameters.m_BiasEnabled = biasEnabled; + data.m_Parameters.m_DataLayout = layout.GetDataLayout(); + + std::unique_ptr workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult DepthwiseConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, + float qScale, + int32_t qOffset, + bool biasEnabled, + const armnn::DataLayoutIndexed& layout) +{ + unsigned int depthMultiplier = 2; + + unsigned int inputHeight = 8; + unsigned int inputWidth = 16; + unsigned int inputChannels = 2; + unsigned int inputBatchSize = 1; + + unsigned int kernelHeight = 5; + unsigned int kernelWidth = 3; + + unsigned int outputHeight = inputHeight - kernelHeight + 1 + 2; + unsigned int outputWidth = (inputWidth - kernelWidth + 1)/2; + unsigned int outputChannels = inputChannels * depthMultiplier; + unsigned int outputBatchSize = inputBatchSize; + + armnn::TensorInfo inputTensorInfo = GetTensorInfo( + inputBatchSize, inputChannels, inputHeight, inputWidth, layout); + armnn::TensorInfo outputTensorInfo = GetTensorInfo( + outputBatchSize, outputChannels, outputHeight, outputWidth, layout); + armnn::TensorInfo kernelDesc = GetTensorInfo( + depthMultiplier, inputChannels, kernelHeight, kernelWidth, layout); + armnn::TensorInfo biasDesc({outputChannels}, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if(armnn::IsQuantizedType()) + { + inputTensorInfo.SetQuantizationScale(qScale); + inputTensorInfo.SetQuantizationOffset(qOffset); + outputTensorInfo.SetQuantizationScale(qScale); + outputTensorInfo.SetQuantizationOffset(qOffset); + kernelDesc.SetQuantizationScale(qScale); + kernelDesc.SetQuantizationOffset(qOffset); + biasDesc.SetQuantizationScale(qScale*qScale); + biasDesc.SetQuantizationOffset(0); + } + + // NOTE: originalInputData is in NCHW format + std::vector originalInputData = std::vector( + QuantizedVector(inputTensorInfo.GetQuantizationScale(), inputTensorInfo.GetQuantizationOffset(), { + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 + })); + std::vector inputData = originalInputData; + // at this point if we require it permute the input data + const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, originalInputData.data(), inputData.data()); + } + auto input = MakeTensor(inputTensorInfo, inputData); + + std::vector biasV(QuantizedVector(biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), + {0, 2, 1, -1})); + auto bias = MakeTensor(biasDesc, biasV); + + std::vector originalKernelData = std::vector( + QuantizedVector(kernelDesc.GetQuantizationScale(), kernelDesc.GetQuantizationOffset(), { + 1, 1, 1, + 1, -1, 1, + 1, 1, 1, + 1, 1, 1, + 1, 1, 1, + + 2, 2, 2, + 2, 2, 2, + 2, 2, 2, + 2, 2, 2, + 2, 2, 2, + + 0, 0, 0, + 0, -1, 0, + 0, 0, 0, + 0, 0, 0, + 0, 0, 0, + + 0, 0, 0, + 0, 0, 0, + 0, 1, 0, + 0, 0, 0, + 0, 0, 0 + })); + std::vector kernelData = originalKernelData; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernelData.data(), kernelData.data()); + } + auto kernel = MakeTensor(kernelDesc, kernelData); + + // Manually calculated. + std::vector originalOutputImage = std::vector( + QuantizedVector(outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), { + 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, + 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, + 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, + 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, + 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, + 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, + + -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, + -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, + -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, + -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, + + 8.0f, 8.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 8.0f, 8.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f + })); + + // Optionally apply bias to output image. + if(biasEnabled) + { + ApplyBias(originalOutputImage, + outputTensorInfo.GetQuantizationScale(), + outputTensorInfo.GetQuantizationOffset(), + biasV, + biasDesc.GetQuantizationScale(), + biasDesc.GetQuantizationOffset(), + outputWidth, + outputHeight); + } + + LayerTestResult ret(outputTensorInfo); + std::vector outputImage = originalOutputImage; + if (layout.GetDataLayout() == armnn::DataLayout::NHWC) + { + armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, originalOutputImage.data(), outputImage.data()); + } + + ret.outputExpected = MakeTensor(outputTensorInfo, outputImage); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::DepthwiseConvolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. + data.m_Parameters.m_StrideX = 2; + data.m_Parameters.m_StrideY = 1; + data.m_Parameters.m_PadLeft = 0; + data.m_Parameters.m_PadRight = 0; + data.m_Parameters.m_PadTop = 1; + data.m_Parameters.m_PadBottom = 1; + data.m_Parameters.m_BiasEnabled = biasEnabled; + data.m_Parameters.m_DataLayout = layout.GetDataLayout(); + + std::unique_ptr workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult DepthwiseConvolution2dNhwcTestImpl(armnn::IWorkloadFactory& workloadFactory, + const boost::multi_array& input, + const boost::multi_array& kernel, + const boost::multi_array& bias, + const boost::multi_array& outputExpected, + float qScale, + int32_t qOffset, + uint32_t padLeft = 0, + uint32_t padTop = 0, + uint32_t padRight = 0, + uint32_t padBottom = 0, + uint32_t strideX = 1, + uint32_t strideY = 1) +{ + unsigned int inputNum = boost::numeric_cast(input.shape()[0]); + unsigned int inputChannels = boost::numeric_cast(input.shape()[3]); + unsigned int inputHeight = boost::numeric_cast(input.shape()[1]); + unsigned int inputWidth = boost::numeric_cast(input.shape()[2]); + + unsigned int kernelChanMul = boost::numeric_cast(kernel.shape()[0]); + unsigned int kernelChannels = boost::numeric_cast(kernel.shape()[3]); + unsigned int kernelHeight = boost::numeric_cast(kernel.shape()[1]); + unsigned int kernelWidth = boost::numeric_cast(kernel.shape()[2]); + + unsigned int outputNum = boost::numeric_cast(outputExpected.shape()[0]); + unsigned int outputChannels = boost::numeric_cast(outputExpected.shape()[3]); + unsigned int outputHeight = boost::numeric_cast(outputExpected.shape()[1]); + unsigned int outputWidth = boost::numeric_cast(outputExpected.shape()[2]); + + // Creates the tensors. + armnn::TensorInfo inputTensorInfo({inputNum, inputHeight, inputWidth, inputChannels}, armnn::GetDataType()); + armnn::TensorInfo outputTensorInfo({outputNum, outputHeight, outputWidth, outputChannels}, + armnn::GetDataType()); + armnn::TensorInfo kernelDesc({kernelChanMul, kernelHeight, kernelWidth, kernelChannels}, armnn::GetDataType()); + armnn::TensorInfo biasDesc({static_cast(bias.size())}, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if (armnn::IsQuantizedType()) + { + inputTensorInfo.SetQuantizationScale(qScale); + inputTensorInfo.SetQuantizationOffset(qOffset); + outputTensorInfo.SetQuantizationScale(qScale); + outputTensorInfo.SetQuantizationOffset(qOffset); + kernelDesc.SetQuantizationScale(qScale); + kernelDesc.SetQuantizationOffset(qOffset); + biasDesc.SetQuantizationScale(qScale*qScale); + biasDesc.SetQuantizationOffset(0); + } + + // Construct the input data. + std::vector inputData; + inputData.assign(input.data(), input.data() + inputHeight*inputWidth*inputChannels); + auto batchedInput = MakeTensor(inputTensorInfo, inputData); + + // Construct the output data, with bias applied, as appropriate. + std::vector outputData; + outputData.assign(outputExpected.data(), outputExpected.data() + outputHeight*outputWidth*outputChannels); + + LayerTestResult ret(outputTensorInfo); + ret.outputExpected = MakeTensor(outputTensorInfo, outputData); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + armnn::DepthwiseConvolution2dQueueDescriptor data; + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs. + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padLeft; + data.m_Parameters.m_PadRight = padRight; + data.m_Parameters.m_PadTop = padTop; + data.m_Parameters.m_PadBottom = padBottom; + data.m_Parameters.m_DataLayout = armnn::DataLayout::NHWC; + + armnn::WorkloadInfo info; + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + std::unique_ptr workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); + + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + + return ret; +} + +template +LayerTestResult Convolution1dTestImpl(armnn::IWorkloadFactory& workloadFactory, + float qScale, + int32_t qOffset, + bool biasEnabled) +{ + using B = typename FullyConnectedBiasTypeForInputType::Type; + + // Until we have a specialist 1D convolution layer, we can fake one using + // 2D convolution with the final dimension set to 1. + // I don't anticipate this being particularly slow, given that convolution is implemented + // as a matrix multiplication, at which point dimension doesn't matter. + + unsigned int batchSize = 1; + unsigned int inputChannels = 2; + unsigned int outputChannels = 3; + unsigned int inputSize = 5; // The 1D size (could view as 'width' or 'height'). + unsigned int kernelSize = 3; + unsigned int padSize = 2; + unsigned int stride = 1; + unsigned int outputSize = 7; // (inputSize + 2 * padSize - kernelSize + 1) / stride. + + armnn::TensorInfo inputInfo({batchSize, inputChannels, inputSize, 1}, armnn::GetDataType()); + armnn::TensorInfo outputInfo({batchSize, outputChannels, outputSize, 1}, armnn::GetDataType()); + armnn::TensorInfo kernelInfo({outputChannels, inputChannels, kernelSize, 1}, armnn::GetDataType()); + armnn::TensorInfo biasInfo({outputChannels}, armnn::GetDataType()); + + // Set quantization parameters if the requested type is a quantized type. + if(armnn::IsQuantizedType()) + { + inputInfo.SetQuantizationScale(qScale); + inputInfo.SetQuantizationOffset(qOffset); + outputInfo.SetQuantizationScale(qScale); + outputInfo.SetQuantizationOffset(qOffset); + kernelInfo.SetQuantizationScale(qScale); + kernelInfo.SetQuantizationOffset(qOffset); + biasInfo.SetQuantizationScale(inputInfo.GetQuantizationScale()*kernelInfo.GetQuantizationScale()); + biasInfo.SetQuantizationOffset(0); + } + + std::vector inputData( + QuantizedVector(inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(), { + 5.0f, -2.0f, 2.5f, 0.0f, 1.0f, + -3.0f, 3.2f, 5.0f, 2.0f, 3.0f, + })); + + std::vector kernelData( + QuantizedVector(kernelInfo.GetQuantizationScale(), kernelInfo.GetQuantizationOffset(), { + 1.0f, 0.0f, 0.0f, + 0.0f, 2.0f, -1.5f, + + 0.0f, 0.0f, 0.0f, + 0.2f, 0.2f, 0.2f, + + 0.5f, 0.0f, 0.5f, + 0.0f, -1.0f, 0.0f + })); + + std::vector biasData( + QuantizedVector(biasInfo.GetQuantizationScale(), biasInfo.GetQuantizationOffset(), { + 1.0f, 0.0f, 0.0f + })); + + std::vector outputData( + QuantizedVector(outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), { + 4.5f, -10.8f, 5.0f + 6.4f - 7.5f, -2.0f + 10.0f -3.0f, 2.5f + 4.0f - 4.5f, 6.0f, 1.0f, + -0.6f, -0.6f + 0.64f, -0.6f + 0.64f + 1.0f, 0.64f + 1.0f + 0.4f, 1.0f + 0.4f + 0.6f, 0.4f + 0.6f, 0.6f, + 2.5f, -1.0f + 3.0f, 1.25f - 3.2f + 2.5f, -1.0f - 5.0f, 1.25f + 0.5f - 2.0f, -3.0f, 0.5f + })); + + // Optionally apply bias to output image. + if(biasEnabled) + { + ApplyBias(outputData, outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), + biasData, biasInfo.GetQuantizationScale(), biasInfo.GetQuantizationOffset(), + 1, outputSize); + } + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputInfo); + + armnn::Convolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelInfo); + armnn::ScopedCpuTensorHandle biasTensor(biasInfo); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); + AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); + + AddInputToWorkload(data, info, inputInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputInfo, outputHandle.get()); + + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; + data.m_Parameters.m_StrideX = 1; + data.m_Parameters.m_StrideY = stride; + data.m_Parameters.m_PadLeft = 0; + data.m_Parameters.m_PadRight = 0; + data.m_Parameters.m_PadTop = padSize; + data.m_Parameters.m_PadBottom = padSize; + data.m_Parameters.m_BiasEnabled = biasEnabled; + + std::unique_ptr workload = workloadFactory.CreateConvolution2d(data, info); + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), inputData.data()); + + workloadFactory.Finalize(); + workload->Execute(); + + // Output + LayerTestResult ret(outputInfo); + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + ret.outputExpected = MakeTensor(outputInfo, outputData); + return ret; +} + + + +template +LayerTestResult CompareConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, + armnn::IWorkloadFactory& refWorkloadFactory) +{ + unsigned int inputHeight = 8; + unsigned int inputWidth = 16; + unsigned int inputChannels = 3; + unsigned int inputNum = 5; + + unsigned int kernelHeight = 3; + unsigned int kernelWidth = 3; + + unsigned int strideX = 2; + unsigned int strideY = 3; + unsigned int padX = 1; + unsigned int padY = 1; + + unsigned int outputNum = inputNum; + unsigned int outputChannels = 2; + unsigned int outputHeight = (inputHeight + 2 * padY - kernelHeight + strideY) / strideY; + unsigned int outputWidth = (inputWidth + 2 * padX - kernelWidth + strideX) / strideX; + + armnn::TensorInfo inputTensorInfo; + armnn::TensorInfo outputTensorInfo; + armnn::TensorInfo kernelDesc; + armnn::TensorInfo biasDesc; + + unsigned int inputShape[] = {inputNum, inputChannels, inputHeight, inputWidth}; + unsigned int outputShape[] = {outputNum, outputChannels, outputHeight, outputWidth}; + unsigned int kernelShape[] = {outputChannels, inputChannels, kernelHeight, kernelWidth}; + unsigned int biasShape[] = {outputChannels}; + + inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::GetDataType()); + outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::GetDataType()); + kernelDesc = armnn::TensorInfo(4, kernelShape, armnn::GetDataType()); + biasDesc = armnn::TensorInfo(1, biasShape, armnn::GetDataType()); + + LayerTestResult ret(outputTensorInfo); + + auto input = MakeRandomTensor(inputTensorInfo, 124908); + auto kernel = MakeRandomTensor(kernelDesc, 891234); + auto bias = MakeRandomTensor(biasDesc, 1028); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::Convolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padX; + data.m_Parameters.m_PadRight = padX; + data.m_Parameters.m_PadTop = padY; + data.m_Parameters.m_PadBottom = padY; + data.m_Parameters.m_BiasEnabled = true; + + std::unique_ptr outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); + std::unique_ptr inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); + + armnn::Convolution2dQueueDescriptor refData = data; + armnn::WorkloadInfo refInfo = info; + SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); + SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); + + std::unique_ptr workload = workloadFactory.CreateConvolution2d(data, info); + std::unique_ptr workloadRef = refWorkloadFactory.CreateConvolution2d(refData, refInfo); + + outputHandleRef->Allocate(); + inputHandleRef->Allocate(); + + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); + CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + refWorkloadFactory.Finalize(); + workloadRef->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); + + return ret; +} + +template +LayerTestResult CompareDepthwiseConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, + armnn::IWorkloadFactory& refWorkloadFactory, + const armnn::DataLayoutIndexed& layout) +{ + unsigned int inputHeight = 8; + unsigned int inputWidth = 16; + unsigned int inputChannels = 3; + unsigned int inputNum = 5; + + unsigned int kernelHeight = 3; + unsigned int kernelWidth = 3; + unsigned int channelMultiplier = 1; + + unsigned int strideX = 2; + unsigned int strideY = 3; + unsigned int padX = 1; + unsigned int padY = 1; + + unsigned int outputNum = inputNum; + unsigned int outputChannels = inputChannels * channelMultiplier; + unsigned int outputHeight = (inputHeight + 2 * padY - kernelHeight + strideY) / strideY; + unsigned int outputWidth = (inputWidth + 2 * padX - kernelWidth + strideX) / strideX; + + armnn::TensorInfo inputTensorInfo; + armnn::TensorInfo outputTensorInfo; + armnn::TensorInfo kernelDesc; + armnn::TensorInfo biasDesc; + + + std::vector inputShape; + std::vector outputShape; + std::vector kernelShape; + std::vector biasShape= { outputChannels }; + switch (layout.GetDataLayout()) + { + case armnn::DataLayout::NCHW: + inputShape = { inputNum, inputChannels, inputHeight, inputWidth }; + outputShape = { outputNum, outputChannels, outputHeight, outputWidth }; + kernelShape = { channelMultiplier, inputChannels, kernelHeight, kernelWidth }; + break; + case armnn::DataLayout ::NHWC: + inputShape = { inputNum, inputHeight, inputWidth, inputChannels }; + outputShape = { outputNum, outputHeight, outputWidth, outputChannels }; + kernelShape = { channelMultiplier, kernelHeight, kernelWidth, inputChannels }; + break; + default: + throw armnn::InvalidArgumentException("unknown data layout [" + + std::to_string(static_cast(layout.GetDataLayout())) + "]"); + } + + float inputsQScale = armnn::IsQuantizedType() ? 1.0f : 0; + float outputQScale = armnn::IsQuantizedType() ? 2.0f : 0; + int32_t qOffset = 0; + + inputTensorInfo = armnn::TensorInfo(4, inputShape.data(), armnn::GetDataType(), inputsQScale, qOffset); + outputTensorInfo = armnn::TensorInfo(4, outputShape.data(), armnn::GetDataType(), outputQScale, qOffset); + kernelDesc = armnn::TensorInfo(4, kernelShape.data(), armnn::GetDataType(), inputsQScale, qOffset); + biasDesc = armnn::TensorInfo( + 1, biasShape.data(), armnn::GetBiasDataType(armnn::GetDataType()), inputsQScale, qOffset); + + LayerTestResult ret(outputTensorInfo); + + auto input = MakeRandomTensor(inputTensorInfo, 124908, 0.0f, 255.0f); + auto kernel = MakeRandomTensor(kernelDesc, 891234, 0.0f, 255.0f); + auto bias = MakeRandomTensor::Type, 1>( + biasDesc, 1028, 0.0f, 255.0f); + + std::unique_ptr inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); + std::unique_ptr outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); + + armnn::DepthwiseConvolution2dQueueDescriptor data; + armnn::WorkloadInfo info; + armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); + armnn::ScopedCpuTensorHandle biasTensor(biasDesc); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); + AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); + + AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); + AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; + data.m_Parameters.m_StrideX = strideX; + data.m_Parameters.m_StrideY = strideY; + data.m_Parameters.m_PadLeft = padX; + data.m_Parameters.m_PadRight = padX; + data.m_Parameters.m_PadTop = padY; + data.m_Parameters.m_PadBottom = padY; + data.m_Parameters.m_BiasEnabled = true; + data.m_Parameters.m_DataLayout = layout.GetDataLayout(); + + std::unique_ptr outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); + std::unique_ptr inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); + + armnn::DepthwiseConvolution2dQueueDescriptor refData = data; + armnn::WorkloadInfo refInfo = info; + SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); + SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); + + std::unique_ptr workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); + std::unique_ptr workloadRef = refWorkloadFactory.CreateDepthwiseConvolution2d(refData, refInfo); + + outputHandleRef->Allocate(); + inputHandleRef->Allocate(); + + inputHandle->Allocate(); + outputHandle->Allocate(); + + CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); + CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); + + workloadFactory.Finalize(); + workload->Execute(); + refWorkloadFactory.Finalize(); + workloadRef->Execute(); + + CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); + CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); + + return ret; +} -- cgit v1.2.1