telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | #pragma once |
| 6 | |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 7 | #include <string> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 8 | #include <armnn/ArmNN.hpp> |
| 9 | #include <armnn/Tensor.hpp> |
| 10 | #include <armnn/TypesUtils.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 11 | |
David Beck | ac42efd | 2018-09-26 17:41:13 +0100 | [diff] [blame] | 12 | #include <test/TensorHelpers.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 13 | #include "QuantizeHelper.hpp" |
| 14 | |
David Beck | ac42efd | 2018-09-26 17:41:13 +0100 | [diff] [blame] | 15 | #include <backends/CpuTensorHandle.hpp> |
| 16 | #include <backends/WorkloadFactory.hpp> |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 17 | #include "Permute.hpp" |
| 18 | #include <boost/numeric/conversion/cast.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 19 | |
| 20 | // Mapping from input type to bias type for fully connected layers. |
| 21 | // float => float, uint8_t => int32_t |
| 22 | template<typename T> |
| 23 | struct FullyConnectedBiasTypeForInputType; |
| 24 | |
| 25 | template<> |
| 26 | struct FullyConnectedBiasTypeForInputType<float> |
| 27 | { |
| 28 | using Type = float; |
| 29 | }; |
| 30 | |
| 31 | template<> |
| 32 | struct FullyConnectedBiasTypeForInputType<uint8_t> |
| 33 | { |
| 34 | using Type = int32_t; |
| 35 | }; |
| 36 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 37 | // Modifies a std::vector in-place using a specified bias. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 38 | template<typename T, typename B> |
| 39 | void ApplyBias(std::vector<T>& v, float vScale, int32_t vOffset, |
| 40 | const std::vector<B>& bias, float bScale, int32_t bOffset, uint32_t w, uint32_t h) |
| 41 | { |
| 42 | BOOST_ASSERT_MSG((armnn::IsQuantizedType<T>() && vScale != 0.0f) || (!armnn::IsQuantizedType<T>()), |
| 43 | "Invalid type and parameter combination."); |
| 44 | BOOST_ASSERT_MSG((armnn::IsQuantizedType<B>() && bScale != 0.0f) || (!armnn::IsQuantizedType<B>()), |
| 45 | "Invalid type and parameter combination."); |
| 46 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 47 | // Note we need to dequantize and re-quantize the image value and the bias. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 48 | for (uint32_t i = 0; i < bias.size(); ++i) |
| 49 | { |
| 50 | float dBias = SelectiveDequantize(bias[i], bScale, bOffset); |
| 51 | for (uint32_t y = 0; y < h; ++y) |
| 52 | { |
| 53 | for (uint32_t x = 0; x < w; ++x) |
| 54 | { |
| 55 | uint32_t offset = (i * h + y) * w + x; |
| 56 | BOOST_ASSERT(offset < v.size()); |
| 57 | T& outRef = v[offset]; |
| 58 | float dOutput = SelectiveDequantize(outRef, vScale, vOffset); |
| 59 | outRef = SelectiveQuantize<T>(dOutput + dBias, vScale, vOffset); |
| 60 | } |
| 61 | } |
| 62 | } |
| 63 | } |
| 64 | |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 65 | template<typename T> |
| 66 | armnn::TensorInfo GetTensorInfo(unsigned int numberOfBatches, |
| 67 | unsigned int numberOfChannels, |
| 68 | unsigned int height, |
| 69 | unsigned int width, |
| 70 | const armnn::DataLayoutIndexed& layout) |
| 71 | { |
| 72 | switch (layout.GetDataLayout()) |
| 73 | { |
| 74 | case armnn::DataLayout::NCHW: |
| 75 | return armnn::TensorInfo({numberOfBatches, numberOfChannels, height, width}, armnn::GetDataType<T>()); |
| 76 | case armnn::DataLayout ::NHWC: |
| 77 | return armnn::TensorInfo({numberOfBatches, height, width, numberOfChannels}, armnn::GetDataType<T>()); |
| 78 | default: |
| 79 | throw armnn::InvalidArgumentException("unknown data layout [" |
| 80 | + std::to_string(static_cast<int>(layout.GetDataLayout())) + "]"); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | |
| 85 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 86 | template<typename T, typename B> |
| 87 | LayerTestResult<T, 4> SimpleConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 88 | const boost::multi_array<T, 4>& originalInput, |
| 89 | const boost::multi_array<T, 4>& originalKernel, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 90 | const boost::multi_array<B, 1>& bias, |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 91 | const boost::multi_array<T, 4>& originalOutputExpected, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 92 | float qScale, |
| 93 | int32_t qOffset, |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 94 | const armnn::DataLayoutIndexed& layout = armnn::DataLayout::NCHW, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 95 | uint32_t padLeft = 0, |
| 96 | uint32_t padTop = 0, |
| 97 | uint32_t padRight = 0, |
| 98 | uint32_t padBottom = 0) |
| 99 | { |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 100 | unsigned int inputHeight = boost::numeric_cast<unsigned int>(originalInput.shape()[2]); |
| 101 | unsigned int inputWidth = boost::numeric_cast<unsigned int>(originalInput.shape()[3]); |
| 102 | unsigned int inputChannels = boost::numeric_cast<unsigned int>(originalInput.shape()[1]); |
| 103 | unsigned int inputNum = boost::numeric_cast<unsigned int>(originalInput.shape()[0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 104 | |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 105 | unsigned int outputHeight = boost::numeric_cast<unsigned int>(originalOutputExpected.shape()[2]); |
| 106 | unsigned int outputWidth = boost::numeric_cast<unsigned int>(originalOutputExpected.shape()[3]); |
| 107 | unsigned int outputChannels = boost::numeric_cast<unsigned int>(originalOutputExpected.shape()[1]); |
| 108 | unsigned int outputNum = boost::numeric_cast<unsigned int>(originalOutputExpected.shape()[0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 109 | |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 110 | unsigned int kernelHeight = boost::numeric_cast<unsigned int>(originalKernel.shape()[2]); |
| 111 | unsigned int kernelWidth = boost::numeric_cast<unsigned int>(originalKernel.shape()[3]); |
| 112 | unsigned int kernelChannels = boost::numeric_cast<unsigned int>(originalKernel.shape()[1]); |
| 113 | unsigned int kernelDepthMul = boost::numeric_cast<unsigned int>(originalKernel.shape()[0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 114 | |
| 115 | bool biasEnabled = bias.size() > 0; |
| 116 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 117 | // This function currently assumes 1 batch of input/output (and duplicates this into 2 batches). |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 118 | BOOST_ASSERT(inputNum == 1); |
| 119 | BOOST_ASSERT(outputNum == 1); |
| 120 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 121 | // If a bias is used, its size must equal the number of output channels. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 122 | BOOST_ASSERT(!biasEnabled || bias.size() == outputChannels); |
| 123 | |
| 124 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 125 | // Note these tensors will use two (identical) batches. |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 126 | // NOTE: if layout is unknown we will get an exception at this point. |
| 127 | armnn::TensorInfo inputTensorInfo = GetTensorInfo<T>(2*inputNum, inputChannels, inputHeight, inputWidth, layout); |
| 128 | armnn::TensorInfo outputTensorInfo = GetTensorInfo<T>( |
| 129 | 2*outputNum, outputChannels, outputHeight, outputWidth, layout); |
| 130 | armnn::TensorInfo kernelDesc = GetTensorInfo<T>(kernelDepthMul, kernelChannels, kernelHeight, kernelWidth, layout); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 131 | armnn::TensorInfo biasDesc({static_cast<unsigned int>(bias.size())}, armnn::GetDataType<B>()); |
| 132 | |
| 133 | // Set quantization parameters if the requested type is a quantized type. |
| 134 | if(armnn::IsQuantizedType<T>()) |
| 135 | { |
| 136 | inputTensorInfo.SetQuantizationScale(qScale); |
| 137 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 138 | outputTensorInfo.SetQuantizationScale(qScale); |
| 139 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 140 | kernelDesc.SetQuantizationScale(qScale); |
| 141 | kernelDesc.SetQuantizationOffset(qOffset); |
| 142 | biasDesc.SetQuantizationScale(qScale*qScale); |
| 143 | biasDesc.SetQuantizationOffset(0); |
| 144 | } |
| 145 | |
| 146 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 147 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 148 | // Construct input data - two batches of the same input image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 149 | std::vector<T> inputImage; |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 150 | inputImage.assign(originalInput.data(), originalInput.data() + 1*inputChannels*inputHeight*inputWidth); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 151 | std::vector<T> inputData; |
| 152 | inputData.insert(inputData.end(), inputImage.begin(), inputImage.end()); |
| 153 | inputData.insert(inputData.end(), inputImage.begin(), inputImage.end()); |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 154 | |
| 155 | // at this point if we require it permute the input data |
| 156 | const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; |
| 157 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 158 | { |
| 159 | std::vector<T> tmp(inputData.size()); |
| 160 | armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data()); |
| 161 | inputData = tmp; |
| 162 | } |
| 163 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 164 | auto batchedInput = MakeTensor<T, 4>(inputTensorInfo, inputData); |
| 165 | |
| 166 | std::vector<T> outputImage; |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 167 | outputImage.assign(originalOutputExpected.data(), |
| 168 | originalOutputExpected.data() + outputChannels*outputHeight*outputWidth); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 169 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 170 | // Apply bias to output image if it is enabled. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 171 | if(biasEnabled) |
| 172 | { |
| 173 | std::vector<T> biasV; |
| 174 | biasV.assign(bias.data(), bias.data() + outputChannels); |
| 175 | ApplyBias(outputImage, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), |
| 176 | biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 177 | outputWidth, outputHeight); |
| 178 | } |
| 179 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 180 | // Construct expected output data - two identical images. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 181 | std::vector<T> outputData; |
| 182 | outputData.insert(outputData.end(), outputImage.begin(), outputImage.end()); |
| 183 | outputData.insert(outputData.end(), outputImage.begin(), outputImage.end()); |
| 184 | |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 185 | // at this point if we require it permute the expected output |
| 186 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 187 | { |
| 188 | std::vector<T> tmp(outputData.size()); |
| 189 | armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputData.data(), tmp.data()); |
| 190 | outputData = tmp; |
| 191 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 192 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputData); |
| 193 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 194 | // Todo: nontrivial padding and strides. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 195 | uint32_t strideX = 1; |
| 196 | uint32_t strideY = 1; |
| 197 | |
| 198 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 199 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 200 | |
| 201 | armnn::Convolution2dQueueDescriptor data; |
| 202 | armnn::WorkloadInfo info; |
| 203 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 204 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 205 | // Permute the kernel if necessary |
| 206 | boost::multi_array<T, 4> kernel = boost::multi_array<T, 4>(originalKernel); |
| 207 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 208 | { |
| 209 | armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data()); |
| 210 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 211 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 212 | |
| 213 | if(biasEnabled) |
| 214 | { |
| 215 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 216 | } |
| 217 | |
| 218 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 219 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 220 | |
| 221 | data.m_Weight = &weightsTensor; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 222 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 223 | data.m_Parameters.m_StrideX = strideX; |
| 224 | data.m_Parameters.m_StrideY = strideY; |
| 225 | data.m_Parameters.m_PadLeft = padLeft; |
| 226 | data.m_Parameters.m_PadRight = padRight; |
| 227 | data.m_Parameters.m_PadTop = padTop; |
| 228 | data.m_Parameters.m_PadBottom = padBottom; |
| 229 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
jimfly01 | 0a088a6 | 2018-10-25 17:05:05 +0100 | [diff] [blame] | 230 | data.m_Parameters.m_DataLayout = layout.GetDataLayout(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 231 | |
| 232 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvolution2d(data, info); |
| 233 | inputHandle->Allocate(); |
| 234 | outputHandle->Allocate(); |
| 235 | |
| 236 | CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); |
| 237 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 238 | workloadFactory.Finalize(); |
| 239 | workload->Execute(); |
| 240 | |
| 241 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 242 | |
| 243 | return ret; |
| 244 | } |
| 245 | |
| 246 | template<typename T, typename B> |
Francis Murtagh | d59116e | 2018-10-04 16:03:07 +0100 | [diff] [blame] | 247 | LayerTestResult<T, 4> SimpleConvolution2dNhwcTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 248 | const boost::multi_array<T, 4>& input, |
| 249 | const boost::multi_array<T, 4>& kernel, |
| 250 | const boost::multi_array<B, 1>& bias, |
| 251 | const boost::multi_array<T, 4>& outputExpected, |
| 252 | armnn::DataLayout dataLayout, |
| 253 | float qScale, |
| 254 | int32_t qOffset, |
| 255 | uint32_t padLeft = 1, |
| 256 | uint32_t padTop = 1, |
| 257 | uint32_t padRight = 1, |
| 258 | uint32_t padBottom = 1, |
| 259 | uint32_t strideX = 1, |
| 260 | uint32_t strideY = 1) |
| 261 | { |
| 262 | unsigned int inputNum = boost::numeric_cast<unsigned int>(input.shape()[0]); |
| 263 | unsigned int inputChannels = boost::numeric_cast<unsigned int>(input.shape()[3]); |
| 264 | unsigned int inputHeight = boost::numeric_cast<unsigned int>(input.shape()[1]); |
| 265 | unsigned int inputWidth = boost::numeric_cast<unsigned int>(input.shape()[2]); |
| 266 | |
| 267 | unsigned int kernelChanMul = boost::numeric_cast<unsigned int>(kernel.shape()[0]); |
| 268 | unsigned int kernelChannels = boost::numeric_cast<unsigned int>(kernel.shape()[3]); |
| 269 | unsigned int kernelHeight = boost::numeric_cast<unsigned int>(kernel.shape()[1]); |
| 270 | unsigned int kernelWidth = boost::numeric_cast<unsigned int>(kernel.shape()[2]); |
| 271 | |
| 272 | unsigned int outputNum = boost::numeric_cast<unsigned int>(outputExpected.shape()[0]); |
| 273 | unsigned int outputChannels = boost::numeric_cast<unsigned int>(outputExpected.shape()[3]); |
| 274 | unsigned int outputHeight = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]); |
| 275 | unsigned int outputWidth = boost::numeric_cast<unsigned int>(outputExpected.shape()[2]); |
| 276 | |
| 277 | bool biasEnabled = bias.size() > 0; |
| 278 | |
| 279 | // Creates the tensors. |
| 280 | armnn::TensorInfo inputTensorInfo({inputNum, inputHeight, inputWidth, inputChannels}, armnn::GetDataType<T>()); |
| 281 | armnn::TensorInfo outputTensorInfo({outputNum, outputHeight, outputWidth, outputChannels}, |
| 282 | armnn::GetDataType<T>()); |
| 283 | armnn::TensorInfo kernelDesc({kernelChanMul, kernelHeight, kernelWidth, kernelChannels}, armnn::GetDataType<T>()); |
| 284 | armnn::TensorInfo biasDesc({static_cast<unsigned int>(bias.size())}, armnn::GetDataType<B>()); |
| 285 | |
| 286 | // Construct the input data. |
| 287 | std::vector<T> inputData; |
| 288 | inputData.assign(input.data(), input.data() + inputHeight*inputWidth*inputChannels); |
| 289 | auto batchedInput = MakeTensor<T, 4>(inputTensorInfo, inputData); |
| 290 | |
| 291 | // Construct the output data, with bias applied, as appropriate. |
| 292 | std::vector<T> outputData; |
| 293 | outputData.assign(outputExpected.data(), outputExpected.data() + outputHeight*outputWidth*outputChannels); |
| 294 | |
| 295 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 296 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputData); |
| 297 | |
| 298 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 299 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 300 | |
| 301 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 302 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 303 | |
| 304 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 305 | |
| 306 | armnn::Convolution2dQueueDescriptor data; |
| 307 | |
| 308 | data.m_Weight = &weightsTensor; |
| 309 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - can be a source of bugs. |
| 310 | data.m_Parameters.m_StrideX = strideX; |
| 311 | data.m_Parameters.m_StrideY = strideY; |
| 312 | data.m_Parameters.m_PadLeft = padLeft; |
| 313 | data.m_Parameters.m_PadRight = padRight; |
| 314 | data.m_Parameters.m_PadTop = padTop; |
| 315 | data.m_Parameters.m_PadBottom = padBottom; |
| 316 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 317 | data.m_Parameters.m_DataLayout = dataLayout; |
| 318 | |
| 319 | armnn::WorkloadInfo info; |
| 320 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 321 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 322 | |
| 323 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvolution2d(data, info); |
| 324 | inputHandle->Allocate(); |
| 325 | outputHandle->Allocate(); |
| 326 | |
| 327 | CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); |
| 328 | |
| 329 | workloadFactory.Finalize(); |
| 330 | workload->Execute(); |
| 331 | |
| 332 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 333 | |
| 334 | return ret; |
| 335 | } |
| 336 | |
| 337 | template<typename T, typename B> |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 338 | LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 339 | const boost::multi_array<T, 4>& input, |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 340 | const boost::multi_array<T, 4>& originalKernel, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 341 | const boost::multi_array<B, 1>& bias, |
| 342 | const boost::multi_array<T, 4>& outputExpected, |
| 343 | float qScale, |
| 344 | int32_t qOffset, |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 345 | const armnn::DataLayoutIndexed& layout, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 346 | uint32_t padLeft = 0, |
| 347 | uint32_t padTop = 0, |
| 348 | uint32_t padRight = 0, |
| 349 | uint32_t padBottom = 0, |
| 350 | uint32_t strideX = 1, |
| 351 | uint32_t strideY = 1) |
| 352 | { |
| 353 | unsigned int inputNum = boost::numeric_cast<unsigned int>(input.shape()[0]); |
| 354 | unsigned int inputChannels = boost::numeric_cast<unsigned int>(input.shape()[1]); |
| 355 | unsigned int inputHeight = boost::numeric_cast<unsigned int>(input.shape()[2]); |
| 356 | unsigned int inputWidth = boost::numeric_cast<unsigned int>(input.shape()[3]); |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 357 | unsigned int kernelChanMul = boost::numeric_cast<unsigned int>(originalKernel.shape()[0]); |
| 358 | unsigned int kernelChannels = boost::numeric_cast<unsigned int>(originalKernel.shape()[1]); |
| 359 | unsigned int kernelHeight = boost::numeric_cast<unsigned int>(originalKernel.shape()[2]); |
| 360 | unsigned int kernelWidth = boost::numeric_cast<unsigned int>(originalKernel.shape()[3]); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 361 | unsigned int outputNum = boost::numeric_cast<unsigned int>(outputExpected.shape()[0]); |
| 362 | unsigned int outputChannels = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]); |
| 363 | unsigned int outputHeight = boost::numeric_cast<unsigned int>(outputExpected.shape()[2]); |
| 364 | unsigned int outputWidth = boost::numeric_cast<unsigned int>(outputExpected.shape()[3]); |
| 365 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 366 | // If a bias is used, its size must equal the number of output channels. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 367 | bool biasEnabled = bias.size() > 0; |
| 368 | BOOST_ASSERT(!biasEnabled || bias.size() == outputChannels); |
| 369 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 370 | // Creates the tensors. |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 371 | armnn::TensorInfo inputTensorInfo = GetTensorInfo<T>(inputNum, inputChannels, inputHeight, inputWidth, layout); |
| 372 | armnn::TensorInfo outputTensorInfo = GetTensorInfo<T>(outputNum, outputChannels, outputHeight, outputWidth, layout); |
| 373 | armnn::TensorInfo kernelDesc = GetTensorInfo<T>(kernelChanMul, kernelChannels, kernelHeight, kernelWidth, layout); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 374 | armnn::TensorInfo biasDesc({static_cast<unsigned int>(bias.size())}, armnn::GetDataType<B>()); |
| 375 | |
| 376 | // Set quantization parameters if the requested type is a quantized type. |
| 377 | if (armnn::IsQuantizedType<T>()) |
| 378 | { |
| 379 | inputTensorInfo.SetQuantizationScale(qScale); |
| 380 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 381 | outputTensorInfo.SetQuantizationScale(qScale); |
| 382 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 383 | kernelDesc.SetQuantizationScale(qScale); |
| 384 | kernelDesc.SetQuantizationOffset(qOffset); |
| 385 | biasDesc.SetQuantizationScale(qScale*qScale); |
| 386 | biasDesc.SetQuantizationOffset(0); |
| 387 | } |
| 388 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 389 | // Construct the input data. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 390 | std::vector<T> inputData; |
| 391 | inputData.assign(input.data(), input.data() + inputChannels*inputHeight*inputWidth); |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 392 | |
| 393 | // At this point if we require it permute the input data |
| 394 | const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 }; |
| 395 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 396 | { |
| 397 | std::vector<T> tmp(inputData.size()); |
| 398 | armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data()); |
| 399 | inputData = tmp; |
| 400 | } |
| 401 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 402 | auto batchedInput = MakeTensor<T, 4>(inputTensorInfo, inputData); |
| 403 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 404 | // Construct the output data, with bias applied, as appropriate. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 405 | std::vector<T> outputData; |
| 406 | outputData.assign(outputExpected.data(), outputExpected.data() + outputChannels*outputHeight*outputWidth); |
| 407 | if (biasEnabled) |
| 408 | { |
| 409 | std::vector<T> biasV; |
| 410 | biasV.assign(bias.data(), bias.data() + outputChannels); |
| 411 | ApplyBias(outputData, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), |
| 412 | biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 413 | outputWidth, outputHeight); |
| 414 | } |
| 415 | |
| 416 | LayerTestResult<T, 4> ret(outputTensorInfo); |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 417 | |
| 418 | // At this point if we require it permute the expected output |
| 419 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 420 | { |
| 421 | std::vector<T> tmp(outputData.size()); |
| 422 | armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputData.data(), tmp.data()); |
| 423 | outputData = tmp; |
| 424 | } |
| 425 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 426 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputData); |
| 427 | |
| 428 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 429 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 430 | |
| 431 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 432 | |
| 433 | // Permute the kernel if necessary |
| 434 | boost::multi_array<T, 4> kernel = boost::multi_array<T, 4>(originalKernel); |
| 435 | if (layout.GetDataLayout() == armnn::DataLayout::NHWC) |
| 436 | { |
| 437 | armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data()); |
| 438 | } |
| 439 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 440 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 441 | |
| 442 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 443 | if (biasEnabled) |
| 444 | { |
| 445 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 446 | } |
| 447 | |
| 448 | armnn::DepthwiseConvolution2dQueueDescriptor data; |
| 449 | data.m_Weight = &weightsTensor; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 450 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 451 | data.m_Parameters.m_StrideX = strideX; |
| 452 | data.m_Parameters.m_StrideY = strideY; |
| 453 | data.m_Parameters.m_PadLeft = padLeft; |
| 454 | data.m_Parameters.m_PadRight = padRight; |
| 455 | data.m_Parameters.m_PadTop = padTop; |
| 456 | data.m_Parameters.m_PadBottom = padBottom; |
| 457 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
jimfly01 | 382a91d | 2018-10-26 15:55:50 +0100 | [diff] [blame] | 458 | data.m_Parameters.m_DataLayout = layout.GetDataLayout(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 459 | |
| 460 | armnn::WorkloadInfo info; |
| 461 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 462 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 463 | |
| 464 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); |
| 465 | inputHandle->Allocate(); |
| 466 | outputHandle->Allocate(); |
| 467 | |
| 468 | CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); |
| 469 | |
| 470 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 471 | workload->Execute(); |
| 472 | |
| 473 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 474 | |
| 475 | return ret; |
| 476 | } |
| 477 | |
| 478 | template<typename T, typename B> |
| 479 | LayerTestResult<T, 4> DepthwiseConvolution2dDepthMul1TestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 480 | float qScale, |
| 481 | int32_t qOffset, |
| 482 | bool biasEnabled) |
| 483 | { |
| 484 | unsigned int inputHeight = 3; |
| 485 | unsigned int inputWidth = 3; |
| 486 | unsigned int inputChannels = 2; |
| 487 | unsigned int inputNum = 1; |
| 488 | |
| 489 | unsigned int kernelHeight = 3; |
| 490 | unsigned int kernelWidth = 3; |
| 491 | unsigned int kernelChannels = inputChannels; |
| 492 | |
| 493 | unsigned int outputHeight = 1; |
| 494 | unsigned int outputWidth = 1; |
| 495 | unsigned int outputChannels = kernelChannels; |
| 496 | unsigned int outputNum = inputNum; |
| 497 | |
| 498 | armnn::TensorInfo inputTensorInfo({ inputNum, inputChannels, inputHeight, inputWidth }, armnn::GetDataType<T>()); |
| 499 | armnn::TensorInfo outputTensorInfo({ outputNum, outputChannels, outputHeight, outputWidth }, |
| 500 | armnn::GetDataType<T>()); |
| 501 | armnn::TensorInfo kernelDesc({ 1, outputChannels, kernelHeight, kernelWidth }, armnn::GetDataType<T>()); |
| 502 | armnn::TensorInfo biasDesc({ outputChannels }, armnn::GetDataType<B>()); |
| 503 | |
| 504 | // Set quantization parameters if the requested type is a quantized type. |
| 505 | if(armnn::IsQuantizedType<T>()) |
| 506 | { |
| 507 | inputTensorInfo.SetQuantizationScale(qScale); |
| 508 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 509 | outputTensorInfo.SetQuantizationScale(qScale); |
| 510 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 511 | kernelDesc.SetQuantizationScale(qScale); |
| 512 | kernelDesc.SetQuantizationOffset(qOffset); |
| 513 | biasDesc.SetQuantizationScale(qScale*qScale); |
| 514 | biasDesc.SetQuantizationOffset(0); |
| 515 | } |
| 516 | |
| 517 | auto input = MakeTensor<T, 4>(inputTensorInfo, std::vector<T>( |
| 518 | QuantizedVector<T>(inputTensorInfo.GetQuantizationScale(), inputTensorInfo.GetQuantizationOffset(), { |
| 519 | 1.f, 2.f, 1.f, |
| 520 | 2.f, 1.f, 2.f, |
| 521 | 1.f, 2.f, 1.f, |
| 522 | |
| 523 | 1.f, 2.f, 1.f, |
| 524 | 2.f, 1.f, 2.f, |
| 525 | 1.f, 2.f, 1.f, |
| 526 | }))); |
| 527 | |
| 528 | std::vector<B> biasV(QuantizedVector<B>(biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 529 | {0, 2})); |
| 530 | auto bias = MakeTensor<B, 1>(biasDesc, biasV); |
| 531 | |
| 532 | auto kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 533 | QuantizedVector<T>(kernelDesc.GetQuantizationScale(), kernelDesc.GetQuantizationOffset(), { |
| 534 | 1.f, 0.f, 1.f, |
| 535 | 0.f, 0.f, 0.f, |
| 536 | -1.f, 0.f, -1.f, |
| 537 | |
| 538 | 1.f, 0.f, 1.f, |
| 539 | 0.f, 0.f, 0.f, |
| 540 | -1.f, 0.f, -1.f, |
| 541 | }))); |
| 542 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 543 | // Manually calculated. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 544 | std::vector<T> outputImage( |
| 545 | QuantizedVector<T>(outputTensorInfo.GetQuantizationScale(), |
| 546 | outputTensorInfo.GetQuantizationOffset(), |
| 547 | {0.f, 0.f}) |
| 548 | ); |
| 549 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 550 | // Optionally apply bias to output image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 551 | if(biasEnabled) |
| 552 | { |
| 553 | ApplyBias(outputImage, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), |
| 554 | biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 555 | outputWidth, outputHeight); |
| 556 | } |
| 557 | |
| 558 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 559 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputImage); |
| 560 | |
| 561 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 562 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 563 | |
| 564 | armnn::DepthwiseConvolution2dQueueDescriptor data; |
| 565 | armnn::WorkloadInfo info; |
| 566 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 567 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 568 | |
| 569 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 570 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 571 | |
| 572 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 573 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 574 | |
| 575 | data.m_Weight = &weightsTensor; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 576 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 577 | data.m_Parameters.m_StrideX = 1; |
| 578 | data.m_Parameters.m_StrideY = 1; |
| 579 | data.m_Parameters.m_PadLeft = 0; |
| 580 | data.m_Parameters.m_PadRight = 0; |
| 581 | data.m_Parameters.m_PadTop = 0; |
| 582 | data.m_Parameters.m_PadBottom = 0; |
| 583 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 584 | |
| 585 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); |
| 586 | inputHandle->Allocate(); |
| 587 | outputHandle->Allocate(); |
| 588 | |
| 589 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 590 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 591 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 592 | workload->Execute(); |
| 593 | |
| 594 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 595 | |
| 596 | return ret; |
| 597 | } |
| 598 | |
| 599 | template<typename T, typename B> |
| 600 | LayerTestResult<T, 4> DepthwiseConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 601 | float qScale, |
| 602 | int32_t qOffset, |
| 603 | bool biasEnabled) |
| 604 | { |
| 605 | unsigned int depthMultiplier = 2; |
| 606 | |
| 607 | unsigned int inputHeight = 8; |
| 608 | unsigned int inputWidth = 16; |
| 609 | unsigned int inputChannels = 2; |
| 610 | unsigned int inputBatchSize = 1; |
| 611 | |
| 612 | unsigned int kernelHeight = 5; |
| 613 | unsigned int kernelWidth = 3; |
| 614 | |
| 615 | unsigned int outputHeight = inputHeight - kernelHeight + 1 + 2; |
| 616 | unsigned int outputWidth = (inputWidth - kernelWidth + 1)/2; |
| 617 | unsigned int outputChannels = inputChannels * depthMultiplier; |
| 618 | unsigned int outputBatchSize = inputBatchSize; |
| 619 | |
| 620 | armnn::TensorInfo inputTensorInfo({inputBatchSize, inputChannels, inputHeight, inputWidth}, |
| 621 | armnn::GetDataType<T>()); |
| 622 | armnn::TensorInfo outputTensorInfo({outputBatchSize, outputChannels, outputHeight, outputWidth}, |
| 623 | armnn::GetDataType<T>()); |
| 624 | armnn::TensorInfo kernelDesc({depthMultiplier, inputChannels, kernelHeight, kernelWidth}, armnn::GetDataType<T>()); |
| 625 | armnn::TensorInfo biasDesc({outputChannels}, armnn::GetDataType<B>()); |
| 626 | |
| 627 | // Set quantization parameters if the requested type is a quantized type. |
| 628 | if(armnn::IsQuantizedType<T>()) |
| 629 | { |
| 630 | inputTensorInfo.SetQuantizationScale(qScale); |
| 631 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 632 | outputTensorInfo.SetQuantizationScale(qScale); |
| 633 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 634 | kernelDesc.SetQuantizationScale(qScale); |
| 635 | kernelDesc.SetQuantizationOffset(qOffset); |
| 636 | biasDesc.SetQuantizationScale(qScale*qScale); |
| 637 | biasDesc.SetQuantizationOffset(0); |
| 638 | } |
| 639 | |
| 640 | auto input = MakeTensor<T, 4>(inputTensorInfo, std::vector<T>( |
| 641 | QuantizedVector<T>(inputTensorInfo.GetQuantizationScale(), inputTensorInfo.GetQuantizationOffset(), { |
| 642 | 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, |
| 643 | 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, |
| 644 | 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, |
| 645 | 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, |
| 646 | 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, |
| 647 | 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, |
| 648 | 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, |
| 649 | 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, |
| 650 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 651 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 652 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 653 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 654 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 655 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 656 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 657 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 |
| 658 | }))); |
| 659 | |
| 660 | std::vector<B> biasV(QuantizedVector<B>(biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 661 | {0, 2, 1, -1})); |
| 662 | auto bias = MakeTensor<B, 1>(biasDesc, biasV); |
| 663 | |
| 664 | auto kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 665 | QuantizedVector<T>(kernelDesc.GetQuantizationScale(), kernelDesc.GetQuantizationOffset(), { |
| 666 | 1, 1, 1, |
| 667 | 1, -1, 1, |
| 668 | 1, 1, 1, |
| 669 | 1, 1, 1, |
| 670 | 1, 1, 1, |
| 671 | |
| 672 | 2, 2, 2, |
| 673 | 2, 2, 2, |
| 674 | 2, 2, 2, |
| 675 | 2, 2, 2, |
| 676 | 2, 2, 2, |
| 677 | |
| 678 | 0, 0, 0, |
| 679 | 0, -1, 0, |
| 680 | 0, 0, 0, |
| 681 | 0, 0, 0, |
| 682 | 0, 0, 0, |
| 683 | |
| 684 | 0, 0, 0, |
| 685 | 0, 0, 0, |
| 686 | 0, 1, 0, |
| 687 | 0, 0, 0, |
| 688 | 0, 0, 0 |
| 689 | }))); |
| 690 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 691 | // Manually calculated. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 692 | std::vector<T> outputImage = std::vector<T>( |
| 693 | QuantizedVector<T>(outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), { |
| 694 | 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, 3.5f, |
| 695 | 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, 6.0f, |
| 696 | 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, |
| 697 | 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, |
| 698 | 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, 6.5f, |
| 699 | 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, |
| 700 | |
| 701 | -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, |
| 702 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 703 | -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, |
| 704 | -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, |
| 705 | -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, |
| 706 | -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, -0.5f, |
| 707 | |
| 708 | 8.0f, 8.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 709 | 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 710 | 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 711 | 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 712 | 10.0f, 10.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 713 | 8.0f, 8.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 714 | |
| 715 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 716 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 717 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 718 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 719 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 720 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f |
| 721 | })); |
| 722 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 723 | // Optionally apply bias to output image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 724 | if(biasEnabled) |
| 725 | { |
| 726 | ApplyBias(outputImage, outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), |
| 727 | biasV, biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset(), |
| 728 | outputWidth, outputHeight); |
| 729 | } |
| 730 | |
| 731 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 732 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputImage); |
| 733 | |
| 734 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 735 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 736 | |
| 737 | armnn::DepthwiseConvolution2dQueueDescriptor data; |
| 738 | armnn::WorkloadInfo info; |
| 739 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 740 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 741 | |
| 742 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 743 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 744 | |
| 745 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 746 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 747 | |
| 748 | data.m_Weight = &weightsTensor; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 749 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 750 | data.m_Parameters.m_StrideX = 2; |
| 751 | data.m_Parameters.m_StrideY = 1; |
| 752 | data.m_Parameters.m_PadLeft = 0; |
| 753 | data.m_Parameters.m_PadRight = 0; |
| 754 | data.m_Parameters.m_PadTop = 1; |
| 755 | data.m_Parameters.m_PadBottom = 1; |
| 756 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 757 | |
| 758 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); |
| 759 | inputHandle->Allocate(); |
| 760 | outputHandle->Allocate(); |
| 761 | |
| 762 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 763 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 764 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 765 | workload->Execute(); |
| 766 | |
| 767 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 768 | |
| 769 | return ret; |
| 770 | } |
| 771 | |
Nikhil Raj | cec6b65 | 2018-10-12 13:51:57 +0100 | [diff] [blame] | 772 | template<typename T, typename B> |
| 773 | LayerTestResult<T, 4> DepthwiseConvolution2dNhwcTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 774 | const boost::multi_array<T, 4>& input, |
| 775 | const boost::multi_array<T, 4>& kernel, |
| 776 | const boost::multi_array<B, 1>& bias, |
| 777 | const boost::multi_array<T, 4>& outputExpected, |
| 778 | float qScale, |
| 779 | int32_t qOffset, |
| 780 | uint32_t padLeft = 0, |
| 781 | uint32_t padTop = 0, |
| 782 | uint32_t padRight = 0, |
| 783 | uint32_t padBottom = 0, |
| 784 | uint32_t strideX = 1, |
| 785 | uint32_t strideY = 1) |
| 786 | { |
| 787 | unsigned int inputNum = boost::numeric_cast<unsigned int>(input.shape()[0]); |
| 788 | unsigned int inputChannels = boost::numeric_cast<unsigned int>(input.shape()[3]); |
| 789 | unsigned int inputHeight = boost::numeric_cast<unsigned int>(input.shape()[1]); |
| 790 | unsigned int inputWidth = boost::numeric_cast<unsigned int>(input.shape()[2]); |
| 791 | |
| 792 | unsigned int kernelChanMul = boost::numeric_cast<unsigned int>(kernel.shape()[0]); |
| 793 | unsigned int kernelChannels = boost::numeric_cast<unsigned int>(kernel.shape()[3]); |
| 794 | unsigned int kernelHeight = boost::numeric_cast<unsigned int>(kernel.shape()[1]); |
| 795 | unsigned int kernelWidth = boost::numeric_cast<unsigned int>(kernel.shape()[2]); |
| 796 | |
| 797 | unsigned int outputNum = boost::numeric_cast<unsigned int>(outputExpected.shape()[0]); |
| 798 | unsigned int outputChannels = boost::numeric_cast<unsigned int>(outputExpected.shape()[3]); |
| 799 | unsigned int outputHeight = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]); |
| 800 | unsigned int outputWidth = boost::numeric_cast<unsigned int>(outputExpected.shape()[2]); |
| 801 | |
| 802 | // Creates the tensors. |
| 803 | armnn::TensorInfo inputTensorInfo({inputNum, inputHeight, inputWidth, inputChannels}, armnn::GetDataType<T>()); |
| 804 | armnn::TensorInfo outputTensorInfo({outputNum, outputHeight, outputWidth, outputChannels}, |
| 805 | armnn::GetDataType<T>()); |
| 806 | armnn::TensorInfo kernelDesc({kernelChanMul, kernelHeight, kernelWidth, kernelChannels}, armnn::GetDataType<T>()); |
| 807 | armnn::TensorInfo biasDesc({static_cast<unsigned int>(bias.size())}, armnn::GetDataType<B>()); |
| 808 | |
| 809 | // Set quantization parameters if the requested type is a quantized type. |
| 810 | if (armnn::IsQuantizedType<T>()) |
| 811 | { |
| 812 | inputTensorInfo.SetQuantizationScale(qScale); |
| 813 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 814 | outputTensorInfo.SetQuantizationScale(qScale); |
| 815 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 816 | kernelDesc.SetQuantizationScale(qScale); |
| 817 | kernelDesc.SetQuantizationOffset(qOffset); |
| 818 | biasDesc.SetQuantizationScale(qScale*qScale); |
| 819 | biasDesc.SetQuantizationOffset(0); |
| 820 | } |
| 821 | |
| 822 | // Construct the input data. |
| 823 | std::vector<T> inputData; |
| 824 | inputData.assign(input.data(), input.data() + inputHeight*inputWidth*inputChannels); |
| 825 | auto batchedInput = MakeTensor<T, 4>(inputTensorInfo, inputData); |
| 826 | |
| 827 | // Construct the output data, with bias applied, as appropriate. |
| 828 | std::vector<T> outputData; |
| 829 | outputData.assign(outputExpected.data(), outputExpected.data() + outputHeight*outputWidth*outputChannels); |
| 830 | |
| 831 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 832 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputData); |
| 833 | |
| 834 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 835 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 836 | |
| 837 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 838 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 839 | |
| 840 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 841 | |
| 842 | armnn::DepthwiseConvolution2dQueueDescriptor data; |
| 843 | data.m_Weight = &weightsTensor; |
| 844 | data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs. |
| 845 | data.m_Parameters.m_StrideX = strideX; |
| 846 | data.m_Parameters.m_StrideY = strideY; |
| 847 | data.m_Parameters.m_PadLeft = padLeft; |
| 848 | data.m_Parameters.m_PadRight = padRight; |
| 849 | data.m_Parameters.m_PadTop = padTop; |
| 850 | data.m_Parameters.m_PadBottom = padBottom; |
| 851 | data.m_Parameters.m_DataLayout = armnn::DataLayout::NHWC; |
| 852 | |
| 853 | armnn::WorkloadInfo info; |
| 854 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 855 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 856 | |
| 857 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); |
| 858 | |
| 859 | inputHandle->Allocate(); |
| 860 | outputHandle->Allocate(); |
| 861 | |
| 862 | CopyDataToITensorHandle(inputHandle.get(), &batchedInput[0][0][0][0]); |
| 863 | |
| 864 | workloadFactory.Finalize(); |
| 865 | workload->Execute(); |
| 866 | |
| 867 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 868 | |
| 869 | return ret; |
| 870 | } |
| 871 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 872 | template<typename T> |
| 873 | LayerTestResult<T,4> Convolution1dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 874 | float qScale, |
| 875 | int32_t qOffset, |
| 876 | bool biasEnabled) |
| 877 | { |
| 878 | using B = typename FullyConnectedBiasTypeForInputType<T>::Type; |
| 879 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 880 | // Until we have a specialist 1D convolution layer, we can fake one using |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 881 | // 2D convolution with the final dimension set to 1. |
| 882 | // I don't anticipate this being particularly slow, given that convolution is implemented |
| 883 | // as a matrix multiplication, at which point dimension doesn't matter. |
| 884 | |
| 885 | unsigned int batchSize = 1; |
| 886 | unsigned int inputChannels = 2; |
| 887 | unsigned int outputChannels = 3; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 888 | unsigned int inputSize = 5; // The 1D size (could view as 'width' or 'height'). |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 889 | unsigned int kernelSize = 3; |
| 890 | unsigned int padSize = 2; |
| 891 | unsigned int stride = 1; |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 892 | unsigned int outputSize = 7; // (inputSize + 2 * padSize - kernelSize + 1) / stride. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 893 | |
| 894 | armnn::TensorInfo inputInfo({batchSize, inputChannels, inputSize, 1}, armnn::GetDataType<T>()); |
| 895 | armnn::TensorInfo outputInfo({batchSize, outputChannels, outputSize, 1}, armnn::GetDataType<T>()); |
| 896 | armnn::TensorInfo kernelInfo({outputChannels, inputChannels, kernelSize, 1}, armnn::GetDataType<T>()); |
| 897 | armnn::TensorInfo biasInfo({outputChannels}, armnn::GetDataType<B>()); |
| 898 | |
| 899 | // Set quantization parameters if the requested type is a quantized type. |
| 900 | if(armnn::IsQuantizedType<T>()) |
| 901 | { |
| 902 | inputInfo.SetQuantizationScale(qScale); |
| 903 | inputInfo.SetQuantizationOffset(qOffset); |
| 904 | outputInfo.SetQuantizationScale(qScale); |
| 905 | outputInfo.SetQuantizationOffset(qOffset); |
| 906 | kernelInfo.SetQuantizationScale(qScale); |
| 907 | kernelInfo.SetQuantizationOffset(qOffset); |
| 908 | biasInfo.SetQuantizationScale(inputInfo.GetQuantizationScale()*kernelInfo.GetQuantizationScale()); |
| 909 | biasInfo.SetQuantizationOffset(0); |
| 910 | } |
| 911 | |
| 912 | std::vector<T> inputData( |
| 913 | QuantizedVector<T>(inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(), { |
| 914 | 5.0f, -2.0f, 2.5f, 0.0f, 1.0f, |
| 915 | -3.0f, 3.2f, 5.0f, 2.0f, 3.0f, |
| 916 | })); |
| 917 | |
| 918 | std::vector<T> kernelData( |
| 919 | QuantizedVector<T>(kernelInfo.GetQuantizationScale(), kernelInfo.GetQuantizationOffset(), { |
| 920 | 1.0f, 0.0f, 0.0f, |
| 921 | 0.0f, 2.0f, -1.5f, |
| 922 | |
| 923 | 0.0f, 0.0f, 0.0f, |
| 924 | 0.2f, 0.2f, 0.2f, |
| 925 | |
| 926 | 0.5f, 0.0f, 0.5f, |
| 927 | 0.0f, -1.0f, 0.0f |
| 928 | })); |
| 929 | |
| 930 | std::vector<B> biasData( |
| 931 | QuantizedVector<B>(biasInfo.GetQuantizationScale(), biasInfo.GetQuantizationOffset(), { |
| 932 | 1.0f, 0.0f, 0.0f |
| 933 | })); |
| 934 | |
| 935 | std::vector<T> outputData( |
| 936 | QuantizedVector<T>(outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), { |
| 937 | 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, |
| 938 | -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, |
| 939 | 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 |
| 940 | })); |
| 941 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 942 | // Optionally apply bias to output image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 943 | if(biasEnabled) |
| 944 | { |
| 945 | ApplyBias(outputData, outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), |
| 946 | biasData, biasInfo.GetQuantizationScale(), biasInfo.GetQuantizationOffset(), |
| 947 | 1, outputSize); |
| 948 | } |
| 949 | |
| 950 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputInfo); |
| 951 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputInfo); |
| 952 | |
| 953 | armnn::Convolution2dQueueDescriptor data; |
| 954 | armnn::WorkloadInfo info; |
| 955 | armnn::ScopedCpuTensorHandle weightsTensor(kernelInfo); |
| 956 | armnn::ScopedCpuTensorHandle biasTensor(biasInfo); |
| 957 | |
| 958 | AllocateAndCopyDataToITensorHandle(&weightsTensor, kernelData.data()); |
| 959 | AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data()); |
| 960 | |
| 961 | AddInputToWorkload(data, info, inputInfo, inputHandle.get()); |
| 962 | AddOutputToWorkload(data, info, outputInfo, outputHandle.get()); |
| 963 | |
| 964 | data.m_Weight = &weightsTensor; |
| 965 | data.m_Bias = &biasTensor; |
| 966 | data.m_Parameters.m_StrideX = 1; |
| 967 | data.m_Parameters.m_StrideY = stride; |
| 968 | data.m_Parameters.m_PadLeft = 0; |
| 969 | data.m_Parameters.m_PadRight = 0; |
| 970 | data.m_Parameters.m_PadTop = padSize; |
| 971 | data.m_Parameters.m_PadBottom = padSize; |
| 972 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 973 | |
| 974 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvolution2d(data, info); |
| 975 | inputHandle->Allocate(); |
| 976 | outputHandle->Allocate(); |
| 977 | |
| 978 | CopyDataToITensorHandle(inputHandle.get(), inputData.data()); |
| 979 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 980 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 981 | workload->Execute(); |
| 982 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 983 | // Output |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 984 | LayerTestResult<T,4> ret(outputInfo); |
| 985 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 986 | ret.outputExpected = MakeTensor<T, 4>(outputInfo, outputData); |
| 987 | return ret; |
| 988 | } |
| 989 | |
| 990 | |
| 991 | |
| 992 | template<typename T> |
| 993 | LayerTestResult<T,4> CompareConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 994 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 995 | { |
| 996 | unsigned int inputHeight = 8; |
| 997 | unsigned int inputWidth = 16; |
| 998 | unsigned int inputChannels = 3; |
| 999 | unsigned int inputNum = 5; |
| 1000 | |
| 1001 | unsigned int kernelHeight = 3; |
| 1002 | unsigned int kernelWidth = 3; |
| 1003 | |
| 1004 | unsigned int strideX = 2; |
| 1005 | unsigned int strideY = 3; |
| 1006 | unsigned int padX = 1; |
| 1007 | unsigned int padY = 1; |
| 1008 | |
| 1009 | unsigned int outputNum = inputNum; |
| 1010 | unsigned int outputChannels = 2; |
| 1011 | unsigned int outputHeight = (inputHeight + 2 * padY - kernelHeight + strideY) / strideY; |
| 1012 | unsigned int outputWidth = (inputWidth + 2 * padX - kernelWidth + strideX) / strideX; |
| 1013 | |
| 1014 | armnn::TensorInfo inputTensorInfo; |
| 1015 | armnn::TensorInfo outputTensorInfo; |
| 1016 | armnn::TensorInfo kernelDesc; |
| 1017 | armnn::TensorInfo biasDesc; |
| 1018 | |
| 1019 | unsigned int inputShape[] = {inputNum, inputChannels, inputHeight, inputWidth}; |
| 1020 | unsigned int outputShape[] = {outputNum, outputChannels, outputHeight, outputWidth}; |
| 1021 | unsigned int kernelShape[] = {outputChannels, inputChannels, kernelHeight, kernelWidth}; |
| 1022 | unsigned int biasShape[] = {outputChannels}; |
| 1023 | |
| 1024 | inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::GetDataType<T>()); |
| 1025 | outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::GetDataType<T>()); |
| 1026 | kernelDesc = armnn::TensorInfo(4, kernelShape, armnn::GetDataType<T>()); |
| 1027 | biasDesc = armnn::TensorInfo(1, biasShape, armnn::GetDataType<T>()); |
| 1028 | |
| 1029 | LayerTestResult<T,4> ret(outputTensorInfo); |
| 1030 | |
| 1031 | auto input = MakeRandomTensor<T, 4>(inputTensorInfo, 124908); |
| 1032 | auto kernel = MakeRandomTensor<T, 4>(kernelDesc, 891234); |
| 1033 | auto bias = MakeRandomTensor<T, 1>(biasDesc, 1028); |
| 1034 | |
| 1035 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1036 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1037 | |
| 1038 | armnn::Convolution2dQueueDescriptor data; |
| 1039 | armnn::WorkloadInfo info; |
| 1040 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 1041 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 1042 | |
| 1043 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 1044 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 1045 | |
| 1046 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 1047 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1048 | data.m_Weight = &weightsTensor; |
| 1049 | data.m_Bias = &biasTensor; |
| 1050 | data.m_Parameters.m_StrideX = strideX; |
| 1051 | data.m_Parameters.m_StrideY = strideY; |
| 1052 | data.m_Parameters.m_PadLeft = padX; |
| 1053 | data.m_Parameters.m_PadRight = padX; |
| 1054 | data.m_Parameters.m_PadTop = padY; |
| 1055 | data.m_Parameters.m_PadBottom = padY; |
| 1056 | data.m_Parameters.m_BiasEnabled = true; |
| 1057 | |
| 1058 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1059 | std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1060 | |
| 1061 | armnn::Convolution2dQueueDescriptor refData = data; |
| 1062 | armnn::WorkloadInfo refInfo = info; |
| 1063 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); |
| 1064 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 1065 | |
| 1066 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvolution2d(data, info); |
| 1067 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateConvolution2d(refData, refInfo); |
| 1068 | |
| 1069 | outputHandleRef->Allocate(); |
| 1070 | inputHandleRef->Allocate(); |
| 1071 | |
| 1072 | inputHandle->Allocate(); |
| 1073 | outputHandle->Allocate(); |
| 1074 | |
| 1075 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 1076 | CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); |
| 1077 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1078 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1079 | workload->Execute(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1080 | refWorkloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1081 | workloadRef->Execute(); |
| 1082 | |
| 1083 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 1084 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 1085 | |
| 1086 | return ret; |
| 1087 | } |
| 1088 | |
| 1089 | template<typename T> |
| 1090 | LayerTestResult<T, 4> CompareDepthwiseConvolution2dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 1091 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 1092 | { |
| 1093 | unsigned int inputHeight = 8; |
| 1094 | unsigned int inputWidth = 16; |
| 1095 | unsigned int inputChannels = 3; |
| 1096 | unsigned int inputNum = 5; |
| 1097 | |
| 1098 | unsigned int kernelHeight = 3; |
| 1099 | unsigned int kernelWidth = 3; |
| 1100 | unsigned int channelMultiplier = 1; |
| 1101 | |
| 1102 | unsigned int strideX = 2; |
| 1103 | unsigned int strideY = 3; |
| 1104 | unsigned int padX = 1; |
| 1105 | unsigned int padY = 1; |
| 1106 | |
| 1107 | unsigned int outputNum = inputNum; |
| 1108 | unsigned int outputChannels = inputChannels * channelMultiplier; |
| 1109 | unsigned int outputHeight = (inputHeight + 2 * padY - kernelHeight + strideY) / strideY; |
| 1110 | unsigned int outputWidth = (inputWidth + 2 * padX - kernelWidth + strideX) / strideX; |
| 1111 | |
| 1112 | armnn::TensorInfo inputTensorInfo; |
| 1113 | armnn::TensorInfo outputTensorInfo; |
| 1114 | armnn::TensorInfo kernelDesc; |
| 1115 | armnn::TensorInfo biasDesc; |
| 1116 | |
| 1117 | unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; |
| 1118 | unsigned int outputShape[] = { outputNum, outputChannels, outputHeight, outputWidth }; |
| 1119 | unsigned int kernelShape[] = { channelMultiplier, inputChannels, kernelHeight, kernelWidth }; |
| 1120 | unsigned int biasShape[] = { outputChannels }; |
| 1121 | |
| 1122 | float inputsQScale = armnn::IsQuantizedType<T>() ? 1.0f : 0; |
| 1123 | float outputQScale = armnn::IsQuantizedType<T>() ? 2.0f : 0; |
| 1124 | int32_t qOffset = 0; |
| 1125 | |
| 1126 | inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::GetDataType<T>(), inputsQScale, qOffset); |
| 1127 | outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::GetDataType<T>(), outputQScale, qOffset); |
| 1128 | kernelDesc = armnn::TensorInfo(4, kernelShape, armnn::GetDataType<T>(), inputsQScale, qOffset); |
| 1129 | biasDesc = armnn::TensorInfo(1, biasShape, armnn::GetBiasDataType(armnn::GetDataType<T>()), inputsQScale, qOffset); |
| 1130 | |
| 1131 | LayerTestResult<T, 4> ret(outputTensorInfo); |
| 1132 | |
| 1133 | auto input = MakeRandomTensor<T, 4>(inputTensorInfo, 124908, 0.0f, 255.0f); |
| 1134 | auto kernel = MakeRandomTensor<T, 4>(kernelDesc, 891234, 0.0f, 255.0f); |
| 1135 | auto bias = MakeRandomTensor<typename FullyConnectedBiasTypeForInputType<T>::Type, 1>(biasDesc, 1028, 0.0f, 255.0f); |
| 1136 | |
| 1137 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1138 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1139 | |
| 1140 | armnn::DepthwiseConvolution2dQueueDescriptor data; |
| 1141 | armnn::WorkloadInfo info; |
| 1142 | armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc); |
| 1143 | armnn::ScopedCpuTensorHandle biasTensor(biasDesc); |
| 1144 | |
| 1145 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]); |
| 1146 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 1147 | |
| 1148 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 1149 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1150 | data.m_Weight = &weightsTensor; |
| 1151 | data.m_Bias = &biasTensor; |
| 1152 | data.m_Parameters.m_StrideX = strideX; |
| 1153 | data.m_Parameters.m_StrideY = strideY; |
| 1154 | data.m_Parameters.m_PadLeft = padX; |
| 1155 | data.m_Parameters.m_PadRight = padX; |
| 1156 | data.m_Parameters.m_PadTop = padY; |
| 1157 | data.m_Parameters.m_PadBottom = padY; |
| 1158 | data.m_Parameters.m_BiasEnabled = true; |
| 1159 | |
| 1160 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1161 | std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1162 | |
| 1163 | armnn::DepthwiseConvolution2dQueueDescriptor refData = data; |
| 1164 | armnn::WorkloadInfo refInfo = info; |
| 1165 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); |
| 1166 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 1167 | |
| 1168 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDepthwiseConvolution2d(data, info); |
| 1169 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateDepthwiseConvolution2d(refData, refInfo); |
| 1170 | |
| 1171 | outputHandleRef->Allocate(); |
| 1172 | inputHandleRef->Allocate(); |
| 1173 | |
| 1174 | inputHandle->Allocate(); |
| 1175 | outputHandle->Allocate(); |
| 1176 | |
| 1177 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 1178 | CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); |
| 1179 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1180 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1181 | workload->Execute(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1182 | refWorkloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1183 | workloadRef->Execute(); |
| 1184 | |
| 1185 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 1186 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 1187 | |
| 1188 | return ret; |
| 1189 | } |