Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 1 | // |
Matthew Sloyan | 21a6a1a | 2022-06-30 17:13:04 +0100 | [diff] [blame] | 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "RefUnidirectionalSequenceLstmWorkload.hpp" |
| 7 | #include "Activation.hpp" |
| 8 | #include "Encoders.hpp" |
| 9 | #include "Decoders.hpp" |
| 10 | #include "Lstm.hpp" |
| 11 | #include "LstmUtils.hpp" |
| 12 | #include "RefWorkloadUtils.hpp" |
| 13 | |
| 14 | #include <armnnUtils/Permute.hpp> |
| 15 | |
| 16 | namespace armnn |
| 17 | { |
| 18 | |
| 19 | RefUnidirectionalSequenceLstmWorkload::RefUnidirectionalSequenceLstmWorkload( |
| 20 | const UnidirectionalSequenceLstmQueueDescriptor& descriptor, |
| 21 | const WorkloadInfo& info) |
Finn Williams | 73c547d | 2022-02-15 20:47:34 +0000 | [diff] [blame] | 22 | : RefBaseWorkload<UnidirectionalSequenceLstmQueueDescriptor>(descriptor, info) |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 23 | , m_InputToInputWeightsTensor (AssignScopedTensorHandle(descriptor.m_InputToInputWeights)) |
| 24 | , m_InputToForgetWeightsTensor (AssignScopedTensorHandle(descriptor.m_InputToForgetWeights)) |
| 25 | , m_InputToCellWeightsTensor (AssignScopedTensorHandle(descriptor.m_InputToCellWeights)) |
| 26 | , m_InputToOutputWeightsTensor (AssignScopedTensorHandle(descriptor.m_InputToOutputWeights)) |
| 27 | , m_RecurrentToInputWeightsTensor (AssignScopedTensorHandle(descriptor.m_RecurrentToInputWeights)) |
| 28 | , m_RecurrentToForgetWeightsTensor(AssignScopedTensorHandle(descriptor.m_RecurrentToForgetWeights)) |
| 29 | , m_RecurrentToCellWeightsTensor (AssignScopedTensorHandle(descriptor.m_RecurrentToCellWeights)) |
| 30 | , m_RecurrentToOutputWeightsTensor(AssignScopedTensorHandle(descriptor.m_RecurrentToOutputWeights)) |
| 31 | , m_CellToInputWeightsTensor (AssignScopedTensorHandle(descriptor.m_CellToInputWeights)) |
| 32 | , m_CellToForgetWeightsTensor (AssignScopedTensorHandle(descriptor.m_CellToForgetWeights)) |
| 33 | , m_CellToOutputWeightsTensor (AssignScopedTensorHandle(descriptor.m_CellToOutputWeights)) |
| 34 | , m_InputGateBiasTensor (AssignScopedTensorHandle(descriptor.m_InputGateBias)) |
| 35 | , m_ForgetGateBiasTensor (AssignScopedTensorHandle(descriptor.m_ForgetGateBias)) |
| 36 | , m_CellBiasTensor (AssignScopedTensorHandle(descriptor.m_CellBias)) |
| 37 | , m_OutputGateBiasTensor (AssignScopedTensorHandle(descriptor.m_OutputGateBias)) |
| 38 | , m_ProjectionWeightsTensor (AssignScopedTensorHandle(descriptor.m_ProjectionWeights)) |
| 39 | , m_ProjectionBiasTensor (AssignScopedTensorHandle(descriptor.m_ProjectionBias)) |
| 40 | , m_InputLayerNormWeights (AssignScopedTensorHandle(descriptor.m_InputLayerNormWeights)) |
| 41 | , m_ForgetLayerNormWeights (AssignScopedTensorHandle(descriptor.m_ForgetLayerNormWeights)) |
| 42 | , m_CellLayerNormWeights (AssignScopedTensorHandle(descriptor.m_CellLayerNormWeights)) |
| 43 | , m_OutputLayerNormWeights (AssignScopedTensorHandle(descriptor.m_OutputLayerNormWeights)) |
| 44 | {} |
| 45 | |
| 46 | void RefUnidirectionalSequenceLstmWorkload::Execute() const |
| 47 | { |
| 48 | Execute(m_Data.m_Inputs, m_Data.m_Outputs); |
| 49 | } |
| 50 | |
Matthew Sloyan | 21a6a1a | 2022-06-30 17:13:04 +0100 | [diff] [blame] | 51 | void RefUnidirectionalSequenceLstmWorkload::ExecuteAsync(ExecutionData& executionData) |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 52 | { |
Matthew Sloyan | 21a6a1a | 2022-06-30 17:13:04 +0100 | [diff] [blame] | 53 | WorkingMemDescriptor* workingMemDescriptor = static_cast<WorkingMemDescriptor*>(executionData.m_Data); |
| 54 | Execute(workingMemDescriptor->m_Inputs, workingMemDescriptor->m_Outputs); |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 55 | } |
| 56 | |
| 57 | void RefUnidirectionalSequenceLstmWorkload::Execute(std::vector<ITensorHandle*> inputs, |
| 58 | std::vector<ITensorHandle*> outputs) const |
| 59 | { |
| 60 | TensorInfo inputInfo = GetTensorInfo(inputs[0]); |
| 61 | const TensorInfo& outputStateInfo = GetTensorInfo(inputs[1]); |
| 62 | const TensorInfo& cellStateInfo = GetTensorInfo(inputs[2]); |
Mike Kelly | 1299496 | 2022-04-21 11:57:09 +0100 | [diff] [blame] | 63 | TensorInfo outputStateOutInfo = GetTensorInfo(outputs[0]); |
| 64 | TensorInfo cellStateOutInfo = GetTensorInfo(outputs[1]); |
| 65 | TensorInfo outputInfo = GetTensorInfo(outputs[2]); |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 66 | TensorShape& inputShape = inputInfo.GetShape(); |
| 67 | TensorShape& outputShape= outputInfo.GetShape(); |
| 68 | auto inputTensor = reinterpret_cast<float*>(inputs[0]->Map()); |
| 69 | |
| 70 | if (!m_Data.m_Parameters.m_TimeMajor) |
| 71 | { |
| 72 | // Permute to time major |
| 73 | const PermutationVector& mappings = {1U, 0U, 2U}; |
| 74 | std::vector<float> inputValue(inputTensor, inputTensor + inputInfo.GetNumElements()); |
| 75 | inputShape = armnnUtils::Permuted(inputInfo.GetShape(), mappings); |
| 76 | inputInfo.SetShape(inputShape); |
| 77 | armnnUtils::Permute(inputShape, mappings, inputValue.data(), inputTensor, sizeof(float)); |
| 78 | |
| 79 | outputShape = armnnUtils::Permuted(outputInfo.GetShape(), mappings); |
| 80 | outputInfo.SetShape(outputShape); |
| 81 | } |
| 82 | unsigned int maxTime = inputShape[0]; |
| 83 | unsigned int batchSize = inputShape[1]; |
| 84 | unsigned int outputSize = outputShape[2]; |
| 85 | unsigned int inputSize = inputShape[2]; |
| 86 | |
| 87 | TensorInfo scratchInfo = outputInfo; |
| 88 | scratchInfo.SetShape({batchSize, cellStateInfo.GetShape()[1]}); |
| 89 | |
| 90 | std::vector<float> inputGateScratchBuffer; |
| 91 | std::vector<float> cellScratchBuffer(scratchInfo.GetNumElements(), 0.); |
| 92 | std::vector<float> forgetGateScratchBuffer(scratchInfo.GetNumElements(), 0.); |
| 93 | std::vector<float> outputGateScratchBuffer(scratchInfo.GetNumElements(), 0.); |
| 94 | |
| 95 | std::vector<float> outputStateOutBuffer(outputStateInfo.GetNumElements(), 0.); |
| 96 | std::vector<float> cellStateOutBuffer(cellStateInfo.GetNumElements(), 0.); |
| 97 | |
| 98 | void* outputStateOutData = outputStateOutBuffer.data(); |
| 99 | void* cellStateOutData = cellStateOutBuffer.data(); |
| 100 | |
| 101 | std::unique_ptr<Encoder<float>> inputGateScratch; |
| 102 | std::unique_ptr<Encoder<float>> cellScratch = MakeEncoder<float>(scratchInfo, cellScratchBuffer.data()); |
| 103 | std::unique_ptr<Encoder<float>> forgetGateScratch = MakeEncoder<float>(scratchInfo, forgetGateScratchBuffer.data()); |
| 104 | std::unique_ptr<Encoder<float>> outputGateScratch = MakeEncoder<float>(scratchInfo, outputGateScratchBuffer.data()); |
| 105 | |
| 106 | std::unique_ptr<Decoder<float>> inputGateScratchDecoder; |
| 107 | std::unique_ptr<Decoder<float>> cellScratchDecoder = MakeDecoder<float>(scratchInfo, cellScratchBuffer.data()); |
| 108 | std::unique_ptr<Decoder<float>> forgetGateScratchDecoder = MakeDecoder<float>(scratchInfo, |
| 109 | forgetGateScratchBuffer.data()); |
| 110 | std::unique_ptr<Decoder<float>> outputGateScratchDecoder = MakeDecoder<float>(scratchInfo, |
| 111 | outputGateScratchBuffer.data()); |
| 112 | |
| 113 | const bool useCifg = m_Data.m_Parameters.m_CifgEnabled; |
| 114 | const bool usePeephole = m_Data.m_Parameters.m_PeepholeEnabled; |
| 115 | const bool useLayerNorm = m_Data.m_Parameters.m_LayerNormEnabled; |
| 116 | |
| 117 | if (!useCifg) |
| 118 | { |
| 119 | inputGateScratchBuffer.resize(scratchInfo.GetNumElements(), 0.); |
| 120 | inputGateScratch = MakeEncoder<float>(scratchInfo, inputGateScratchBuffer.data()); |
| 121 | inputGateScratchDecoder = MakeDecoder<float>(scratchInfo, inputGateScratchBuffer.data()); |
| 122 | } |
| 123 | |
| 124 | std::unique_ptr<Encoder<float>> outputStateOut = MakeEncoder<float>(outputStateInfo, outputStateOutData); |
| 125 | std::unique_ptr<Encoder<float>> cellStateOut = MakeEncoder<float>(cellStateInfo, cellStateOutData); |
| 126 | std::unique_ptr<Decoder<float>> cellStateOutDecoder = MakeDecoder<float>(cellStateInfo, cellStateOutData); |
| 127 | |
| 128 | TensorInfo lstmInputInfo = inputInfo; |
| 129 | TensorShape batchInputShape = TensorShape({batchSize, inputSize}); |
| 130 | lstmInputInfo.SetShape(batchInputShape); |
| 131 | |
| 132 | TensorInfo lstmOutputInfo = outputInfo; |
| 133 | lstmOutputInfo.SetShape({batchSize, outputSize}); |
| 134 | |
| 135 | const TensorShape& inputToOutputWeightsShape = m_InputToOutputWeightsTensor->GetShape(); |
| 136 | const TensorShape& recurrentToOutputWeightsShape = m_RecurrentToOutputWeightsTensor->GetShape(); |
| 137 | unsigned int nOutput = recurrentToOutputWeightsShape[1]; |
| 138 | auto outputStateInData = inputs[1]->Map(); |
| 139 | std::unique_ptr<Decoder<float>> outputStateIn = MakeDecoder<float>(outputStateInfo, outputStateInData); |
| 140 | |
| 141 | auto cellStateInData = inputs[2]->Map(); |
| 142 | std::unique_ptr<Decoder<float>> cellStateIn = MakeDecoder<float>(cellStateInfo, cellStateInData); |
| 143 | |
| 144 | auto currentInputData = reinterpret_cast<float*>(inputs[0]->Map()); |
| 145 | std::unique_ptr<Decoder<float>> inputData = MakeDecoder<float>(lstmInputInfo, currentInputData); |
Mike Kelly | 1299496 | 2022-04-21 11:57:09 +0100 | [diff] [blame] | 146 | auto currentOutputData = reinterpret_cast<float*>(outputs[2]->Map()); |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 147 | std::unique_ptr<Encoder<float>> output = MakeEncoder<float>(lstmOutputInfo, currentOutputData); |
| 148 | std::unique_ptr<Decoder<float>> outputDecoder = MakeDecoder<float>(lstmOutputInfo, currentOutputData); |
| 149 | |
| 150 | std::unique_ptr<Decoder<float>> inputToInputWeightsTensor; |
| 151 | std::unique_ptr<Decoder<float>> inputToForgetWeightsTensor = MakeDecoder<float>( |
| 152 | m_InputToForgetWeightsTensor->GetTensorInfo(), m_InputToForgetWeightsTensor->GetConstTensor<void>()); |
| 153 | std::unique_ptr<Decoder<float>> inputToCellWeightsTensor = MakeDecoder<float>( |
| 154 | m_InputToCellWeightsTensor->GetTensorInfo(), m_InputToCellWeightsTensor->GetConstTensor<void>()); |
| 155 | std::unique_ptr<Decoder<float>> inputToOutputWeightsTensor = MakeDecoder<float>( |
| 156 | m_InputToOutputWeightsTensor->GetTensorInfo(), m_InputToOutputWeightsTensor->GetConstTensor<void>()); |
| 157 | |
| 158 | std::unique_ptr<Decoder<float>> recurrentToInputWeightsTensor; |
| 159 | std::unique_ptr<Decoder<float>> recurrentToForgetWeightsTensor = MakeDecoder<float>( |
| 160 | m_RecurrentToForgetWeightsTensor->GetTensorInfo(), m_RecurrentToForgetWeightsTensor->GetConstTensor<void>()); |
| 161 | std::unique_ptr<Decoder<float>> recurrentToCellWeightsTensor = MakeDecoder<float>( |
| 162 | m_RecurrentToCellWeightsTensor->GetTensorInfo(), m_RecurrentToCellWeightsTensor->GetConstTensor<void>()); |
| 163 | std::unique_ptr<Decoder<float>> recurrentToOutputWeightsTensor = MakeDecoder<float>( |
| 164 | m_RecurrentToOutputWeightsTensor->GetTensorInfo(), m_RecurrentToOutputWeightsTensor->GetConstTensor<void>()); |
| 165 | |
| 166 | std::unique_ptr<Decoder<float>> inputGateBiasTensor; |
| 167 | std::unique_ptr<Decoder<float>> forgetGateBiasTensor = MakeDecoder<float>( |
| 168 | m_ForgetGateBiasTensor->GetTensorInfo(), m_ForgetGateBiasTensor->GetConstTensor<void>()); |
| 169 | std::unique_ptr<Decoder<float>> cellBiasTensor = MakeDecoder<float>( |
| 170 | m_CellBiasTensor->GetTensorInfo(), m_CellBiasTensor->GetConstTensor<void>()); |
| 171 | std::unique_ptr<Decoder<float>> outputGateBiasTensor = MakeDecoder<float>( |
| 172 | m_OutputGateBiasTensor->GetTensorInfo(), m_OutputGateBiasTensor->GetConstTensor<void>()); |
| 173 | |
| 174 | std::unique_ptr<Decoder<float>> cellToInputWeightsTensor; |
| 175 | std::unique_ptr<Decoder<float>> cellToForgetWeightsTensor; |
| 176 | std::unique_ptr<Decoder<float>> cellToOutputWeightsTensor; |
| 177 | |
| 178 | std::unique_ptr<Decoder<float>> projectionWeightsTensor; |
| 179 | std::unique_ptr<Decoder<float>> projectionBiasTensor; |
| 180 | |
| 181 | std::unique_ptr<Decoder<float>> inputLayerNormWeights; |
| 182 | std::unique_ptr<Decoder<float>> forgetLayerNormWeights; |
| 183 | std::unique_ptr<Decoder<float>> cellLayerNormWeights; |
| 184 | std::unique_ptr<Decoder<float>> outputLayerNormWeights; |
| 185 | |
| 186 | if (useLayerNorm) |
| 187 | { |
| 188 | if (!useCifg) |
| 189 | { |
| 190 | inputLayerNormWeights = MakeDecoder<float>( |
| 191 | m_InputLayerNormWeights->GetTensorInfo(), m_InputLayerNormWeights->GetConstTensor<void>()); |
| 192 | } |
| 193 | forgetLayerNormWeights = MakeDecoder<float>( |
| 194 | m_ForgetLayerNormWeights->GetTensorInfo(), m_ForgetLayerNormWeights->GetConstTensor<void>()); |
| 195 | cellLayerNormWeights = MakeDecoder<float>( |
| 196 | m_CellLayerNormWeights->GetTensorInfo(), m_CellLayerNormWeights->GetConstTensor<void>()); |
| 197 | outputLayerNormWeights = MakeDecoder<float>( |
| 198 | m_OutputLayerNormWeights->GetTensorInfo(), m_OutputLayerNormWeights->GetConstTensor<void>()); |
| 199 | } |
| 200 | |
| 201 | if (!useCifg) |
| 202 | { |
| 203 | inputToInputWeightsTensor = MakeDecoder<float>( |
| 204 | m_InputToInputWeightsTensor->GetTensorInfo(), m_InputToInputWeightsTensor->GetConstTensor<void>()); |
| 205 | inputGateBiasTensor = MakeDecoder<float>( |
| 206 | m_InputGateBiasTensor->GetTensorInfo(), m_InputGateBiasTensor->GetConstTensor<void>()); |
| 207 | recurrentToInputWeightsTensor = MakeDecoder<float>( |
| 208 | m_RecurrentToInputWeightsTensor->GetTensorInfo(), m_RecurrentToInputWeightsTensor->GetConstTensor<void>()); |
| 209 | } |
| 210 | |
| 211 | if (usePeephole) |
| 212 | { |
| 213 | cellToForgetWeightsTensor = MakeDecoder<float>( |
| 214 | m_CellToForgetWeightsTensor->GetTensorInfo(), m_CellToForgetWeightsTensor->GetConstTensor<void>()); |
| 215 | cellToOutputWeightsTensor = MakeDecoder<float>( |
| 216 | m_CellToOutputWeightsTensor->GetTensorInfo(), m_CellToOutputWeightsTensor->GetConstTensor<void>()); |
| 217 | } |
| 218 | |
| 219 | if (!useCifg && usePeephole) |
| 220 | { |
| 221 | cellToInputWeightsTensor = MakeDecoder<float>( |
| 222 | m_CellToInputWeightsTensor->GetTensorInfo(), m_CellToInputWeightsTensor->GetConstTensor<void>()); |
| 223 | } |
| 224 | |
| 225 | if (m_Data.m_Parameters.m_ProjectionEnabled) |
| 226 | { |
| 227 | projectionWeightsTensor = MakeDecoder<float>( |
| 228 | m_ProjectionWeightsTensor->GetTensorInfo(), m_ProjectionWeightsTensor->GetConstTensor<void>()); |
| 229 | if (m_ProjectionBiasTensor) |
| 230 | { |
| 231 | projectionBiasTensor = MakeDecoder<float>( |
| 232 | m_ProjectionBiasTensor->GetTensorInfo(), m_ProjectionBiasTensor->GetConstTensor<void>()); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | unsigned int batchInputSize = batchSize * inputSize; |
| 237 | unsigned int batchOutputSize = batchSize * nOutput; |
| 238 | |
| 239 | for (unsigned int t = 0; t < maxTime; ++t) |
| 240 | { |
| 241 | LstmImpl(m_Data.m_Parameters, |
| 242 | lstmInputInfo, |
| 243 | lstmOutputInfo, |
| 244 | inputToOutputWeightsShape, |
| 245 | recurrentToOutputWeightsShape, |
| 246 | inputData, |
| 247 | outputStateIn, |
| 248 | cellStateIn, |
| 249 | outputStateOut, |
| 250 | cellStateOut, |
| 251 | output, |
| 252 | cellStateOutDecoder, |
| 253 | outputDecoder, |
| 254 | inputToInputWeightsTensor, |
| 255 | inputToForgetWeightsTensor, |
| 256 | inputToCellWeightsTensor, |
| 257 | inputToOutputWeightsTensor, |
| 258 | recurrentToInputWeightsTensor, |
| 259 | recurrentToForgetWeightsTensor, |
| 260 | recurrentToCellWeightsTensor, |
| 261 | recurrentToOutputWeightsTensor, |
| 262 | cellToInputWeightsTensor, |
| 263 | cellToForgetWeightsTensor, |
| 264 | cellToOutputWeightsTensor, |
| 265 | inputGateBiasTensor, |
| 266 | forgetGateBiasTensor, |
| 267 | cellBiasTensor, |
| 268 | outputGateBiasTensor, |
| 269 | projectionWeightsTensor, |
| 270 | projectionBiasTensor, |
| 271 | inputLayerNormWeights, |
| 272 | forgetLayerNormWeights, |
| 273 | cellLayerNormWeights, |
| 274 | outputLayerNormWeights, |
| 275 | inputGateScratch, |
| 276 | cellScratch, |
| 277 | forgetGateScratch, |
| 278 | outputGateScratch, |
| 279 | inputGateScratchDecoder, |
| 280 | cellScratchDecoder, |
| 281 | forgetGateScratchDecoder, |
| 282 | outputGateScratchDecoder, |
| 283 | m_LayerNormEpsilon); |
| 284 | |
| 285 | currentInputData += batchInputSize; |
| 286 | inputData = MakeDecoder<float>(lstmInputInfo, currentInputData); |
| 287 | currentOutputData += batchOutputSize; |
| 288 | output = MakeEncoder<float>(lstmOutputInfo, currentOutputData); |
| 289 | outputDecoder = MakeDecoder<float>(lstmOutputInfo, currentOutputData); |
| 290 | |
| 291 | // Assign output state out to the next output state in |
| 292 | outputStateIn = MakeDecoder<float>(outputStateInfo, outputStateOutData); |
| 293 | |
| 294 | // Assign cell state out to the next cell state in |
| 295 | cellStateIn = MakeDecoder<float>(cellStateInfo, cellStateOutData); |
| 296 | } |
| 297 | |
| 298 | if (!m_Data.m_Parameters.m_TimeMajor) |
| 299 | { |
| 300 | // Permute Output back to batch major |
| 301 | const PermutationVector& mappings = {1U, 0U, 2U}; |
Mike Kelly | 1299496 | 2022-04-21 11:57:09 +0100 | [diff] [blame] | 302 | auto outputData = reinterpret_cast<float*>(outputs[2]->Map()); |
Narumol Prangnawarat | e5339e7 | 2021-07-28 17:33:28 +0100 | [diff] [blame] | 303 | std::vector<float> outputValue(outputData, outputData + outputInfo.GetNumElements()); |
| 304 | outputShape = armnnUtils::Permuted(outputInfo.GetShape(), mappings); |
| 305 | outputInfo.SetShape(outputShape); |
| 306 | armnnUtils::Permute(outputShape, mappings, outputValue.data(), outputData, sizeof(float)); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | } //namespace armnn |