Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 1 | // |
Ryan OShea | 238ecd9 | 2023-03-07 11:44:23 +0000 | [diff] [blame] | 2 | // Copyright © 2021, 2023 Arm Ltd and Contributors. All rights reserved. |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include "TestUtils.hpp" |
| 9 | |
| 10 | #include <armnn_delegate.hpp> |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 11 | #include <DelegateTestInterpreter.hpp> |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 12 | |
| 13 | #include <flatbuffers/flatbuffers.h> |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 14 | #include <tensorflow/lite/kernels/register.h> |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 15 | #include <tensorflow/lite/version.h> |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 16 | |
| 17 | #include <schema_generated.h> |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 18 | |
| 19 | #include <doctest/doctest.h> |
| 20 | |
| 21 | namespace |
| 22 | { |
| 23 | |
| 24 | template <typename T> |
| 25 | std::vector<char> CreateLstmTfLiteModel(tflite::TensorType tensorType, |
| 26 | int32_t batchSize, |
| 27 | int32_t inputSize, |
| 28 | int32_t outputSize, |
| 29 | int32_t numUnits, |
| 30 | bool hasInputToInputWeights, |
| 31 | const std::vector<T>& inputToInputWeights, |
| 32 | const std::vector<T>& inputToForgetWeights, |
| 33 | const std::vector<T>& inputToCellWeights, |
| 34 | const std::vector<T>& inputToOutputWeights, |
| 35 | bool hasRecurrentToInputWeights, |
| 36 | const std::vector<T>& recurrentToInputWeights, |
| 37 | const std::vector<T>& recurrentToForgetWeights, |
| 38 | const std::vector<T>& recurrentToCellWeights, |
| 39 | const std::vector<T>& recurrentToOutputWeights, |
| 40 | bool hasCellToInputWeights, |
| 41 | const std::vector<T>& cellToInputWeights, |
| 42 | bool hasCellToForgetWeights, |
| 43 | const std::vector<T>& cellToForgetWeights, |
| 44 | bool hasCellToOutputWeights, |
| 45 | const std::vector<T>& cellToOutputWeights, |
| 46 | bool hasInputGateBias, |
| 47 | const std::vector<T>& inputGateBias, |
| 48 | const std::vector<T>& forgetGateBias, |
| 49 | const std::vector<T>& cellBias, |
| 50 | const std::vector<T>& outputGateBias, |
| 51 | bool hasProjectionWeights, |
| 52 | const std::vector<T>& projectionWeights, |
| 53 | bool hasProjectionBias, |
| 54 | const std::vector<T>& projectionBias, |
| 55 | bool hasInputLayerNormWeights, |
| 56 | const std::vector<T>& inputLayerNormWeights, |
| 57 | bool hasForgetLayerNormWeights, |
| 58 | const std::vector<T>& forgetLayerNormWeights, |
| 59 | bool hasCellLayerNormWeights, |
| 60 | const std::vector<T>& cellLayerNormWeights, |
| 61 | bool hasOutputLayerNormWeights, |
| 62 | const std::vector<T>& outputLayerNormWeights, |
| 63 | tflite::ActivationFunctionType activationFunction, |
| 64 | float clippingThresCell, |
| 65 | float clippingThresProj, |
| 66 | float quantScale = 1.0f, |
| 67 | int quantOffset = 0, |
| 68 | float outputQuantScale = 2.0f, |
| 69 | int outputQuantOffset = 0) |
| 70 | { |
| 71 | |
| 72 | std::vector <int32_t> tensorInfo0 {}; |
| 73 | std::vector <int32_t> tensorInfo4 {numUnits}; |
| 74 | std::vector <int32_t> tensorInfo8 {numUnits, static_cast<int32_t>(2)}; |
| 75 | std::vector <int32_t> tensorInfo16 {numUnits, static_cast<int32_t>(4)}; |
| 76 | |
| 77 | std::vector<int32_t> inputShape {batchSize , inputSize}; |
| 78 | std::vector<int32_t> outputShape {batchSize , outputSize}; |
| 79 | |
| 80 | std::vector<int32_t> outputStateInDimensions{batchSize, outputSize}; |
| 81 | std::vector<int32_t> cellStateInDimensions{batchSize, numUnits}; |
| 82 | |
| 83 | std::vector<int> operatorInputs; |
| 84 | using namespace tflite; |
| 85 | flatbuffers::FlatBufferBuilder flatBufferBuilder; |
| 86 | std::vector<flatbuffers::Offset<tflite::Buffer>> buffers; |
| 87 | std::vector<flatbuffers::Offset<Tensor>> tensors; |
| 88 | |
| 89 | auto quantizationParameters = |
| 90 | CreateQuantizationParameters(flatBufferBuilder, |
| 91 | 0, |
| 92 | 0, |
| 93 | flatBufferBuilder.CreateVector<float>({ quantScale }), |
| 94 | flatBufferBuilder.CreateVector<int64_t>({ quantOffset })); |
| 95 | |
| 96 | auto outputQuantizationParameters = |
| 97 | CreateQuantizationParameters(flatBufferBuilder, |
| 98 | 0, |
| 99 | 0, |
| 100 | flatBufferBuilder.CreateVector<float>({ outputQuantScale }), |
| 101 | flatBufferBuilder.CreateVector<int64_t>({ outputQuantOffset })); |
| 102 | |
Ryan OShea | 238ecd9 | 2023-03-07 11:44:23 +0000 | [diff] [blame] | 103 | buffers.push_back(CreateBuffer(flatBufferBuilder)); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 104 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 105 | flatBufferBuilder.CreateVector<int32_t>(inputShape.data(), |
| 106 | inputShape.size()), |
| 107 | tensorType, |
| 108 | buffers.size() - 1, |
| 109 | flatBufferBuilder.CreateString("input_0"), |
| 110 | quantizationParameters)); |
| 111 | operatorInputs.push_back(buffers.size() - 1); |
| 112 | |
| 113 | if (hasInputToInputWeights) |
| 114 | { |
| 115 | buffers.push_back( |
| 116 | CreateBuffer(flatBufferBuilder, |
| 117 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToInputWeights.data()), |
| 118 | sizeof(T) * inputToInputWeights.size()))); |
| 119 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 120 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(), |
| 121 | tensorInfo8.size()), |
| 122 | tensorType, |
| 123 | buffers.size() - 1, |
| 124 | flatBufferBuilder.CreateString("inputToInputWeights"), |
| 125 | outputQuantizationParameters)); |
| 126 | operatorInputs.push_back(buffers.size() - 1); |
| 127 | } |
| 128 | else |
| 129 | { |
| 130 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 131 | } |
| 132 | |
| 133 | buffers.push_back( |
| 134 | CreateBuffer(flatBufferBuilder, |
| 135 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToForgetWeights.data()), |
| 136 | sizeof(T) * inputToForgetWeights.size()))); |
| 137 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 138 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(), |
| 139 | tensorInfo8.size()), |
| 140 | tensorType, |
| 141 | buffers.size() - 1, |
| 142 | flatBufferBuilder.CreateString("inputToForgetWeights"), |
| 143 | outputQuantizationParameters)); |
| 144 | operatorInputs.push_back(buffers.size() - 1); |
| 145 | |
| 146 | buffers.push_back( |
| 147 | CreateBuffer(flatBufferBuilder, |
| 148 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToCellWeights.data()), |
| 149 | sizeof(T) * inputToCellWeights.size()))); |
| 150 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 151 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(), |
| 152 | tensorInfo8.size()), |
| 153 | tensorType, |
| 154 | buffers.size() - 1, |
| 155 | flatBufferBuilder.CreateString("inputToCellWeights"), |
| 156 | outputQuantizationParameters)); |
| 157 | operatorInputs.push_back(buffers.size() - 1); |
| 158 | |
| 159 | buffers.push_back( |
| 160 | CreateBuffer(flatBufferBuilder, |
| 161 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(inputToOutputWeights.data()), |
| 162 | sizeof(T) * inputToOutputWeights.size()))); |
| 163 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 164 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo8.data(), |
| 165 | tensorInfo8.size()), |
| 166 | tensorType, |
| 167 | buffers.size() - 1, |
| 168 | flatBufferBuilder.CreateString("inputToOutputWeights"), |
| 169 | outputQuantizationParameters)); |
| 170 | operatorInputs.push_back(buffers.size() - 1); |
| 171 | |
| 172 | if (hasRecurrentToInputWeights) |
| 173 | { |
| 174 | buffers.push_back(CreateBuffer( |
| 175 | flatBufferBuilder, |
| 176 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(recurrentToInputWeights.data()), |
| 177 | sizeof(T) * recurrentToInputWeights.size()))); |
| 178 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 179 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(), |
| 180 | tensorInfo16.size()), |
| 181 | tensorType, |
| 182 | buffers.size() - 1, |
| 183 | flatBufferBuilder.CreateString("recurrentToInputWeights"), |
| 184 | outputQuantizationParameters)); |
| 185 | operatorInputs.push_back(buffers.size() - 1); |
| 186 | } |
| 187 | else |
| 188 | { |
| 189 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 190 | } |
| 191 | |
| 192 | buffers.push_back( |
| 193 | CreateBuffer(flatBufferBuilder, |
| 194 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToForgetWeights.data()), |
| 195 | sizeof(T) * recurrentToForgetWeights.size()))); |
| 196 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 197 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(), |
| 198 | tensorInfo16.size()), |
| 199 | tensorType, |
| 200 | buffers.size() - 1, |
| 201 | flatBufferBuilder.CreateString("recurrentToForgetWeights"), |
| 202 | outputQuantizationParameters)); |
| 203 | operatorInputs.push_back(buffers.size() - 1); |
| 204 | |
| 205 | buffers.push_back( |
| 206 | CreateBuffer(flatBufferBuilder, |
| 207 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToCellWeights.data()), |
| 208 | sizeof(T) * recurrentToCellWeights.size()))); |
| 209 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 210 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(), |
| 211 | tensorInfo16.size()), |
| 212 | tensorType, |
| 213 | buffers.size() - 1, |
| 214 | flatBufferBuilder.CreateString("recurrentToCellWeights"), |
| 215 | outputQuantizationParameters)); |
| 216 | operatorInputs.push_back(buffers.size() - 1); |
| 217 | |
| 218 | buffers.push_back( |
| 219 | CreateBuffer(flatBufferBuilder, |
| 220 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(recurrentToOutputWeights.data()), |
| 221 | sizeof(T) * recurrentToOutputWeights.size()))); |
| 222 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 223 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo16.data(), |
| 224 | tensorInfo16.size()), |
| 225 | tensorType, |
| 226 | buffers.size() - 1 , |
| 227 | flatBufferBuilder.CreateString("recurrentToOutputWeights"), |
| 228 | outputQuantizationParameters)); |
| 229 | operatorInputs.push_back(buffers.size() - 1); |
| 230 | |
| 231 | if (hasCellToInputWeights) |
| 232 | { |
| 233 | buffers.push_back( |
| 234 | CreateBuffer(flatBufferBuilder, |
| 235 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToInputWeights.data()), |
| 236 | sizeof(T) * cellToInputWeights.size()))); |
| 237 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 238 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 239 | tensorInfo4.size()), |
| 240 | tensorType, |
| 241 | buffers.size() - 1, |
| 242 | flatBufferBuilder.CreateString("cellToInputWeights"), |
| 243 | outputQuantizationParameters)); |
| 244 | operatorInputs.push_back(buffers.size() - 1); |
| 245 | } |
| 246 | else |
| 247 | { |
| 248 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 249 | } |
| 250 | |
| 251 | if (hasCellToForgetWeights) |
| 252 | { |
| 253 | buffers.push_back( |
| 254 | CreateBuffer(flatBufferBuilder, |
| 255 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToForgetWeights.data()), |
| 256 | sizeof(T) * cellToForgetWeights.size()))); |
| 257 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 258 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 259 | tensorInfo4.size()), |
| 260 | tensorType, |
| 261 | buffers.size() - 1, |
| 262 | flatBufferBuilder.CreateString("cellToForgetWeights"), |
| 263 | outputQuantizationParameters)); |
| 264 | operatorInputs.push_back(buffers.size() - 1); |
| 265 | } |
| 266 | else |
| 267 | { |
| 268 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 269 | } |
| 270 | |
| 271 | if (hasCellToOutputWeights) |
| 272 | { |
| 273 | buffers.push_back( |
| 274 | CreateBuffer(flatBufferBuilder, |
| 275 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(cellToOutputWeights.data()), |
| 276 | sizeof(T) * cellToOutputWeights.size()))); |
| 277 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 278 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 279 | tensorInfo4.size()), |
| 280 | tensorType, |
| 281 | buffers.size() - 1, |
| 282 | flatBufferBuilder.CreateString("cellToOutputWeights"), |
| 283 | outputQuantizationParameters)); |
| 284 | operatorInputs.push_back(buffers.size() - 1); |
| 285 | } |
| 286 | else |
| 287 | { |
| 288 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 289 | } |
| 290 | |
| 291 | if (hasInputGateBias) |
| 292 | { |
| 293 | buffers.push_back( |
| 294 | CreateBuffer(flatBufferBuilder, |
| 295 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(inputGateBias.data()), |
| 296 | sizeof(T) * inputGateBias.size()))); |
| 297 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 298 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 299 | tensorInfo4.size()), |
| 300 | tensorType, |
| 301 | buffers.size() - 1, |
| 302 | flatBufferBuilder.CreateString("inputGateBias"), |
| 303 | outputQuantizationParameters)); |
| 304 | operatorInputs.push_back(buffers.size() - 1); |
| 305 | } |
| 306 | else |
| 307 | { |
| 308 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 309 | } |
| 310 | |
| 311 | buffers.push_back( |
| 312 | CreateBuffer(flatBufferBuilder, |
| 313 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(forgetGateBias.data()), |
| 314 | sizeof(T) * forgetGateBias.size()))); |
| 315 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 316 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 317 | tensorInfo4.size()), |
| 318 | tensorType, |
| 319 | buffers.size() - 1, |
| 320 | flatBufferBuilder.CreateString("forgetGateBias"), |
| 321 | outputQuantizationParameters)); |
| 322 | operatorInputs.push_back(buffers.size() - 1); |
| 323 | |
| 324 | buffers.push_back( |
| 325 | CreateBuffer(flatBufferBuilder, |
| 326 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(cellBias.data()), |
| 327 | sizeof(T) * cellBias.size()))); |
| 328 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 329 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 330 | tensorInfo4.size()), |
| 331 | tensorType, |
| 332 | buffers.size() - 1, |
| 333 | flatBufferBuilder.CreateString("cellBias"), |
| 334 | outputQuantizationParameters)); |
| 335 | operatorInputs.push_back(buffers.size() - 1); |
| 336 | |
| 337 | buffers.push_back( |
| 338 | CreateBuffer(flatBufferBuilder, |
| 339 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(outputGateBias.data()), |
| 340 | sizeof(T) * outputGateBias.size()))); |
| 341 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 342 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 343 | tensorInfo4.size()), |
| 344 | tensorType, |
| 345 | buffers.size() - 1, |
| 346 | flatBufferBuilder.CreateString("outputGateBias"), |
| 347 | outputQuantizationParameters)); |
| 348 | operatorInputs.push_back(buffers.size() - 1); |
| 349 | |
| 350 | if (hasProjectionWeights) |
| 351 | { |
| 352 | buffers.push_back( |
| 353 | CreateBuffer(flatBufferBuilder, |
| 354 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(projectionWeights.data()), |
| 355 | sizeof(T) * projectionWeights.size()))); |
| 356 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 357 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 358 | tensorInfo4.size()), |
| 359 | tensorType, |
| 360 | buffers.size() - 1, |
| 361 | flatBufferBuilder.CreateString("outputGateBias"), |
| 362 | outputQuantizationParameters)); |
| 363 | operatorInputs.push_back(buffers.size() - 1); |
| 364 | } |
| 365 | else |
| 366 | { |
| 367 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 368 | } |
| 369 | |
| 370 | if (hasProjectionBias) |
| 371 | { |
| 372 | buffers.push_back( |
| 373 | CreateBuffer(flatBufferBuilder, |
| 374 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(projectionBias.data()), |
| 375 | sizeof(T) * projectionBias.size()))); |
| 376 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 377 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 378 | tensorInfo4.size()), |
| 379 | tensorType, |
| 380 | buffers.size() - 1, |
| 381 | flatBufferBuilder.CreateString("projectionBias"), |
| 382 | outputQuantizationParameters)); |
| 383 | operatorInputs.push_back(buffers.size() - 1); |
| 384 | } |
| 385 | else |
| 386 | { |
| 387 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 388 | } |
| 389 | |
Ryan OShea | 238ecd9 | 2023-03-07 11:44:23 +0000 | [diff] [blame] | 390 | buffers.push_back(CreateBuffer(flatBufferBuilder)); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 391 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 392 | flatBufferBuilder.CreateVector<int32_t>(outputStateInDimensions.data(), |
| 393 | outputStateInDimensions.size()), |
| 394 | tensorType, |
| 395 | buffers.size() - 1, |
| 396 | flatBufferBuilder.CreateString("outputStateInInfo"), |
| 397 | outputQuantizationParameters, |
| 398 | true)); |
| 399 | operatorInputs.push_back(buffers.size() - 1); |
| 400 | |
Ryan OShea | 238ecd9 | 2023-03-07 11:44:23 +0000 | [diff] [blame] | 401 | buffers.push_back(CreateBuffer(flatBufferBuilder)); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 402 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 403 | flatBufferBuilder.CreateVector<int32_t>(cellStateInDimensions.data(), |
| 404 | cellStateInDimensions.size()), |
| 405 | tensorType, |
| 406 | buffers.size() - 1, |
| 407 | flatBufferBuilder.CreateString("cellStateInInfo"), |
| 408 | outputQuantizationParameters, |
| 409 | true)); |
| 410 | operatorInputs.push_back(buffers.size() - 1); |
| 411 | |
| 412 | if (hasInputLayerNormWeights) |
| 413 | { |
| 414 | buffers.push_back( |
| 415 | CreateBuffer(flatBufferBuilder, |
| 416 | flatBufferBuilder.CreateVector( |
| 417 | reinterpret_cast<const uint8_t *>(inputLayerNormWeights.data()), |
| 418 | sizeof(T) * inputLayerNormWeights.size()))); |
| 419 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 420 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 421 | tensorInfo4.size()), |
| 422 | tensorType, |
| 423 | buffers.size() - 1, |
| 424 | flatBufferBuilder.CreateString("inputLayerNormWeights"), |
| 425 | outputQuantizationParameters)); |
| 426 | operatorInputs.push_back(buffers.size() - 1); |
| 427 | } |
| 428 | else |
| 429 | { |
| 430 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 431 | } |
| 432 | |
| 433 | if (hasForgetLayerNormWeights) |
| 434 | { |
| 435 | buffers.push_back( |
| 436 | CreateBuffer(flatBufferBuilder, |
| 437 | flatBufferBuilder.CreateVector( |
| 438 | reinterpret_cast<const uint8_t *>(forgetLayerNormWeights.data()), |
| 439 | sizeof(T) * forgetLayerNormWeights.size()))); |
| 440 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 441 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 442 | tensorInfo4.size()), |
| 443 | tensorType, |
| 444 | buffers.size() - 1, |
| 445 | flatBufferBuilder.CreateString("forgetLayerNormWeights"), |
| 446 | outputQuantizationParameters)); |
| 447 | operatorInputs.push_back(buffers.size() - 1); |
| 448 | } |
| 449 | else |
| 450 | { |
| 451 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 452 | } |
| 453 | |
| 454 | if (hasCellLayerNormWeights) |
| 455 | { |
| 456 | buffers.push_back( |
| 457 | CreateBuffer(flatBufferBuilder, |
| 458 | flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t *>(cellLayerNormWeights.data()), |
| 459 | sizeof(T) * cellLayerNormWeights.size()))); |
| 460 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 461 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 462 | tensorInfo4.size()), |
| 463 | tensorType, |
| 464 | buffers.size() - 1, |
| 465 | flatBufferBuilder.CreateString("cellLayerNormWeights"), |
| 466 | outputQuantizationParameters)); |
| 467 | operatorInputs.push_back(buffers.size() - 1); |
| 468 | } |
| 469 | else |
| 470 | { |
| 471 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 472 | } |
| 473 | |
| 474 | if (hasOutputLayerNormWeights) |
| 475 | { |
| 476 | buffers.push_back( |
| 477 | CreateBuffer(flatBufferBuilder, |
| 478 | flatBufferBuilder.CreateVector( |
| 479 | reinterpret_cast<const uint8_t *>(outputLayerNormWeights.data()), |
| 480 | sizeof(T) * outputLayerNormWeights.size()))); |
| 481 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 482 | flatBufferBuilder.CreateVector<int32_t>(tensorInfo4.data(), |
| 483 | tensorInfo4.size()), |
| 484 | tensorType, |
| 485 | buffers.size() - 1, |
| 486 | flatBufferBuilder.CreateString("outputLayerNormWeights"), |
| 487 | outputQuantizationParameters)); |
| 488 | operatorInputs.push_back(buffers.size() - 1); |
| 489 | } |
| 490 | else |
| 491 | { |
| 492 | operatorInputs.push_back(kTfLiteOptionalTensor); |
| 493 | } |
| 494 | int outputBufferId = buffers.size(); |
Ryan OShea | 238ecd9 | 2023-03-07 11:44:23 +0000 | [diff] [blame] | 495 | buffers.push_back(CreateBuffer(flatBufferBuilder)); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 496 | tensors.push_back(CreateTensor(flatBufferBuilder, |
| 497 | flatBufferBuilder.CreateVector<int32_t>(outputShape.data(), |
| 498 | outputShape.size()), |
| 499 | tensorType, |
| 500 | outputBufferId, |
| 501 | flatBufferBuilder.CreateString("output"), |
| 502 | outputQuantizationParameters)); |
| 503 | std::vector<int> operatorOutputs; |
| 504 | operatorOutputs.push_back(buffers.size() - 1); |
| 505 | |
| 506 | // create operator |
| 507 | tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_LSTMOptions; |
| 508 | flatbuffers::Offset<void> operatorBuiltinOptions = |
| 509 | CreateLSTMOptions(flatBufferBuilder, |
| 510 | activationFunction, |
| 511 | clippingThresCell, |
| 512 | clippingThresProj).Union(); |
| 513 | |
| 514 | flatbuffers::Offset <Operator> lstmOperator = |
| 515 | CreateOperator(flatBufferBuilder, |
| 516 | 0, |
| 517 | flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()), |
| 518 | flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()), |
| 519 | operatorBuiltinOptionsType, operatorBuiltinOptions); |
| 520 | |
| 521 | flatbuffers::Offset <SubGraph> subgraph = |
| 522 | CreateSubGraph(flatBufferBuilder, |
| 523 | flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), |
| 524 | flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()), |
| 525 | flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()), |
| 526 | flatBufferBuilder.CreateVector(&lstmOperator, 1)); |
| 527 | |
| 528 | flatbuffers::Offset <flatbuffers::String> modelDescription = |
| 529 | flatBufferBuilder.CreateString("ArmnnDelegate: LSTM Operator Model"); |
| 530 | flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, |
| 531 | tflite::BuiltinOperator_LSTM); |
| 532 | |
| 533 | flatbuffers::Offset <Model> flatbufferModel = |
| 534 | CreateModel(flatBufferBuilder, |
| 535 | TFLITE_SCHEMA_VERSION, |
| 536 | flatBufferBuilder.CreateVector(&operatorCode, 1), |
| 537 | flatBufferBuilder.CreateVector(&subgraph, 1), |
| 538 | modelDescription, |
| 539 | flatBufferBuilder.CreateVector(buffers.data(), buffers.size())); |
| 540 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 541 | flatBufferBuilder.Finish(flatbufferModel, armnnDelegate::FILE_IDENTIFIER); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 542 | |
| 543 | return std::vector<char>(flatBufferBuilder.GetBufferPointer(), |
| 544 | flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); |
| 545 | } |
| 546 | |
| 547 | template <typename T> |
| 548 | void LstmTestImpl(std::vector<armnn::BackendId>& backends, |
| 549 | tflite::TensorType tensorType, |
| 550 | int32_t batchSize, |
| 551 | int32_t inputSize, |
| 552 | int32_t outputSize, |
| 553 | int32_t numUnits, |
| 554 | bool hasInputToInputWeights, |
| 555 | const std::vector<T>& inputToInputWeights, |
| 556 | const std::vector<T>& inputToForgetWeights, |
| 557 | const std::vector<T>& inputToCellWeights, |
| 558 | const std::vector<T>& inputToOutputWeights, |
| 559 | bool hasRecurrentToInputWeights, |
| 560 | const std::vector<T>& recurrentToInputWeights, |
| 561 | const std::vector<T>& recurrentToForgetWeights, |
| 562 | const std::vector<T>& recurrentToCellWeights, |
| 563 | const std::vector<T>& recurrentToOutputWeights, |
| 564 | bool hasCellToInputWeights, |
| 565 | const std::vector<T>& cellToInputWeights, |
| 566 | bool hasCellToForgetWeights, |
| 567 | const std::vector<T>& cellToForgetWeights, |
| 568 | bool hasCellToOutputWeights, |
| 569 | const std::vector<T>& cellToOutputWeights, |
| 570 | bool hasInputGateBias, |
| 571 | const std::vector<T>& inputGateBias, |
| 572 | const std::vector<T>& forgetGateBias, |
| 573 | const std::vector<T>& cellBias, |
| 574 | const std::vector<T>& outputGateBias, |
| 575 | bool hasProjectionWeights, |
| 576 | const std::vector<T>& projectionWeights, |
| 577 | bool hasProjectionBias, |
| 578 | const std::vector<T>& projectionBias, |
| 579 | bool hasInputLayerNormWeights, |
| 580 | const std::vector<T>& inputLayerNormWeights, |
| 581 | bool hasForgetLayerNormWeights, |
| 582 | const std::vector<T>& forgetLayerNormWeights, |
| 583 | bool hasCellLayerNormWeights, |
| 584 | const std::vector<T>& cellLayerNormWeights, |
| 585 | bool hasOutputLayerNormWeights, |
| 586 | const std::vector<T>& outputLayerNormWeights, |
| 587 | std::vector<T>& inputValues, |
| 588 | std::vector<T>& expectedOutputValues, |
| 589 | tflite::ActivationFunctionType activationFunction, |
| 590 | float clippingThresCell, |
| 591 | float clippingThresProj) |
| 592 | { |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 593 | using namespace delegateTestInterpreter; |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 594 | |
| 595 | std::vector<char> modelBuffer = CreateLstmTfLiteModel(tensorType, |
| 596 | batchSize, |
| 597 | inputSize, |
| 598 | outputSize, |
| 599 | numUnits, |
| 600 | hasInputToInputWeights, |
| 601 | inputToInputWeights, |
| 602 | inputToForgetWeights, |
| 603 | inputToCellWeights, |
| 604 | inputToOutputWeights, |
| 605 | hasRecurrentToInputWeights, |
| 606 | recurrentToInputWeights, |
| 607 | recurrentToForgetWeights, |
| 608 | recurrentToCellWeights, |
| 609 | recurrentToOutputWeights, |
| 610 | hasCellToInputWeights, |
| 611 | cellToInputWeights, |
| 612 | hasCellToForgetWeights, |
| 613 | cellToForgetWeights, |
| 614 | hasCellToOutputWeights, |
| 615 | cellToOutputWeights, |
| 616 | hasInputGateBias, |
| 617 | inputGateBias, |
| 618 | forgetGateBias, |
| 619 | cellBias, |
| 620 | outputGateBias, |
| 621 | hasProjectionWeights, |
| 622 | projectionWeights, |
| 623 | hasProjectionBias, |
| 624 | projectionBias, |
| 625 | hasInputLayerNormWeights, |
| 626 | inputLayerNormWeights, |
| 627 | hasForgetLayerNormWeights, |
| 628 | forgetLayerNormWeights, |
| 629 | hasCellLayerNormWeights, |
| 630 | cellLayerNormWeights, |
| 631 | hasOutputLayerNormWeights, |
| 632 | outputLayerNormWeights, |
| 633 | activationFunction, |
| 634 | clippingThresCell, |
| 635 | clippingThresProj); |
| 636 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 637 | std::vector<int32_t> expectedOutputShape {batchSize , outputSize}; |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 638 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 639 | // Setup interpreter with just TFLite Runtime. |
| 640 | auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer); |
| 641 | CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk); |
| 642 | CHECK(tfLiteInterpreter.FillInputTensor<T>(inputValues, 0) == kTfLiteOk); |
| 643 | CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk); |
| 644 | std::vector<T> tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<T>(0); |
| 645 | std::vector<int32_t> tfLiteOutputShape = tfLiteInterpreter.GetOutputShape(0); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 646 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 647 | // Setup interpreter with Arm NN Delegate applied. |
| 648 | auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, backends); |
| 649 | CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk); |
| 650 | CHECK(armnnInterpreter.FillInputTensor<T>(inputValues, 0) == kTfLiteOk); |
| 651 | CHECK(armnnInterpreter.Invoke() == kTfLiteOk); |
| 652 | std::vector<T> armnnOutputValues = armnnInterpreter.GetOutputResult<T>(0); |
| 653 | std::vector<int32_t> armnnOutputShape = armnnInterpreter.GetOutputShape(0); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 654 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 655 | armnnDelegate::CompareOutputData<T>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues); |
| 656 | armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, expectedOutputShape); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 657 | |
Matthew Sloyan | ebe392d | 2023-03-30 10:12:08 +0100 | [diff] [blame] | 658 | tfLiteInterpreter.Cleanup(); |
| 659 | armnnInterpreter.Cleanup(); |
Mike Kelly | 8ae17b3 | 2021-02-17 13:45:50 +0000 | [diff] [blame] | 660 | } |
| 661 | |
| 662 | } // anonymous namespace |