Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "BatchMatMulTestImpl.hpp" |
| 7 | |
| 8 | #include <armnn/backends/IBackendInternal.hpp> |
| 9 | #include <armnn/backends/Workload.hpp> |
| 10 | #include <armnn/backends/WorkloadData.hpp> |
| 11 | #include <armnn/backends/WorkloadFactory.hpp> |
| 12 | |
| 13 | #include <armnnTestUtils/WorkloadTestUtils.hpp> |
| 14 | #include <armnnUtils/QuantizeHelper.hpp> |
| 15 | #include <armnnTestUtils/TensorCopyUtils.hpp> |
| 16 | #include <armnn/Optional.hpp> |
| 17 | |
| 18 | |
| 19 | template<armnn::DataType ArmnnType, typename T, std::size_t NumDims> |
| 20 | LayerTestResult<T, NumDims> BatchMatMulTestImpl( |
| 21 | armnn::IWorkloadFactory& workloadFactory, |
| 22 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 23 | const armnn::ITensorHandleFactory& tensorHandleFactory, |
| 24 | armnn::BatchMatMulDescriptor descriptor, |
| 25 | const std::vector<T>& inputX, |
| 26 | const std::vector<T>& inputY, |
| 27 | const std::vector<T>& outputExpected, |
| 28 | const armnn::TensorInfo& inputXInfo, |
| 29 | const armnn::TensorInfo& inputYInfo, |
| 30 | const armnn::TensorInfo& outputInfo) |
| 31 | { |
| 32 | std::vector<T> outputActual(outputInfo.GetNumElements()); |
| 33 | |
| 34 | std::unique_ptr<armnn::ITensorHandle> inputXHandle = tensorHandleFactory.CreateTensorHandle(inputXInfo); |
| 35 | std::unique_ptr<armnn::ITensorHandle> inputYHandle = tensorHandleFactory.CreateTensorHandle(inputYInfo); |
| 36 | std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputInfo); |
| 37 | |
| 38 | armnn::BatchMatMulQueueDescriptor queueDescriptor; |
| 39 | queueDescriptor.m_Parameters = descriptor; |
| 40 | armnn::WorkloadInfo workloadInfo; |
| 41 | |
| 42 | AddInputToWorkload(queueDescriptor, workloadInfo, inputXInfo, inputXHandle.get()); |
| 43 | AddInputToWorkload(queueDescriptor, workloadInfo, inputYInfo, inputYHandle.get()); |
| 44 | AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get()); |
| 45 | |
| 46 | auto workload = workloadFactory.CreateWorkload(armnn::LayerType::BatchMatMul, queueDescriptor, workloadInfo); |
| 47 | |
| 48 | inputXHandle->Allocate(); |
| 49 | inputYHandle->Allocate(); |
| 50 | outputHandle->Allocate(); |
| 51 | |
| 52 | CopyDataToITensorHandle(inputXHandle.get(), inputX.data()); |
| 53 | CopyDataToITensorHandle(inputYHandle.get(), inputY.data()); |
| 54 | |
| 55 | workload->PostAllocationConfigure(); |
| 56 | ExecuteWorkload(*workload, memoryManager); |
| 57 | |
| 58 | CopyDataFromITensorHandle(outputActual.data(), outputHandle.get()); |
| 59 | |
| 60 | return LayerTestResult<T, NumDims>(outputActual, |
| 61 | outputExpected, |
| 62 | outputHandle->GetShape(), |
| 63 | outputInfo.GetShape()); |
| 64 | } |
| 65 | |
| 66 | template<armnn::DataType ArmnnType, typename T> |
| 67 | LayerTestResult<T, 2> BatchMatMul2DSimpleTest( |
| 68 | armnn::IWorkloadFactory& workloadFactory, |
| 69 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 70 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 71 | { |
| 72 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 73 | |
| 74 | float qScale = 0.0f; |
| 75 | int32_t qOffset = 0; |
| 76 | |
| 77 | switch(ArmnnType) |
| 78 | { |
| 79 | case armnn::DataType::QAsymmS8: |
| 80 | case armnn::DataType::QAsymmU8: |
| 81 | case armnn::DataType::QSymmS16: |
| 82 | qScale = 1.0f; |
| 83 | break; |
| 84 | default: |
| 85 | break; |
| 86 | } |
| 87 | |
| 88 | armnn::TensorInfo inputXInfo({2,2}, ArmnnType, qScale, qOffset); |
| 89 | armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset); |
| 90 | armnn::TensorInfo outputInfo({2,2}, ArmnnType, qScale, qOffset); |
| 91 | |
| 92 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 93 | 1, 2, |
| 94 | 3, 4 |
| 95 | }, qScale, qOffset); |
| 96 | |
| 97 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 98 | 5, 6, |
| 99 | 7, 8 |
| 100 | }, qScale, qOffset); |
| 101 | |
| 102 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 103 | 19, 22, |
| 104 | 43, 50 |
| 105 | }, qScale, qOffset); |
| 106 | |
| 107 | return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory, |
| 108 | memoryManager, |
| 109 | tensorHandleFactory, |
| 110 | descriptor, |
| 111 | inputX, |
| 112 | inputY, |
| 113 | outputExpected, |
| 114 | inputXInfo, |
| 115 | inputYInfo, |
| 116 | outputInfo); |
| 117 | } |
| 118 | |
| 119 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2> |
| 120 | BatchMatMul2DSimpleTest<armnn::DataType::BFloat16>( |
| 121 | armnn::IWorkloadFactory& workloadFactory, |
| 122 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 123 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 124 | |
| 125 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2> |
| 126 | BatchMatMul2DSimpleTest<armnn::DataType::Float32>( |
| 127 | armnn::IWorkloadFactory& workloadFactory, |
| 128 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 129 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 130 | |
| 131 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2> |
| 132 | BatchMatMul2DSimpleTest<armnn::DataType::Float16>( |
| 133 | armnn::IWorkloadFactory& workloadFactory, |
| 134 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 135 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 136 | |
| 137 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2> |
| 138 | BatchMatMul2DSimpleTest<armnn::DataType::QAsymmS8>( |
| 139 | armnn::IWorkloadFactory& workloadFactory, |
| 140 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 141 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 142 | |
| 143 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2> |
| 144 | BatchMatMul2DSimpleTest<armnn::DataType::QAsymmU8>( |
| 145 | armnn::IWorkloadFactory& workloadFactory, |
| 146 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 147 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 148 | |
| 149 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2> |
| 150 | BatchMatMul2DSimpleTest<armnn::DataType::QSymmS16>( |
| 151 | armnn::IWorkloadFactory& workloadFactory, |
| 152 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 153 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 154 | |
| 155 | template<armnn::DataType ArmnnType, typename T> |
| 156 | LayerTestResult<T, 3> BatchMatMul3DSimpleTest( |
| 157 | armnn::IWorkloadFactory& workloadFactory, |
| 158 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 159 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 160 | { |
| 161 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 162 | |
| 163 | float qScale = 0.0f; |
| 164 | int32_t qOffset = 0; |
| 165 | |
| 166 | switch(ArmnnType) |
| 167 | { |
| 168 | case armnn::DataType::QAsymmS8: |
| 169 | case armnn::DataType::QAsymmU8: |
| 170 | case armnn::DataType::QSymmS16: |
| 171 | qScale = 1.0f; |
| 172 | break; |
| 173 | default: |
| 174 | break; |
| 175 | } |
| 176 | |
| 177 | armnn::TensorInfo inputXInfo({1,2,2}, ArmnnType, qScale, qOffset); |
| 178 | armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset); |
| 179 | armnn::TensorInfo outputInfo({1,2,2}, ArmnnType, qScale, qOffset); |
| 180 | |
| 181 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 182 | 1, 2, |
| 183 | 3, 4 |
| 184 | }, qScale, qOffset); |
| 185 | |
| 186 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 187 | 5, 6, |
| 188 | 7, 8 |
| 189 | }, qScale, qOffset); |
| 190 | |
| 191 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 192 | 19, 22, |
| 193 | 43, 50 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 194 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 195 | |
| 196 | return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory, |
| 197 | memoryManager, |
| 198 | tensorHandleFactory, |
| 199 | descriptor, |
| 200 | inputX, |
| 201 | inputY, |
| 202 | outputExpected, |
| 203 | inputXInfo, |
| 204 | inputYInfo, |
| 205 | outputInfo); |
| 206 | } |
| 207 | |
| 208 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3> |
| 209 | BatchMatMul3DSimpleTest<armnn::DataType::BFloat16>( |
| 210 | armnn::IWorkloadFactory& workloadFactory, |
| 211 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 212 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 213 | |
| 214 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3> |
| 215 | BatchMatMul3DSimpleTest<armnn::DataType::Float32>( |
| 216 | armnn::IWorkloadFactory& workloadFactory, |
| 217 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 218 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 219 | |
| 220 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3> |
| 221 | BatchMatMul3DSimpleTest<armnn::DataType::Float16>( |
| 222 | armnn::IWorkloadFactory& workloadFactory, |
| 223 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 224 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 225 | |
| 226 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3> |
| 227 | BatchMatMul3DSimpleTest<armnn::DataType::QAsymmS8>( |
| 228 | armnn::IWorkloadFactory& workloadFactory, |
| 229 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 230 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 231 | |
| 232 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3> |
| 233 | BatchMatMul3DSimpleTest<armnn::DataType::QAsymmU8>( |
| 234 | armnn::IWorkloadFactory& workloadFactory, |
| 235 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 236 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 237 | |
| 238 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3> |
| 239 | BatchMatMul3DSimpleTest<armnn::DataType::QSymmS16>( |
| 240 | armnn::IWorkloadFactory& workloadFactory, |
| 241 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 242 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 243 | |
| 244 | template<armnn::DataType ArmnnType, typename T> |
| 245 | LayerTestResult<T, 4> BatchMatMulNCHWSimpleTest( |
| 246 | armnn::IWorkloadFactory& workloadFactory, |
| 247 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 248 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 249 | { |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 250 | auto descriptor = armnn::BatchMatMulDescriptor(); // Default arbitrary layout is treated the same as NCHW |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 251 | |
| 252 | float qScale = 0.0f; |
| 253 | int32_t qOffset = 0; |
| 254 | |
| 255 | switch(ArmnnType) |
| 256 | { |
| 257 | case armnn::DataType::QAsymmS8: |
| 258 | case armnn::DataType::QAsymmU8: |
| 259 | case armnn::DataType::QSymmS16: |
| 260 | qScale = 1.0f; |
| 261 | break; |
| 262 | default: |
| 263 | break; |
| 264 | } |
| 265 | |
| 266 | armnn::TensorInfo inputXInfo({1,1,2,2}, ArmnnType, qScale, qOffset); |
| 267 | armnn::TensorInfo inputYInfo({1,1,2,2}, ArmnnType, qScale, qOffset); |
| 268 | armnn::TensorInfo outputInfo({1,1,2,2}, ArmnnType, qScale, qOffset); |
| 269 | |
| 270 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 271 | 1, 2, |
| 272 | 3, 4 |
| 273 | }, qScale, qOffset); |
| 274 | |
| 275 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 276 | 5, 6, |
| 277 | 7, 8 |
| 278 | }, qScale, qOffset); |
| 279 | |
| 280 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 281 | 19, 22, |
| 282 | 43, 50 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 283 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 284 | |
| 285 | return BatchMatMulTestImpl<ArmnnType, T, 4>(workloadFactory, |
| 286 | memoryManager, |
| 287 | tensorHandleFactory, |
| 288 | descriptor, |
| 289 | inputX, |
| 290 | inputY, |
| 291 | outputExpected, |
| 292 | inputXInfo, |
| 293 | inputYInfo, |
| 294 | outputInfo); |
| 295 | } |
| 296 | |
| 297 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 4> |
| 298 | BatchMatMulNCHWSimpleTest<armnn::DataType::BFloat16>( |
| 299 | armnn::IWorkloadFactory& workloadFactory, |
| 300 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 301 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 302 | |
| 303 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4> |
| 304 | BatchMatMulNCHWSimpleTest<armnn::DataType::Float32>( |
| 305 | armnn::IWorkloadFactory& workloadFactory, |
| 306 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 307 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 308 | |
| 309 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 4> |
| 310 | BatchMatMulNCHWSimpleTest<armnn::DataType::Float16>( |
| 311 | armnn::IWorkloadFactory& workloadFactory, |
| 312 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 313 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 314 | |
| 315 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 4> |
| 316 | BatchMatMulNCHWSimpleTest<armnn::DataType::QAsymmS8>( |
| 317 | armnn::IWorkloadFactory& workloadFactory, |
| 318 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 319 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 320 | |
| 321 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4> |
| 322 | BatchMatMulNCHWSimpleTest<armnn::DataType::QAsymmU8>( |
| 323 | armnn::IWorkloadFactory& workloadFactory, |
| 324 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 325 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 326 | |
| 327 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 4> |
| 328 | BatchMatMulNCHWSimpleTest<armnn::DataType::QSymmS16>( |
| 329 | armnn::IWorkloadFactory& workloadFactory, |
| 330 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 331 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 332 | |
| 333 | template<armnn::DataType ArmnnType, typename T> |
| 334 | LayerTestResult<T, 4> BatchMatMulNHWCSimpleTest( |
| 335 | armnn::IWorkloadFactory& workloadFactory, |
| 336 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 337 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 338 | { |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 339 | auto descriptor = armnn::BatchMatMulDescriptor(false, |
| 340 | false, |
| 341 | false, |
| 342 | false, |
| 343 | armnn::DataLayout::NHWC, |
| 344 | armnn::DataLayout::NHWC); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 345 | |
| 346 | float qScale = 0.0f; |
| 347 | int32_t qOffset = 0; |
| 348 | |
| 349 | switch(ArmnnType) |
| 350 | { |
| 351 | case armnn::DataType::QAsymmS8: |
| 352 | case armnn::DataType::QAsymmU8: |
| 353 | case armnn::DataType::QSymmS16: |
| 354 | qScale = 1.0f; |
| 355 | break; |
| 356 | default: |
| 357 | break; |
| 358 | } |
| 359 | |
| 360 | armnn::TensorInfo inputXInfo({1,2,2,1}, ArmnnType, qScale, qOffset); |
| 361 | armnn::TensorInfo inputYInfo({1,2,2,1}, ArmnnType, qScale, qOffset); |
| 362 | armnn::TensorInfo outputInfo({1,2,2,1}, ArmnnType, qScale, qOffset); |
| 363 | |
| 364 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 365 | 1, 2, |
| 366 | 3, 4 |
| 367 | }, qScale, qOffset); |
| 368 | |
| 369 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 370 | 5, 6, |
| 371 | 7, 8 |
| 372 | }, qScale, qOffset); |
| 373 | |
| 374 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 375 | 19, 22, |
| 376 | 43, 50 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 377 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 378 | |
| 379 | return BatchMatMulTestImpl<ArmnnType, T, 4>(workloadFactory, |
| 380 | memoryManager, |
| 381 | tensorHandleFactory, |
| 382 | descriptor, |
| 383 | inputX, |
| 384 | inputY, |
| 385 | outputExpected, |
| 386 | inputXInfo, |
| 387 | inputYInfo, |
| 388 | outputInfo); |
| 389 | } |
| 390 | |
| 391 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 4> |
| 392 | BatchMatMulNHWCSimpleTest<armnn::DataType::BFloat16>( |
| 393 | armnn::IWorkloadFactory& workloadFactory, |
| 394 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 395 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 396 | |
| 397 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4> |
| 398 | BatchMatMulNHWCSimpleTest<armnn::DataType::Float32>( |
| 399 | armnn::IWorkloadFactory& workloadFactory, |
| 400 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 401 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 402 | |
| 403 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 4> |
| 404 | BatchMatMulNHWCSimpleTest<armnn::DataType::Float16>( |
| 405 | armnn::IWorkloadFactory& workloadFactory, |
| 406 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 407 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 408 | |
| 409 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 4> |
| 410 | BatchMatMulNHWCSimpleTest<armnn::DataType::QAsymmS8>( |
| 411 | armnn::IWorkloadFactory& workloadFactory, |
| 412 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 413 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 414 | |
| 415 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4> |
| 416 | BatchMatMulNHWCSimpleTest<armnn::DataType::QAsymmU8>( |
| 417 | armnn::IWorkloadFactory& workloadFactory, |
| 418 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 419 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 420 | |
| 421 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 4> |
| 422 | BatchMatMulNHWCSimpleTest<armnn::DataType::QSymmS16>( |
| 423 | armnn::IWorkloadFactory& workloadFactory, |
| 424 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 425 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 426 | |
| 427 | template<armnn::DataType ArmnnType, typename T> |
| 428 | LayerTestResult<T, 3> BatchMatMul3DBatchTest( |
| 429 | armnn::IWorkloadFactory& workloadFactory, |
| 430 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 431 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 432 | { |
| 433 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 434 | |
| 435 | float qScale = 0.0f; |
| 436 | int32_t qOffset = 0; |
| 437 | |
| 438 | switch(ArmnnType) |
| 439 | { |
| 440 | case armnn::DataType::QAsymmS8: |
| 441 | case armnn::DataType::QAsymmU8: |
| 442 | case armnn::DataType::QSymmS16: |
| 443 | qScale = 1.0f; |
| 444 | break; |
| 445 | default: |
| 446 | break; |
| 447 | } |
| 448 | |
| 449 | armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 450 | armnn::TensorInfo inputYInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 451 | armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 452 | |
| 453 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 454 | 1, 2, |
| 455 | 3, 4, |
| 456 | |
| 457 | 9, 10, |
| 458 | 11, 12 |
| 459 | }, qScale, qOffset); |
| 460 | |
| 461 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 462 | 5, 6, |
| 463 | 7, 8, |
| 464 | |
| 465 | 13, 14, |
| 466 | 15, 16 |
| 467 | }, qScale, qOffset); |
| 468 | |
| 469 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 470 | 19, 22, |
| 471 | 43, 50, |
| 472 | |
| 473 | 267, 286, |
| 474 | 323, 346 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 475 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 476 | |
| 477 | return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory, |
| 478 | memoryManager, |
| 479 | tensorHandleFactory, |
| 480 | descriptor, |
| 481 | inputX, |
| 482 | inputY, |
| 483 | outputExpected, |
| 484 | inputXInfo, |
| 485 | inputYInfo, |
| 486 | outputInfo); |
| 487 | } |
| 488 | |
| 489 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3> |
| 490 | BatchMatMul3DBatchTest<armnn::DataType::BFloat16>( |
| 491 | armnn::IWorkloadFactory& workloadFactory, |
| 492 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 493 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 494 | |
| 495 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3> |
| 496 | BatchMatMul3DBatchTest<armnn::DataType::Float32>( |
| 497 | armnn::IWorkloadFactory& workloadFactory, |
| 498 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 499 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 500 | |
| 501 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3> |
| 502 | BatchMatMul3DBatchTest<armnn::DataType::Float16>( |
| 503 | armnn::IWorkloadFactory& workloadFactory, |
| 504 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 505 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 506 | |
| 507 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3> |
| 508 | BatchMatMul3DBatchTest<armnn::DataType::QAsymmS8>( |
| 509 | armnn::IWorkloadFactory& workloadFactory, |
| 510 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 511 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 512 | |
| 513 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3> |
| 514 | BatchMatMul3DBatchTest<armnn::DataType::QAsymmU8>( |
| 515 | armnn::IWorkloadFactory& workloadFactory, |
| 516 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 517 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 518 | |
| 519 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3> |
| 520 | BatchMatMul3DBatchTest<armnn::DataType::QSymmS16>( |
| 521 | armnn::IWorkloadFactory& workloadFactory, |
| 522 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 523 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 524 | |
| 525 | template<armnn::DataType ArmnnType, typename T> |
| 526 | LayerTestResult<T, 3> BatchMatMul3DBroadcastTest( |
| 527 | armnn::IWorkloadFactory& workloadFactory, |
| 528 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 529 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 530 | { |
| 531 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 532 | |
| 533 | float qScale = 0.0f; |
| 534 | int32_t qOffset = 0; |
| 535 | |
| 536 | switch(ArmnnType) |
| 537 | { |
| 538 | case armnn::DataType::QAsymmS8: |
| 539 | case armnn::DataType::QAsymmU8: |
| 540 | case armnn::DataType::QSymmS16: |
| 541 | qScale = 1.0f; |
| 542 | break; |
| 543 | default: |
| 544 | break; |
| 545 | } |
| 546 | |
| 547 | armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 548 | armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset); |
| 549 | armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 550 | |
| 551 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 552 | 1, 2, |
| 553 | 3, 4, |
| 554 | |
| 555 | 9, 10, |
| 556 | 11, 12 |
| 557 | }, qScale, qOffset); |
| 558 | |
| 559 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 560 | 13, 14, |
| 561 | 15, 16 |
| 562 | }, qScale, qOffset); |
| 563 | |
| 564 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 565 | 43, 46, |
| 566 | 99, 106, |
| 567 | |
| 568 | 267, 286, |
| 569 | 323, 346 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 570 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 571 | |
| 572 | return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory, |
| 573 | memoryManager, |
| 574 | tensorHandleFactory, |
| 575 | descriptor, |
| 576 | inputX, |
| 577 | inputY, |
| 578 | outputExpected, |
| 579 | inputXInfo, |
| 580 | inputYInfo, |
| 581 | outputInfo); |
| 582 | } |
| 583 | |
| 584 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3> |
| 585 | BatchMatMul3DBroadcastTest<armnn::DataType::BFloat16>( |
| 586 | armnn::IWorkloadFactory& workloadFactory, |
| 587 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 588 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 589 | |
| 590 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3> |
| 591 | BatchMatMul3DBroadcastTest<armnn::DataType::Float32>( |
| 592 | armnn::IWorkloadFactory& workloadFactory, |
| 593 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 594 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 595 | |
| 596 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3> |
| 597 | BatchMatMul3DBroadcastTest<armnn::DataType::Float16>( |
| 598 | armnn::IWorkloadFactory& workloadFactory, |
| 599 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 600 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 601 | |
| 602 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3> |
| 603 | BatchMatMul3DBroadcastTest<armnn::DataType::QAsymmS8>( |
| 604 | armnn::IWorkloadFactory& workloadFactory, |
| 605 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 606 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 607 | |
| 608 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3> |
| 609 | BatchMatMul3DBroadcastTest<armnn::DataType::QAsymmU8>( |
| 610 | armnn::IWorkloadFactory& workloadFactory, |
| 611 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 612 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 613 | |
| 614 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3> |
| 615 | BatchMatMul3DBroadcastTest<armnn::DataType::QSymmS16>( |
| 616 | armnn::IWorkloadFactory& workloadFactory, |
| 617 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 618 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 619 | |
| 620 | template<armnn::DataType ArmnnType, typename T> |
| 621 | LayerTestResult<T, 3> BatchMatMul3D2DBroadcastTest( |
| 622 | armnn::IWorkloadFactory& workloadFactory, |
| 623 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 624 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 625 | { |
| 626 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 627 | |
| 628 | float qScale = 0.0f; |
| 629 | int32_t qOffset = 0; |
| 630 | |
| 631 | switch(ArmnnType) |
| 632 | { |
| 633 | case armnn::DataType::QAsymmS8: |
| 634 | case armnn::DataType::QAsymmU8: |
| 635 | case armnn::DataType::QSymmS16: |
| 636 | qScale = 1.0f; |
| 637 | break; |
| 638 | default: |
| 639 | break; |
| 640 | } |
| 641 | |
| 642 | armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 643 | armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset); |
| 644 | armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset); |
| 645 | |
| 646 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 647 | 1, 2, |
| 648 | 3, 4, |
| 649 | |
| 650 | 9, 10, |
| 651 | 11, 12 |
| 652 | }, qScale, qOffset); |
| 653 | |
| 654 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 655 | 13, 14, |
| 656 | 15, 16 |
| 657 | }, qScale, qOffset); |
| 658 | |
| 659 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 660 | 43, 46, |
| 661 | 99, 106, |
| 662 | |
| 663 | 267, 286, |
| 664 | 323, 346 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 665 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 666 | |
| 667 | return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory, |
| 668 | memoryManager, |
| 669 | tensorHandleFactory, |
| 670 | descriptor, |
| 671 | inputX, |
| 672 | inputY, |
| 673 | outputExpected, |
| 674 | inputXInfo, |
| 675 | inputYInfo, |
| 676 | outputInfo); |
| 677 | } |
| 678 | |
| 679 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3> |
| 680 | BatchMatMul3D2DBroadcastTest<armnn::DataType::BFloat16>( |
| 681 | armnn::IWorkloadFactory& workloadFactory, |
| 682 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 683 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 684 | |
| 685 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3> |
| 686 | BatchMatMul3D2DBroadcastTest<armnn::DataType::Float32>( |
| 687 | armnn::IWorkloadFactory& workloadFactory, |
| 688 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 689 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 690 | |
| 691 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3> |
| 692 | BatchMatMul3D2DBroadcastTest<armnn::DataType::Float16>( |
| 693 | armnn::IWorkloadFactory& workloadFactory, |
| 694 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 695 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 696 | |
| 697 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3> |
| 698 | BatchMatMul3D2DBroadcastTest<armnn::DataType::QAsymmS8>( |
| 699 | armnn::IWorkloadFactory& workloadFactory, |
| 700 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 701 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 702 | |
| 703 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3> |
| 704 | BatchMatMul3D2DBroadcastTest<armnn::DataType::QAsymmU8>( |
| 705 | armnn::IWorkloadFactory& workloadFactory, |
| 706 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 707 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 708 | |
| 709 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3> |
| 710 | BatchMatMul3D2DBroadcastTest<armnn::DataType::QSymmS16>( |
| 711 | armnn::IWorkloadFactory& workloadFactory, |
| 712 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 713 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 714 | |
| 715 | template<armnn::DataType ArmnnType, typename T> |
| 716 | LayerTestResult<T, 5> BatchMatMulNDHWCNHWCTest( |
| 717 | armnn::IWorkloadFactory& workloadFactory, |
| 718 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 719 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 720 | { |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 721 | auto descriptor = armnn::BatchMatMulDescriptor(false, |
| 722 | false, |
| 723 | false, |
| 724 | false, |
| 725 | armnn::DataLayout::NDHWC, |
| 726 | armnn::DataLayout::NHWC); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 727 | |
| 728 | float qScale = 0.0f; |
| 729 | int32_t qOffset = 0; |
| 730 | |
| 731 | switch(ArmnnType) |
| 732 | { |
| 733 | case armnn::DataType::QAsymmS8: |
| 734 | case armnn::DataType::QAsymmU8: |
| 735 | case armnn::DataType::QSymmS16: |
| 736 | qScale = 1.0f; |
| 737 | break; |
| 738 | default: |
| 739 | break; |
| 740 | } |
| 741 | |
| 742 | armnn::TensorInfo inputXInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset); |
| 743 | armnn::TensorInfo inputYInfo({1,2,2,2}, ArmnnType, qScale, qOffset); |
| 744 | armnn::TensorInfo outputInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset); |
| 745 | |
| 746 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 747 | 1, 20, |
| 748 | 3, 22, |
| 749 | |
| 750 | 2, 21, |
| 751 | 4, 23 |
| 752 | }, qScale, qOffset); |
| 753 | |
| 754 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 755 | 5, 24, |
| 756 | 7, 26, |
| 757 | |
| 758 | 6, 25, |
| 759 | 8, 27 |
| 760 | }, qScale, qOffset); |
| 761 | |
| 762 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 763 | 23, 1030, |
| 764 | 31, 1114, |
| 765 | |
| 766 | 34, 1079, |
| 767 | 46, 1167 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 768 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 769 | |
| 770 | return BatchMatMulTestImpl<ArmnnType, T, 5>(workloadFactory, |
| 771 | memoryManager, |
| 772 | tensorHandleFactory, |
| 773 | descriptor, |
| 774 | inputX, |
| 775 | inputY, |
| 776 | outputExpected, |
| 777 | inputXInfo, |
| 778 | inputYInfo, |
| 779 | outputInfo); |
| 780 | } |
| 781 | |
| 782 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 5> |
| 783 | BatchMatMulNDHWCNHWCTest<armnn::DataType::BFloat16>( |
| 784 | armnn::IWorkloadFactory& workloadFactory, |
| 785 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 786 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 787 | |
| 788 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 5> |
| 789 | BatchMatMulNDHWCNHWCTest<armnn::DataType::Float32>( |
| 790 | armnn::IWorkloadFactory& workloadFactory, |
| 791 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 792 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 793 | |
| 794 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 5> |
| 795 | BatchMatMulNDHWCNHWCTest<armnn::DataType::Float16>( |
| 796 | armnn::IWorkloadFactory& workloadFactory, |
| 797 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 798 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 799 | |
| 800 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 5> |
| 801 | BatchMatMulNDHWCNHWCTest<armnn::DataType::QAsymmS8>( |
| 802 | armnn::IWorkloadFactory& workloadFactory, |
| 803 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 804 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 805 | |
| 806 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 5> |
| 807 | BatchMatMulNDHWCNHWCTest<armnn::DataType::QAsymmU8>( |
| 808 | armnn::IWorkloadFactory& workloadFactory, |
| 809 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 810 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 811 | |
| 812 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 5> |
| 813 | BatchMatMulNDHWCNHWCTest<armnn::DataType::QSymmS16>( |
| 814 | armnn::IWorkloadFactory& workloadFactory, |
| 815 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 816 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 817 | |
| 818 | template<armnn::DataType ArmnnType, typename T> |
| 819 | LayerTestResult<T, 2> BatchMatMul2DTinyTest( |
| 820 | armnn::IWorkloadFactory& workloadFactory, |
| 821 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 822 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 823 | { |
| 824 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 825 | |
| 826 | float qScale = 0.0f; |
| 827 | int32_t qOffset = 0; |
| 828 | |
| 829 | switch(ArmnnType) |
| 830 | { |
| 831 | case armnn::DataType::QAsymmS8: |
| 832 | case armnn::DataType::QAsymmU8: |
| 833 | case armnn::DataType::QSymmS16: |
| 834 | qScale = 1.0f; |
| 835 | break; |
| 836 | default: |
| 837 | break; |
| 838 | } |
| 839 | |
| 840 | armnn::TensorInfo inputXInfo({1,1}, ArmnnType, qScale, qOffset); |
| 841 | armnn::TensorInfo inputYInfo({1,1}, ArmnnType, qScale, qOffset); |
| 842 | armnn::TensorInfo outputInfo({1,1}, ArmnnType, qScale, qOffset); |
| 843 | |
| 844 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 845 | 3 |
| 846 | }, qScale, qOffset); |
| 847 | |
| 848 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 849 | 5 |
| 850 | }, qScale, qOffset); |
| 851 | |
| 852 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 853 | 15 |
| 854 | }, qScale, qOffset); |
| 855 | |
| 856 | return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory, |
| 857 | memoryManager, |
| 858 | tensorHandleFactory, |
| 859 | descriptor, |
| 860 | inputX, |
| 861 | inputY, |
| 862 | outputExpected, |
| 863 | inputXInfo, |
| 864 | inputYInfo, |
| 865 | outputInfo); |
| 866 | } |
| 867 | |
| 868 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2> |
| 869 | BatchMatMul2DTinyTest<armnn::DataType::BFloat16>( |
| 870 | armnn::IWorkloadFactory& workloadFactory, |
| 871 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 872 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 873 | |
| 874 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2> |
| 875 | BatchMatMul2DTinyTest<armnn::DataType::Float32>( |
| 876 | armnn::IWorkloadFactory& workloadFactory, |
| 877 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 878 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 879 | |
| 880 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2> |
| 881 | BatchMatMul2DTinyTest<armnn::DataType::Float16>( |
| 882 | armnn::IWorkloadFactory& workloadFactory, |
| 883 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 884 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 885 | |
| 886 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2> |
| 887 | BatchMatMul2DTinyTest<armnn::DataType::QAsymmS8>( |
| 888 | armnn::IWorkloadFactory& workloadFactory, |
| 889 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 890 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 891 | |
| 892 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2> |
| 893 | BatchMatMul2DTinyTest<armnn::DataType::QAsymmU8>( |
| 894 | armnn::IWorkloadFactory& workloadFactory, |
| 895 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 896 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 897 | |
| 898 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2> |
| 899 | BatchMatMul2DTinyTest<armnn::DataType::QSymmS16>( |
| 900 | armnn::IWorkloadFactory& workloadFactory, |
| 901 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 902 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 903 | |
| 904 | template<armnn::DataType ArmnnType, typename T> |
| 905 | LayerTestResult<T, 3> BatchMatMul3DNonSquareTest( |
| 906 | armnn::IWorkloadFactory& workloadFactory, |
| 907 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 908 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 909 | { |
| 910 | auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing |
| 911 | |
| 912 | float qScale = 0.0f; |
| 913 | int32_t qOffset = 0; |
| 914 | |
| 915 | switch(ArmnnType) |
| 916 | { |
| 917 | case armnn::DataType::QAsymmS8: |
| 918 | case armnn::DataType::QAsymmU8: |
| 919 | case armnn::DataType::QSymmS16: |
| 920 | qScale = 1.0f; |
| 921 | break; |
| 922 | default: |
| 923 | break; |
| 924 | } |
| 925 | |
| 926 | armnn::TensorInfo inputXInfo({2,5,3}, ArmnnType, qScale, qOffset); |
| 927 | armnn::TensorInfo inputYInfo({2,3,4}, ArmnnType, qScale, qOffset); |
| 928 | armnn::TensorInfo outputInfo({2,5,4}, ArmnnType, qScale, qOffset); |
| 929 | |
| 930 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 931 | 8, 8, 4, |
| 932 | 6, 1, 3, |
| 933 | 8, 8, 3, |
| 934 | 8, 9, 8, |
| 935 | 5, 4, 4, |
| 936 | |
| 937 | 1, 8, 5, |
| 938 | 7, 1, 1, |
| 939 | 8, 7, 9, |
| 940 | 3, 2, 7, |
| 941 | 8, 5, 3 |
| 942 | }, qScale, qOffset); |
| 943 | |
| 944 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 945 | 6, 2, 3, 2, |
| 946 | 6, 2, 2, 8, |
| 947 | 3, 7, 8, 1, |
| 948 | |
| 949 | 7, 2, 9, 5, |
| 950 | 2, 3, 1, 3, |
| 951 | 2, 7, 7, 5 |
| 952 | }, qScale, qOffset); |
| 953 | |
| 954 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 955 | 108, 60, 72, 84, |
| 956 | 51, 35, 44, 23, |
| 957 | 105, 53, 64, 83, |
| 958 | 126, 90, 106, 96, |
| 959 | 66, 46, 55, 46, |
| 960 | |
| 961 | 33, 61, 52, 54, |
| 962 | 53, 24, 71, 43, |
| 963 | 88, 100, 142, 106, |
| 964 | 39, 61, 78, 56, |
| 965 | 72, 52, 98, 70 |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 966 | }, qScale, qOffset); |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 967 | |
| 968 | return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory, |
| 969 | memoryManager, |
| 970 | tensorHandleFactory, |
| 971 | descriptor, |
| 972 | inputX, |
| 973 | inputY, |
| 974 | outputExpected, |
| 975 | inputXInfo, |
| 976 | inputYInfo, |
| 977 | outputInfo); |
| 978 | } |
| 979 | |
| 980 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3> |
| 981 | BatchMatMul3DNonSquareTest<armnn::DataType::BFloat16>( |
| 982 | armnn::IWorkloadFactory& workloadFactory, |
| 983 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 984 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 985 | |
| 986 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3> |
| 987 | BatchMatMul3DNonSquareTest<armnn::DataType::Float32>( |
| 988 | armnn::IWorkloadFactory& workloadFactory, |
| 989 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 990 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 991 | |
| 992 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3> |
| 993 | BatchMatMul3DNonSquareTest<armnn::DataType::Float16>( |
| 994 | armnn::IWorkloadFactory& workloadFactory, |
| 995 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 996 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 997 | |
| 998 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3> |
| 999 | BatchMatMul3DNonSquareTest<armnn::DataType::QAsymmS8>( |
| 1000 | armnn::IWorkloadFactory& workloadFactory, |
| 1001 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1002 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1003 | |
| 1004 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3> |
| 1005 | BatchMatMul3DNonSquareTest<armnn::DataType::QAsymmU8>( |
| 1006 | armnn::IWorkloadFactory& workloadFactory, |
| 1007 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1008 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1009 | |
| 1010 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3> |
| 1011 | BatchMatMul3DNonSquareTest<armnn::DataType::QSymmS16>( |
| 1012 | armnn::IWorkloadFactory& workloadFactory, |
| 1013 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
Samuel Yap | dc8ed9d | 2022-08-08 14:07:42 +0100 | [diff] [blame] | 1014 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1015 | |
| 1016 | template<armnn::DataType ArmnnType, typename T> |
| 1017 | LayerTestResult<T, 2> BatchMatMul2DTranspSimpleTest( |
| 1018 | armnn::IWorkloadFactory& workloadFactory, |
| 1019 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1020 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 1021 | { |
| 1022 | auto descriptor = armnn::BatchMatMulDescriptor(true, |
| 1023 | false, |
| 1024 | false, |
| 1025 | false); |
| 1026 | |
| 1027 | float qScale = 0.0f; |
| 1028 | int32_t qOffset = 0; |
| 1029 | |
| 1030 | switch(ArmnnType) |
| 1031 | { |
| 1032 | case armnn::DataType::QAsymmS8: |
| 1033 | case armnn::DataType::QAsymmU8: |
| 1034 | case armnn::DataType::QSymmS16: |
| 1035 | qScale = 1.0f; |
| 1036 | break; |
| 1037 | default: |
| 1038 | break; |
| 1039 | } |
| 1040 | |
| 1041 | armnn::TensorInfo inputXInfo({2,3}, ArmnnType, qScale, qOffset); |
| 1042 | armnn::TensorInfo inputYInfo({2,3}, ArmnnType, qScale, qOffset); |
| 1043 | armnn::TensorInfo outputInfo({3,3}, ArmnnType, qScale, qOffset); |
| 1044 | |
| 1045 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 1046 | 1, 2, 3, |
| 1047 | 4, 5, 6 |
| 1048 | }, qScale, qOffset); |
| 1049 | |
| 1050 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 1051 | 7, 8, 9, |
| 1052 | 10, 11, 12 |
| 1053 | }, qScale, qOffset); |
| 1054 | |
| 1055 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 1056 | 47, 52, 57, |
| 1057 | 64, 71, 78, |
| 1058 | 81, 90, 99 |
| 1059 | }, qScale, qOffset); |
| 1060 | |
| 1061 | return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory, |
| 1062 | memoryManager, |
| 1063 | tensorHandleFactory, |
| 1064 | descriptor, |
| 1065 | inputX, |
| 1066 | inputY, |
| 1067 | outputExpected, |
| 1068 | inputXInfo, |
| 1069 | inputYInfo, |
| 1070 | outputInfo); |
| 1071 | } |
| 1072 | |
| 1073 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2> |
| 1074 | BatchMatMul2DTranspSimpleTest<armnn::DataType::BFloat16>( |
| 1075 | armnn::IWorkloadFactory& workloadFactory, |
| 1076 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1077 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1078 | |
| 1079 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2> |
| 1080 | BatchMatMul2DTranspSimpleTest<armnn::DataType::Float32>( |
| 1081 | armnn::IWorkloadFactory& workloadFactory, |
| 1082 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1083 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1084 | |
| 1085 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2> |
| 1086 | BatchMatMul2DTranspSimpleTest<armnn::DataType::Float16>( |
| 1087 | armnn::IWorkloadFactory& workloadFactory, |
| 1088 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1089 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1090 | |
| 1091 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2> |
| 1092 | BatchMatMul2DTranspSimpleTest<armnn::DataType::QAsymmS8>( |
| 1093 | armnn::IWorkloadFactory& workloadFactory, |
| 1094 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1095 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1096 | |
| 1097 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2> |
| 1098 | BatchMatMul2DTranspSimpleTest<armnn::DataType::QAsymmU8>( |
| 1099 | armnn::IWorkloadFactory& workloadFactory, |
| 1100 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1101 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1102 | |
| 1103 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2> |
| 1104 | BatchMatMul2DTranspSimpleTest<armnn::DataType::QSymmS16>( |
| 1105 | armnn::IWorkloadFactory& workloadFactory, |
| 1106 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1107 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1108 | |
| 1109 | template<armnn::DataType ArmnnType, typename T> |
| 1110 | LayerTestResult<T, 2> BatchMatMul2DAdjointSimpleTest( |
| 1111 | armnn::IWorkloadFactory& workloadFactory, |
| 1112 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1113 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 1114 | { |
| 1115 | auto descriptor = armnn::BatchMatMulDescriptor(false, |
| 1116 | false, |
| 1117 | true, |
| 1118 | false); |
| 1119 | |
| 1120 | float qScale = 0.0f; |
| 1121 | int32_t qOffset = 0; |
| 1122 | |
| 1123 | switch(ArmnnType) |
| 1124 | { |
| 1125 | case armnn::DataType::QAsymmS8: |
| 1126 | case armnn::DataType::QAsymmU8: |
| 1127 | case armnn::DataType::QSymmS16: |
| 1128 | qScale = 1.0f; |
| 1129 | break; |
| 1130 | default: |
| 1131 | break; |
| 1132 | } |
| 1133 | |
| 1134 | armnn::TensorInfo inputXInfo({3,3}, ArmnnType, qScale, qOffset); |
| 1135 | armnn::TensorInfo inputYInfo({3,3}, ArmnnType, qScale, qOffset); |
| 1136 | armnn::TensorInfo outputInfo({3,3}, ArmnnType, qScale, qOffset); |
| 1137 | |
| 1138 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 1139 | 3, 1, 1, |
| 1140 | 1, 3, -1, |
| 1141 | 2, 4, 1 |
| 1142 | }, qScale, qOffset); |
| 1143 | |
| 1144 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 1145 | 1, 0, 0, |
| 1146 | 0, 1, 0, |
| 1147 | 0, 0, 1 |
| 1148 | }, qScale, qOffset); |
| 1149 | |
| 1150 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 1151 | 7, 3, -4, |
| 1152 | -3, 1, 4, |
| 1153 | -2, -10, 8 |
| 1154 | }, qScale, qOffset); |
| 1155 | |
| 1156 | switch (ArmnnType) |
| 1157 | { |
| 1158 | case armnn::DataType::QAsymmU8: |
| 1159 | outputExpected = armnnUtils::QuantizedVector<T>({ |
| 1160 | 3, 3, 0, |
| 1161 | 0, 1, 1, |
| 1162 | 0, 0, 8 |
| 1163 | }, qScale, qOffset); |
| 1164 | break; |
| 1165 | default: |
| 1166 | break; |
| 1167 | } |
| 1168 | |
| 1169 | return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory, |
| 1170 | memoryManager, |
| 1171 | tensorHandleFactory, |
| 1172 | descriptor, |
| 1173 | inputX, |
| 1174 | inputY, |
| 1175 | outputExpected, |
| 1176 | inputXInfo, |
| 1177 | inputYInfo, |
| 1178 | outputInfo); |
| 1179 | } |
| 1180 | |
| 1181 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2> |
| 1182 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::BFloat16>( |
| 1183 | armnn::IWorkloadFactory& workloadFactory, |
| 1184 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1185 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1186 | |
| 1187 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2> |
| 1188 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::Float32>( |
| 1189 | armnn::IWorkloadFactory& workloadFactory, |
| 1190 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1191 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1192 | |
| 1193 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2> |
| 1194 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::Float16>( |
| 1195 | armnn::IWorkloadFactory& workloadFactory, |
| 1196 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1197 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1198 | |
| 1199 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2> |
| 1200 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::QAsymmS8>( |
| 1201 | armnn::IWorkloadFactory& workloadFactory, |
| 1202 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1203 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1204 | |
| 1205 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2> |
| 1206 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::QAsymmU8>( |
| 1207 | armnn::IWorkloadFactory& workloadFactory, |
| 1208 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1209 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1210 | |
| 1211 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2> |
| 1212 | BatchMatMul2DAdjointSimpleTest<armnn::DataType::QSymmS16>( |
| 1213 | armnn::IWorkloadFactory& workloadFactory, |
| 1214 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1215 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1216 | |
| 1217 | template<armnn::DataType ArmnnType, typename T> |
| 1218 | LayerTestResult<T, 4> BatchMatMulNHWCParamsTest( |
| 1219 | armnn::IWorkloadFactory& workloadFactory, |
| 1220 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1221 | const armnn::ITensorHandleFactory& tensorHandleFactory) |
| 1222 | { |
| 1223 | auto descriptor = armnn::BatchMatMulDescriptor(false, |
| 1224 | true, |
| 1225 | true, |
| 1226 | false, |
| 1227 | armnn::DataLayout::NHWC, |
| 1228 | armnn::DataLayout::NHWC); |
| 1229 | |
| 1230 | float qScale = 0.0f; |
| 1231 | int32_t qOffset = 0; |
| 1232 | |
| 1233 | switch(ArmnnType) |
| 1234 | { |
| 1235 | case armnn::DataType::QAsymmS8: |
| 1236 | case armnn::DataType::QAsymmU8: |
| 1237 | case armnn::DataType::QSymmS16: |
| 1238 | qScale = 1.0f; |
| 1239 | break; |
| 1240 | default: |
| 1241 | break; |
| 1242 | } |
| 1243 | |
| 1244 | armnn::TensorInfo inputXInfo({1,4,4,2}, ArmnnType, qScale, qOffset); |
| 1245 | armnn::TensorInfo inputYInfo({2,2,4,1}, ArmnnType, qScale, qOffset); |
| 1246 | armnn::TensorInfo outputInfo({2,4,2,2}, ArmnnType, qScale, qOffset); |
| 1247 | |
| 1248 | std::vector<T> inputX = armnnUtils::QuantizedVector<T>({ |
| 1249 | 1, -3, 1, 4, 4, 9, 1, 2, |
| 1250 | 2, 4, 2, 2, 10, 7, 6, -5, |
| 1251 | 3, 8, 9, 9, 21, 1, 17, 7, |
| 1252 | 5, 11, 11, 8, 29, 3, 23, 6 |
| 1253 | }, qScale, qOffset); |
| 1254 | |
| 1255 | std::vector<T> inputY = armnnUtils::QuantizedVector<T>({ |
| 1256 | 1, 2, 3, 4, |
| 1257 | 5, 6, 7, 8, |
| 1258 | |
| 1259 | 9, 10, 11, 12, |
| 1260 | 13, 14, 15, 16 |
| 1261 | }, qScale, qOffset); |
| 1262 | |
| 1263 | std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({ |
| 1264 | 28, 625, 140, 585, |
| 1265 | 8, 110, -8, 1662, |
| 1266 | -24, 401, -120, 921, |
| 1267 | 12, 131, 108, -501, |
| 1268 | |
| 1269 | 252, 545, 364, 505, |
| 1270 | -24, 3214, -40, 4766, |
| 1271 | -216, 1441, -312, 1961, |
| 1272 | 204, -1133, 300, -1765 |
| 1273 | }, qScale, qOffset); |
| 1274 | |
| 1275 | switch (ArmnnType) |
| 1276 | { |
| 1277 | case armnn::DataType::QAsymmU8: |
| 1278 | outputExpected = armnnUtils::QuantizedVector<T>({ |
| 1279 | 28, 80, 140, 80, |
| 1280 | 8, 45, 0, 255, |
| 1281 | 0, 18, 0, 18, |
| 1282 | 12, 0, 108, 0, |
| 1283 | |
| 1284 | 252, 80, 255, 80, |
| 1285 | 0, 255, 0, 255, |
| 1286 | 0, 18, 0, 18, |
| 1287 | 204, 0, 255, 0 |
| 1288 | }, qScale, qOffset); |
| 1289 | break; |
| 1290 | default: |
| 1291 | break; |
| 1292 | } |
| 1293 | |
| 1294 | return BatchMatMulTestImpl<ArmnnType, T, 4>(workloadFactory, |
| 1295 | memoryManager, |
| 1296 | tensorHandleFactory, |
| 1297 | descriptor, |
| 1298 | inputX, |
| 1299 | inputY, |
| 1300 | outputExpected, |
| 1301 | inputXInfo, |
| 1302 | inputYInfo, |
| 1303 | outputInfo); |
| 1304 | } |
| 1305 | |
| 1306 | template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 4> |
| 1307 | BatchMatMulNHWCParamsTest<armnn::DataType::BFloat16>( |
| 1308 | armnn::IWorkloadFactory& workloadFactory, |
| 1309 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1310 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1311 | |
| 1312 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4> |
| 1313 | BatchMatMulNHWCParamsTest<armnn::DataType::Float32>( |
| 1314 | armnn::IWorkloadFactory& workloadFactory, |
| 1315 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1316 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1317 | |
| 1318 | template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 4> |
| 1319 | BatchMatMulNHWCParamsTest<armnn::DataType::Float16>( |
| 1320 | armnn::IWorkloadFactory& workloadFactory, |
| 1321 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1322 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1323 | |
| 1324 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 4> |
| 1325 | BatchMatMulNHWCParamsTest<armnn::DataType::QAsymmS8>( |
| 1326 | armnn::IWorkloadFactory& workloadFactory, |
| 1327 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1328 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1329 | |
| 1330 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4> |
| 1331 | BatchMatMulNHWCParamsTest<armnn::DataType::QAsymmU8>( |
| 1332 | armnn::IWorkloadFactory& workloadFactory, |
| 1333 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 1334 | const armnn::ITensorHandleFactory& tensorHandleFactory); |
| 1335 | |
| 1336 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 4> |
| 1337 | BatchMatMulNHWCParamsTest<armnn::DataType::QSymmS16>( |
| 1338 | armnn::IWorkloadFactory& workloadFactory, |
| 1339 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 1340 | const armnn::ITensorHandleFactory& tensorHandleFactory); |