Georgios Pinitas | bdcdc39 | 2021-04-22 16:42:03 +0100 | [diff] [blame] | 1 | { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2 | "common": [ |
| 3 | "src/common/cpuinfo/CpuInfo.cpp", |
| 4 | "src/common/cpuinfo/CpuModel.cpp", |
| 5 | "src/common/cpuinfo/CpuIsaInfo.cpp", |
| 6 | "src/common/utils/LegacySupport.cpp", |
| 7 | "src/common/AllocatorWrapper.cpp", |
| 8 | "src/common/ITensorV2.cpp", |
Georgios Pinitas | 4164814 | 2021-08-03 08:24:00 +0100 | [diff] [blame] | 9 | "src/common/TensorPack.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 10 | "src/common/IOperator.cpp", |
| 11 | "src/core/AccessWindowAutoPadding.cpp", |
| 12 | "src/core/AccessWindowStatic.cpp", |
| 13 | "src/core/AccessWindowTranspose.cpp", |
| 14 | "src/core/Error.cpp", |
| 15 | "src/core/GPUTarget.cpp", |
| 16 | "src/core/Helpers.cpp", |
| 17 | "src/core/IAccessWindow.cpp", |
| 18 | "src/core/IKernel.cpp", |
| 19 | "src/core/ITensor.cpp", |
| 20 | "src/core/ITensorPack.cpp", |
| 21 | "src/core/Rounding.cpp", |
| 22 | "src/core/Size2D.cpp", |
| 23 | "src/core/SubTensorInfo.cpp", |
| 24 | "src/core/TensorInfo.cpp", |
| 25 | "src/core/Utils.cpp", |
| 26 | "src/core/Validate.cpp", |
| 27 | "src/core/Version.cpp", |
| 28 | "src/core/helpers/SoftmaxHelpers.cpp", |
| 29 | "src/core/helpers/WindowHelpers.cpp", |
| 30 | "src/core/utils/AssemblyUtils.cpp", |
| 31 | "src/core/utils/ScaleUtils.cpp", |
| 32 | "src/core/utils/helpers/fft.cpp", |
| 33 | "src/core/utils/helpers/tensor_transform.cpp", |
| 34 | "src/core/utils/io/FileHandler.cpp", |
| 35 | "src/core/utils/misc/MMappedFile.cpp", |
| 36 | "src/core/utils/quantization/AsymmHelpers.cpp", |
| 37 | "src/core/CPP/CPPTypes.cpp", |
| 38 | "src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp", |
| 39 | "src/core/CPP/kernels/CPPNonMaximumSuppressionKernel.cpp", |
| 40 | "src/core/CPP/kernels/CPPPermuteKernel.cpp", |
| 41 | "src/core/CPP/kernels/CPPTopKVKernel.cpp", |
| 42 | "src/core/CPP/kernels/CPPUpsampleKernel.cpp", |
| 43 | "src/runtime/Allocator.cpp", |
| 44 | "src/runtime/BlobLifetimeManager.cpp", |
| 45 | "src/runtime/BlobMemoryPool.cpp", |
| 46 | "src/runtime/ISimpleLifetimeManager.cpp", |
| 47 | "src/runtime/ITensorAllocator.cpp", |
| 48 | "src/runtime/IWeightsManager.cpp", |
| 49 | "src/runtime/IScheduler.cpp", |
| 50 | "src/runtime/Memory.cpp", |
| 51 | "src/runtime/MemoryManagerOnDemand.cpp", |
| 52 | "src/runtime/OffsetLifetimeManager.cpp", |
| 53 | "src/runtime/OffsetMemoryPool.cpp", |
| 54 | "src/runtime/OperatorTensor.cpp", |
| 55 | "src/runtime/PoolManager.cpp", |
| 56 | "src/runtime/RuntimeContext.cpp", |
| 57 | "src/runtime/Scheduler.cpp", |
| 58 | "src/runtime/SchedulerFactory.cpp", |
| 59 | "src/runtime/SchedulerUtils.cpp", |
| 60 | "src/runtime/SubTensor.cpp", |
| 61 | "src/runtime/Tensor.cpp", |
| 62 | "src/runtime/TensorAllocator.cpp", |
| 63 | "src/runtime/Utils.cpp", |
| 64 | "src/runtime/CPP/ICPPSimpleFunction.cpp", |
| 65 | "src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp", |
| 66 | "src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp", |
| 67 | "src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp", |
| 68 | "src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp", |
| 69 | "src/runtime/CPP/functions/CPPPermute.cpp", |
| 70 | "src/runtime/CPP/functions/CPPTopKV.cpp", |
| 71 | "src/runtime/CPP/functions/CPPUpsample.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 72 | ], |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 73 | "logging": [ |
| 74 | "src/core/utils/logging/FilePrinter.cpp", |
| 75 | "src/core/utils/logging/Helpers.cpp", |
| 76 | "src/core/utils/logging/Logger.cpp", |
| 77 | "src/core/utils/logging/LoggerRegistry.cpp" |
| 78 | ], |
| 79 | "scheduler": { |
| 80 | "single": [ "src/runtime/CPP/SingleThreadScheduler.cpp" ], |
| 81 | "threads": [ "src/runtime/CPP/CPPScheduler.cpp" ], |
| 82 | "omp": [ "src/runtime/OMP/OMPScheduler.cpp"] |
| 83 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 84 | "c_api": { |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 85 | "common": [ |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 86 | "src/c/AclContext.cpp", |
Georgios Pinitas | 06ac6e4 | 2021-07-05 08:08:52 +0100 | [diff] [blame] | 87 | "src/c/AclOperator.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 88 | "src/c/AclQueue.cpp", |
| 89 | "src/c/AclTensor.cpp", |
| 90 | "src/c/AclTensorPack.cpp", |
| 91 | "src/c/AclVersion.cpp" |
Georgios Pinitas | bdcdc39 | 2021-04-22 16:42:03 +0100 | [diff] [blame] | 92 | ], |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 93 | "cpu": [ |
| 94 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 95 | "gpu": [ |
| 96 | "src/c/cl/AclOpenClExt.cpp" |
Georgios Pinitas | 4164814 | 2021-08-03 08:24:00 +0100 | [diff] [blame] | 97 | ], |
| 98 | "operators": |
| 99 | [ |
| 100 | "src/c/operators/AclActivation.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 101 | ] |
| 102 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 103 | "high_priority": [ |
| 104 | "Activation", |
| 105 | "DepthwiseConv2d", |
| 106 | "Conv2d", |
| 107 | "Permute", |
| 108 | "Pool2d", |
| 109 | "Reshape" |
| 110 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 111 | "gpu": { |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 112 | "common": [ |
| 113 | "src/core/CL/CLCompileContext.cpp", |
Giorgio Arena | ba2dd82 | 2021-07-28 16:10:03 +0100 | [diff] [blame] | 114 | "src/core/CL/DefaultLWSHeuristics.cpp", |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 115 | "src/core/CL/CLHelpers.cpp", |
| 116 | "src/core/CL/CLKernelLibrary.cpp", |
| 117 | "src/core/CL/CLUtils.cpp", |
| 118 | "src/core/CL/ICLKernel.cpp", |
| 119 | "src/core/CL/ICLSimple2DKernel.cpp", |
| 120 | "src/core/CL/ICLSimple3DKernel.cpp", |
| 121 | "src/core/CL/ICLSimpleKernel.cpp", |
| 122 | "src/core/CL/ICLTensor.cpp", |
| 123 | "src/core/CL/OpenCL.cpp", |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 124 | "src/gpu/cl/ClContext.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 125 | "src/gpu/cl/ClKernelLibrary.cpp", |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 126 | "src/gpu/cl/ClQueue.cpp", |
| 127 | "src/gpu/cl/ClTensor.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 128 | "src/core/CL/kernels/CLFillBorderKernel.cpp", |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 129 | "src/runtime/CL/CLBufferAllocator.cpp", |
| 130 | "src/runtime/CL/CLGEMMHeuristicsHandle.cpp", |
| 131 | "src/runtime/CL/CLHelpers.cpp", |
| 132 | "src/runtime/CL/CLMemory.cpp", |
| 133 | "src/runtime/CL/CLMemoryRegion.cpp", |
| 134 | "src/runtime/CL/CLOperator.cpp", |
| 135 | "src/runtime/CL/CLRuntimeContext.cpp", |
| 136 | "src/runtime/CL/CLScheduler.cpp", |
| 137 | "src/runtime/CL/CLSubTensor.cpp", |
| 138 | "src/runtime/CL/CLTensor.cpp", |
| 139 | "src/runtime/CL/CLTensorAllocator.cpp", |
| 140 | "src/runtime/CL/CLTuner.cpp", |
| 141 | "src/runtime/CL/ICLSimpleFunction.cpp", |
| 142 | "src/runtime/CL/Utils.cpp", |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 143 | "src/runtime/CL/mlgo/HeuristicTree.cpp", |
| 144 | "src/runtime/CL/mlgo/MLGOHeuristics.cpp", |
| 145 | "src/runtime/CL/mlgo/MLGOParser.cpp", |
| 146 | "src/runtime/CL/mlgo/Utils.cpp", |
| 147 | "src/runtime/CL/tuners/CLTuningParametersList.cpp" |
| 148 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 149 | "operators": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 150 | "Activation":{ |
| 151 | "files": { |
| 152 | "common": [ |
| 153 | "src/gpu/cl/kernels/ClActivationKernel.cpp", |
| 154 | "src/gpu/cl/operators/ClActivation.cpp", |
| 155 | "src/runtime/CL/functions/CLActivationLayer.cpp" |
| 156 | ] |
| 157 | } |
| 158 | }, |
| 159 | "ArgMinMax": { |
| 160 | "deps": [ "Reshape" ], |
| 161 | "files": { |
| 162 | "common": [ |
| 163 | "src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp", |
| 164 | "src/runtime/CL/functions/CLArgMinMaxLayer.cpp" |
| 165 | ] |
| 166 | } |
| 167 | }, |
| 168 | "Add": { |
| 169 | "files": { |
| 170 | "common": [ |
| 171 | "src/gpu/cl/kernels/ClElementwiseKernel.cpp", |
| 172 | "src/gpu/cl/operators/ClAdd.cpp" |
| 173 | ] |
| 174 | } |
| 175 | }, |
| 176 | "BatchNormalization": { |
| 177 | "files": { |
| 178 | "common": [ |
| 179 | "src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp", |
| 180 | "src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp", |
| 181 | "src/runtime/CL/functions/CLBatchNormalizationLayer.cpp", |
| 182 | "src/runtime/CL/functions/CLFuseBatchNormalization.cpp" |
| 183 | ] |
| 184 | } |
| 185 | }, |
| 186 | "BatchToSpace": { |
| 187 | "files": { |
| 188 | "common": [ |
| 189 | "src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp", |
| 190 | "src/runtime/CL/functions/CLBatchToSpaceLayer.cpp" |
| 191 | ] |
| 192 | } |
| 193 | }, |
| 194 | "Bitwise": { |
| 195 | "files": { |
| 196 | "common": [ "src/core/CL/kernels/CLBitwiseKernel.cpp" ] |
| 197 | } |
| 198 | }, |
| 199 | "BitwiseAnd": { |
| 200 | "deps": [ "Bitwise" ], |
| 201 | "files": { |
| 202 | "common": [ "src/runtime/CL/functions/CLBitwiseAnd.cpp" ] |
| 203 | } |
| 204 | }, |
| 205 | "BitwiseNot": { |
| 206 | "deps": [ "Bitwise" ], |
| 207 | "files": { |
| 208 | "common": [ "src/runtime/CL/functions/CLBitwiseNot.cpp" ] |
| 209 | } |
| 210 | }, |
| 211 | "BitwiseOr": { |
| 212 | "deps": [ "Bitwise" ], |
| 213 | "files": { |
| 214 | "common": [ "src/runtime/CL/functions/CLBitwiseOr.cpp" ] |
| 215 | } |
| 216 | }, |
| 217 | "BitwiseXor": { |
| 218 | "deps": [ "Bitwise" ], |
| 219 | "files": { |
| 220 | "common": [ "src/runtime/CL/functions/CLBitwiseXor.cpp" ] |
| 221 | } |
| 222 | }, |
| 223 | "BoundingBoxTransform": { |
| 224 | "files": { |
| 225 | "common": [ |
| 226 | "src/core/CL/kernels/CLBoundingBoxTransformKernel.cpp", |
| 227 | "src/runtime/CL/functions/CLBoundingBoxTransform.cpp" |
| 228 | ] |
| 229 | } |
| 230 | }, |
| 231 | "Cast": { |
| 232 | "files": { |
| 233 | "common": [ |
| 234 | "src/gpu/cl/kernels/ClCastKernel.cpp", |
| 235 | "src/gpu/cl/operators/ClCast.cpp", |
| 236 | "src/runtime/CL/functions/CLCast.cpp" |
| 237 | ] |
| 238 | } |
| 239 | }, |
| 240 | "ChannelShuffle": { |
| 241 | "files": { |
| 242 | "common": [ |
| 243 | "src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp", |
| 244 | "src/runtime/CL/functions/CLChannelShuffleLayer.cpp" |
| 245 | ] |
| 246 | } |
| 247 | }, |
| 248 | "Comparison": { |
| 249 | "files": { |
| 250 | "common": [ |
| 251 | "src/core/CL/kernels/CLComparisonKernel.cpp", |
| 252 | "src/runtime/CL/functions/CLComparison.cpp" |
| 253 | ] |
| 254 | } |
| 255 | }, |
| 256 | "Concatenate": { |
| 257 | "files": { |
| 258 | "common": [ |
| 259 | "src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp", |
| 260 | "src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp", |
| 261 | "src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp", |
| 262 | "src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp", |
| 263 | "src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp", |
| 264 | "src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp", |
| 265 | "src/gpu/cl/operators/ClConcatenate.cpp", |
| 266 | "src/runtime/CL/functions/CLConcatenateLayer.cpp" |
| 267 | ] |
| 268 | } |
| 269 | }, |
| 270 | "Conv2d": { |
| 271 | "deps": [ |
| 272 | "Activation", |
| 273 | "ElementwiseBinary", |
| 274 | "FFT2D", |
| 275 | "Gemm", |
| 276 | "Mul", |
| 277 | "Pad", |
| 278 | "Permute", |
| 279 | "Reduction", |
| 280 | "Reshape", |
| 281 | "Reverse", |
| 282 | "Slice" |
| 283 | ], |
| 284 | "files": { |
| 285 | "common": [ |
| 286 | "src/gpu/cl/kernels/ClDirectConv2dKernel.cpp", |
| 287 | "src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp", |
| 288 | "src/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp", |
| 289 | "src/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp", |
| 290 | "src/gpu/cl/kernels/ClIm2ColKernel.cpp", |
Gian Marco Iodice | 76335eb | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 291 | "src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp", |
Gian Marco Iodice | 5d01681 | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 292 | "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 293 | "src/gpu/cl/kernels/ClCol2ImKernel.cpp", |
| 294 | "src/gpu/cl/operators/ClConv2d.cpp", |
| 295 | "src/gpu/cl/operators/ClDirectConv2d.cpp", |
| 296 | "src/gpu/cl/operators/ClGemmConv2d.cpp", |
Gian Marco Iodice | 76335eb | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 297 | "src/gpu/cl/operators/ClIndirectConv2d.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 298 | "src/gpu/cl/operators/ClWinogradConv2d.cpp", |
| 299 | "src/gpu/cl/kernels/ClWeightsReshapeKernel.cpp", |
| 300 | "src/runtime/CL/functions/CLConvolutionLayer.cpp", |
| 301 | "src/runtime/CL/functions/CLDirectConvolutionLayer.cpp", |
| 302 | "src/runtime/CL/functions/CLFFTConvolutionLayer.cpp", |
| 303 | "src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp", |
Gian Marco Iodice | 76335eb | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 304 | "src/runtime/CL/functions/CLIndirectConvolutionLayer.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 305 | "src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp" |
| 306 | ] |
| 307 | } |
| 308 | }, |
Giorgio Arena | 273c28c | 2021-10-14 15:59:15 +0100 | [diff] [blame] | 309 | "Conv3d": { |
| 310 | "files": { |
| 311 | "common": [ |
| 312 | "src/gpu/cl/kernels/ClDirectConv3dKernel.cpp", |
| 313 | "src/gpu/cl/operators/ClDirectConv3d.cpp", |
| 314 | "src/runtime/CL/functions/CLConv3D.cpp" |
| 315 | ] |
| 316 | } |
| 317 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 318 | "Copy": { |
| 319 | "files": { |
| 320 | "common": [ |
| 321 | "src/gpu/cl/kernels/ClCopyKernel.cpp", |
| 322 | "src/gpu/cl/operators/ClCopy.cpp", |
| 323 | "src/runtime/CL/functions/CLCopy.cpp" |
| 324 | ] |
| 325 | } |
| 326 | }, |
| 327 | "CropResize": { |
| 328 | "deps": [ "Copy", "Fill", "Scale" ], |
| 329 | "files": { |
| 330 | "common": [ |
| 331 | "src/gpu/cl/kernels/ClCropKernel.cpp", |
| 332 | "src/gpu/cl/operators/ClCrop.cpp", |
| 333 | "src/runtime/CL/functions/CLCrop.cpp", |
| 334 | "src/runtime/CL/functions/CLCropResize.cpp" |
| 335 | ] |
| 336 | } |
| 337 | }, |
| 338 | "Deconv2d": { |
| 339 | "deps": [ "Conv2d", "Reverse", "Transpose"], |
| 340 | "files": { |
| 341 | "common": [ |
| 342 | "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp", |
| 343 | "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp", |
Gunes Bayir | ec0113d | 2022-11-09 09:26:27 +0000 | [diff] [blame] | 344 | "src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp", |
| 345 | "src/gpu/cl/operators/ClTransposedConvolution.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 346 | "src/runtime/CL/functions/CLDeconvolutionLayer.cpp", |
| 347 | "src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp", |
| 348 | "src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp", |
| 349 | "src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp" |
| 350 | ] |
| 351 | } |
| 352 | }, |
| 353 | "DepthConvert": { |
| 354 | "deps": [ "Cast"], |
| 355 | "files": { |
| 356 | "common": [ "src/runtime/CL/functions/CLDepthConvertLayer.cpp" ] |
| 357 | } |
| 358 | }, |
| 359 | "DepthToSpace": { |
| 360 | "files": { |
| 361 | "common": [ |
| 362 | "src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp", |
| 363 | "src/runtime/CL/functions/CLDepthToSpaceLayer.cpp" |
| 364 | ] |
| 365 | } |
| 366 | }, |
| 367 | "DepthwiseConv2d": { |
| 368 | "deps": [ "Permute" ], |
| 369 | "files": { |
| 370 | "common": [ |
| 371 | "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp", |
| 372 | "src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp" |
| 373 | ] |
| 374 | } |
| 375 | }, |
| 376 | "Dequantize": { |
| 377 | "files": { |
| 378 | "common": [ |
| 379 | "src/gpu/cl/kernels/ClDequantizeKernel.cpp", |
| 380 | "src/gpu/cl/operators/ClDequantize.cpp", |
| 381 | "src/runtime/CL/functions/CLDequantizationLayer.cpp" |
| 382 | ] |
| 383 | } |
| 384 | }, |
| 385 | "ElementwiseBinary": { |
| 386 | "deps": ["Add", "Sub"], |
| 387 | "files": { |
| 388 | "common": [ |
| 389 | "src/gpu/cl/kernels/ClElementwiseKernel.cpp", |
| 390 | "src/gpu/cl/operators/ClElementwiseOperations.cpp", |
| 391 | "src/runtime/CL/functions/CLElementwiseOperations.cpp" |
| 392 | ] |
| 393 | } |
| 394 | }, |
| 395 | "ElementwiseUnary":{ |
| 396 | "files": { |
| 397 | "common": [ |
| 398 | "src/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp", |
| 399 | "src/gpu/cl/operators/ClElementwiseUnary.cpp", |
| 400 | "src/runtime/CL/functions/CLElementwiseUnaryLayer.cpp" |
| 401 | ] |
| 402 | } |
| 403 | }, |
| 404 | "FFT1D": { |
| 405 | "files": { |
| 406 | "common": [ |
| 407 | "src/core/CL/kernels/CLFFTDigitReverseKernel.cpp", |
| 408 | "src/core/CL/kernels/CLFFTRadixStageKernel.cpp", |
| 409 | "src/core/CL/kernels/CLFFTScaleKernel.cpp", |
| 410 | "src/runtime/CL/functions/CLFFT1D.cpp" |
| 411 | ] |
| 412 | } |
| 413 | }, |
| 414 | "FFT2D": { |
| 415 | "deps": [ "FFT1D" ], |
| 416 | "files": { |
| 417 | "common": [ "src/runtime/CL/functions/CLFFT2D.cpp" ] |
| 418 | } |
| 419 | }, |
| 420 | "Fill": { |
| 421 | "files": { |
| 422 | "common": [ |
| 423 | "src/gpu/cl/kernels/ClFillKernel.cpp", |
| 424 | "src/gpu/cl/operators/ClFill.cpp", |
| 425 | "src/runtime/CL/functions/CLFill.cpp" |
| 426 | ] |
| 427 | } |
| 428 | }, |
| 429 | "Flatten": { |
| 430 | "files": { |
| 431 | "common": [ |
| 432 | "src/gpu/cl/operators/ClFlatten.cpp", |
| 433 | "src/runtime/CL/functions/CLFlattenLayer.cpp" |
| 434 | ] |
| 435 | } |
| 436 | }, |
| 437 | "Floor": { |
| 438 | "files": { |
| 439 | "common": [ |
| 440 | "src/gpu/cl/kernels/ClFloorKernel.cpp", |
| 441 | "src/gpu/cl/operators/ClFloor.cpp", |
| 442 | "src/runtime/CL/functions/CLFloor.cpp" |
| 443 | ] |
| 444 | } |
| 445 | }, |
| 446 | "FullyConnected": { |
| 447 | "deps": [ "Flatten", "Gemm", "Transpose"], |
| 448 | "files": { |
| 449 | "common": [ |
| 450 | "src/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp", |
| 451 | "src/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp", |
| 452 | "src/gpu/cl/operators/ClFullyConnected.cpp", |
| 453 | "src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp", |
| 454 | "src/runtime/CL/functions/CLFullyConnectedLayer.cpp" |
| 455 | ] |
| 456 | } |
| 457 | }, |
| 458 | "Gather": { |
| 459 | "files": { |
| 460 | "common": [ |
| 461 | "src/core/CL/kernels/CLGatherKernel.cpp", |
| 462 | "src/runtime/CL/functions/CLGather.cpp"] |
| 463 | } |
| 464 | }, |
| 465 | "Gemm": { |
| 466 | "deps": [ "Cast" ], |
| 467 | "files": { |
| 468 | "common": [ |
| 469 | "src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp", |
| 470 | "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp", |
| 471 | "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeMidgard.cpp", |
| 472 | "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeValhall.cpp", |
| 473 | "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedBifrost.cpp", |
| 474 | "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedValhall.cpp", |
| 475 | "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyBifrost.cpp", |
| 476 | "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp", |
| 477 | "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp", |
| 478 | "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp", |
| 479 | "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.cpp", |
Freddie Liardet | e572dff | 2022-05-16 14:09:10 +0100 | [diff] [blame] | 480 | "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 481 | "src/gpu/cl/kernels/ClGemmLowpOffsetContributionKernel.cpp", |
| 482 | "src/gpu/cl/kernels/ClGemmLowpOffsetContributionOutputStageKernel.cpp", |
| 483 | "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.cpp", |
| 484 | "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp", |
| 485 | "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 486 | "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp", |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 487 | "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 488 | "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp", |
| 489 | "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp", |
| 490 | "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.cpp", |
| 491 | "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.cpp", |
| 492 | "src/gpu/cl/kernels/ClGemmLowpReductionKernel.cpp", |
| 493 | "src/gpu/cl/operators/ClGemm.cpp", |
| 494 | "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp", |
| 495 | "src/gpu/cl/operators/ClGemmLowpOutputStage.cpp", |
| 496 | "src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp", |
| 497 | "src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp", |
| 498 | "src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp", |
| 499 | "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp", |
| 500 | "src/runtime/CL/functions/CLGEMM.cpp", |
| 501 | "src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp", |
Ramy Elgammal | df6a3b0 | 2022-11-30 16:23:10 +0000 | [diff] [blame] | 502 | "src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp", |
| 503 | "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp", |
Gian Marco Iodice | a5cb79f | 2022-12-28 13:53:51 +0000 | [diff] [blame] | 504 | "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp", |
Gian Marco Iodice | 9d3bd41 | 2022-12-30 09:45:00 +0000 | [diff] [blame] | 505 | "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigBifrost.cpp", |
| 506 | "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigValhall.cpp", |
| 507 | "src/runtime/heuristics/dwc_native/ClDWCNativeHeuristicsHelpers.cpp", |
Gian Marco Iodice | a5cb79f | 2022-12-28 13:53:51 +0000 | [diff] [blame] | 508 | "src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp" |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 509 | ] |
| 510 | } |
| 511 | }, |
Ramy Elgammal | 2b6ebfe | 2023-03-09 21:15:37 +0000 | [diff] [blame] | 512 | "MatMul": { |
| 513 | "files": { |
| 514 | "common": [ |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 515 | "src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp", |
Jakub Sujak | 1ed6a14 | 2023-04-13 21:14:42 +0100 | [diff] [blame] | 516 | "src/gpu/cl/kernels/ClMatMulNativeKernel.cpp", |
Ramy Elgammal | f26ea2f | 2023-03-24 11:42:03 +0000 | [diff] [blame] | 517 | "src/gpu/cl/operators/ClMatMul.cpp", |
| 518 | "src/runtime/CL/functions/CLMatMul.cpp" |
Ramy Elgammal | 2b6ebfe | 2023-03-09 21:15:37 +0000 | [diff] [blame] | 519 | ] |
| 520 | } |
| 521 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 522 | "GenerateProposals": { |
| 523 | "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ], |
| 524 | "files": { |
| 525 | "common": [ |
| 526 | "src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp", |
| 527 | "src/runtime/CL/functions/CLGenerateProposalsLayer.cpp" |
| 528 | ] |
| 529 | } |
| 530 | }, |
| 531 | "InstanceNormalize": { |
| 532 | "files": { |
| 533 | "common": [ |
| 534 | "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp", |
| 535 | "src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp" |
| 536 | ] |
| 537 | } |
| 538 | }, |
| 539 | "L2Normalize": { |
| 540 | "deps": [ "Reduction" ], |
| 541 | "files": { |
| 542 | "common": [ |
| 543 | "src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp", |
| 544 | "src/runtime/CL/functions/CLL2NormalizeLayer.cpp" |
| 545 | ] |
| 546 | } |
| 547 | }, |
| 548 | "Logical": { |
| 549 | "files": { |
| 550 | "common": [ |
| 551 | "src/gpu/cl/operators/ClLogicalNot.cpp", |
| 552 | "src/runtime/CL/functions/CLLogicalAnd.cpp", |
| 553 | "src/runtime/CL/functions/CLLogicalNot.cpp", |
| 554 | "src/runtime/CL/functions/CLLogicalOr.cpp" |
| 555 | ] |
| 556 | } |
| 557 | }, |
| 558 | "LSTM": { |
| 559 | "deps": [ |
| 560 | "Activation", |
| 561 | "Concatenate", |
| 562 | "Copy", |
| 563 | "Dequantize", |
| 564 | "ElementwiseBinary", |
| 565 | "Fill", |
| 566 | "FullyConnected", |
| 567 | "Gemm", |
| 568 | "MeanStdDevNormalize", |
| 569 | "Mul", |
| 570 | "Quantize", |
| 571 | "Slice", |
| 572 | "Transpose" |
| 573 | ], |
| 574 | "files": { |
| 575 | "common": [ |
| 576 | "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp", |
| 577 | "src/runtime/CL/functions/CLQLSTMLayer.cpp", |
| 578 | "src/runtime/CL/functions/CLLSTMLayer.cpp", |
| 579 | "src/runtime/CL/functions/CLLSTMLayerQuantized.cpp" |
| 580 | ] |
| 581 | } |
| 582 | }, |
| 583 | "MaxUnpool2d": { |
| 584 | "deps": [ "Fill" ], |
| 585 | "files": { |
| 586 | "common": [ |
| 587 | "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.cpp", |
| 588 | "src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp" |
| 589 | ] |
| 590 | } |
| 591 | }, |
| 592 | "MeanStdDevNormalize": { |
| 593 | "deps": [ "Reduction" ], |
| 594 | "files": { |
| 595 | "common": [ |
| 596 | "src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp", |
| 597 | "src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp", |
| 598 | "src/runtime/CL/functions/CLReduceMean.cpp" |
| 599 | ] |
| 600 | } |
| 601 | }, |
| 602 | "Mul": { |
| 603 | "files": { |
| 604 | "common": [ |
| 605 | "src/gpu/cl/kernels/ClMulKernel.cpp", |
| 606 | "src/gpu/cl/operators/ClMul.cpp", |
| 607 | "src/runtime/CL/functions/CLPixelWiseMultiplication.cpp" |
| 608 | ] |
| 609 | } |
| 610 | }, |
| 611 | "Normalize": { |
| 612 | "files": { |
| 613 | "common": [ |
| 614 | "src/core/CL/kernels/CLNormalizationLayerKernel.cpp", |
| 615 | "src/runtime/CL/functions/CLNormalizationLayer.cpp" |
| 616 | ] |
| 617 | } |
| 618 | }, |
| 619 | "Pad": { |
| 620 | "deps": [ "Copy" ], |
| 621 | "files": { |
| 622 | "common": [ |
| 623 | "src/core/CL/kernels/CLPadLayerKernel.cpp", |
| 624 | "src/runtime/CL/functions/CLPadLayer.cpp" |
| 625 | ] |
| 626 | } |
| 627 | }, |
| 628 | "Permute": { |
| 629 | "files": { |
| 630 | "common": [ |
| 631 | "src/gpu/cl/kernels/ClPermuteKernel.cpp", |
| 632 | "src/gpu/cl/operators/ClPermute.cpp", |
| 633 | "src/runtime/CL/functions/CLPermute.cpp" |
| 634 | ] |
| 635 | } |
| 636 | }, |
| 637 | "Pool2d": { |
| 638 | "files": { |
| 639 | "common": [ |
| 640 | "src/gpu/cl/kernels/ClPool2dKernel.cpp", |
| 641 | "src/gpu/cl/operators/ClPool2d.cpp", |
| 642 | "src/runtime/CL/functions/CLPoolingLayer.cpp" |
| 643 | ] |
| 644 | } |
| 645 | }, |
ramelg01 | 3751569 | 2022-02-26 22:06:20 +0000 | [diff] [blame] | 646 | "Pool3d": { |
| 647 | "files": { |
| 648 | "common": [ |
| 649 | "src/gpu/cl/kernels/ClPool3dKernel.cpp", |
| 650 | "src/gpu/cl/operators/ClPool3d.cpp", |
| 651 | "src/runtime/CL/functions/CLPooling3dLayer.cpp" |
| 652 | ] |
| 653 | } |
| 654 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 655 | "PRelu": { |
| 656 | "deps": [ "ElementwiseBinary" ], |
| 657 | "files": { |
| 658 | "common": [ |
| 659 | "src/gpu/cl/operators/ClPRelu.cpp", |
| 660 | "src/runtime/CL/functions/CLPReluLayer.cpp" |
| 661 | ] |
| 662 | } |
| 663 | }, |
| 664 | "PriorBox": { |
| 665 | "files": { |
| 666 | "common": [ |
| 667 | "src/core/CL/kernels/CLPriorBoxLayerKernel.cpp", |
| 668 | "src/runtime/CL/functions/CLPriorBoxLayer.cpp" |
| 669 | ] |
| 670 | } |
| 671 | }, |
| 672 | "Quantize": { |
| 673 | "files": { |
| 674 | "common": [ |
| 675 | "src/gpu/cl/kernels/ClQuantizeKernel.cpp", |
| 676 | "src/gpu/cl/operators/ClQuantize.cpp", |
| 677 | "src/runtime/CL/functions/CLQuantizationLayer.cpp" |
| 678 | ] |
| 679 | } |
| 680 | }, |
| 681 | "Range": { |
| 682 | "files": { |
| 683 | "common": [ |
| 684 | "src/core/CL/kernels/CLRangeKernel.cpp", |
| 685 | "src/runtime/CL/functions/CLRange.cpp" |
| 686 | ] |
| 687 | } |
| 688 | }, |
| 689 | "Reduction": { |
| 690 | "deps": [ "Reshape" ], |
| 691 | "files": { |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 692 | "common": [ |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 693 | "src/core/CL/kernels/CLReductionOperationKernel.cpp", |
| 694 | "src/runtime/CL/functions/CLReductionOperation.cpp" |
| 695 | ] |
| 696 | } |
| 697 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 698 | "Reorg": { |
| 699 | "files": { |
| 700 | "common": [ |
| 701 | "src/core/CL/kernels/CLReorgLayerKernel.cpp", |
| 702 | "src/runtime/CL/functions/CLReorgLayer.cpp" |
| 703 | ] |
| 704 | } |
| 705 | }, |
| 706 | "Reshape": { |
| 707 | "files": { |
| 708 | "common": [ |
| 709 | "src/gpu/cl/kernels/ClReshapeKernel.cpp", |
| 710 | "src/gpu/cl/operators/ClReshape.cpp", |
| 711 | "src/runtime/CL/functions/CLReshapeLayer.cpp" |
| 712 | ] |
| 713 | } |
| 714 | }, |
| 715 | "Reverse": { |
| 716 | "files": { |
| 717 | "common": [ |
| 718 | "src/core/CL/kernels/CLReverseKernel.cpp", |
| 719 | "src/runtime/CL/functions/CLReverse.cpp" |
| 720 | ] |
| 721 | } |
| 722 | }, |
| 723 | "RNN": { |
| 724 | "deps": [ "Activation", "Cast", "ElementwiseBinary", "FullyConnected", "Gemm"], |
| 725 | "files": { |
| 726 | "common": [ "src/runtime/CL/functions/CLRNNLayer.cpp" ] |
| 727 | } |
| 728 | }, |
| 729 | "ROIAlign": { |
| 730 | "files": { |
| 731 | "common": [ |
| 732 | "src/core/CL/kernels/CLROIAlignLayerKernel.cpp", |
| 733 | "src/runtime/CL/functions/CLROIAlignLayer.cpp" |
| 734 | ] |
| 735 | } |
| 736 | }, |
| 737 | "ROIPool2d": { |
| 738 | "files": { |
| 739 | "common": [ |
| 740 | "src/core/CL/kernels/CLROIPoolingLayerKernel.cpp", |
| 741 | "src/runtime/CL/functions/CLROIPoolingLayer.cpp" |
| 742 | ] |
| 743 | } |
| 744 | }, |
| 745 | "Scale": { |
| 746 | "files": { |
| 747 | "common": [ |
| 748 | "src/gpu/cl/kernels/ClScaleKernel.cpp", |
| 749 | "src/gpu/cl/operators/ClScale.cpp", |
| 750 | "src/runtime/CL/functions/CLScale.cpp" |
| 751 | ] |
| 752 | } |
| 753 | }, |
| 754 | "Select": { |
| 755 | "files": { |
| 756 | "common": [ |
| 757 | "src/core/CL/kernels/CLSelectKernel.cpp", |
| 758 | "src/runtime/CL/functions/CLSelect.cpp" |
| 759 | ] |
| 760 | } |
| 761 | }, |
| 762 | "Slice": { |
| 763 | "deps": [ "StridedSlice" ], |
| 764 | "files": { |
| 765 | "common": [ "src/runtime/CL/functions/CLSlice.cpp" ] |
| 766 | } |
| 767 | }, |
| 768 | "Softmax": { |
| 769 | "deps": [ "Permute" ], |
| 770 | "files": { |
| 771 | "common": [ |
| 772 | "src/gpu/cl/kernels/ClSoftmaxKernel.cpp", |
| 773 | "src/gpu/cl/operators/ClSoftmax.cpp", |
| 774 | "src/runtime/CL/functions/CLSoftmaxLayer.cpp" |
| 775 | ] |
| 776 | } |
| 777 | }, |
| 778 | "SpaceToBatch": { |
| 779 | "files": { |
| 780 | "common": [ |
| 781 | "src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp", |
| 782 | "src/runtime/CL/functions/CLSpaceToBatchLayer.cpp" |
| 783 | ] |
| 784 | } |
| 785 | }, |
| 786 | "SpaceToDepth": { |
| 787 | "files": { |
| 788 | "common": [ |
| 789 | "src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp", |
| 790 | "src/runtime/CL/functions/CLSpaceToDepthLayer.cpp" |
| 791 | ] |
| 792 | } |
| 793 | }, |
| 794 | "Split": { |
| 795 | "deps": [ "StridedSlice" ], |
| 796 | "files": { |
| 797 | "common": [ "src/runtime/CL/functions/CLSplit.cpp" ] |
| 798 | } |
| 799 | }, |
| 800 | "Stack": { |
| 801 | "files": { |
| 802 | "common": [ |
| 803 | "src/core/CL/kernels/CLStackLayerKernel.cpp", |
| 804 | "src/runtime/CL/functions/CLStackLayer.cpp" |
| 805 | ] |
| 806 | } |
| 807 | }, |
| 808 | "StridedSlice": { |
| 809 | "files": { |
| 810 | "common": [ |
| 811 | "src/core/CL/kernels/CLStridedSliceKernel.cpp", |
| 812 | "src/runtime/CL/functions/CLStridedSlice.cpp" |
| 813 | ] |
| 814 | } |
| 815 | }, |
| 816 | "Sub": { |
| 817 | "files": { |
| 818 | "common": [ |
| 819 | "src/gpu/cl/kernels/ClElementwiseKernel.cpp", |
| 820 | "src/gpu/cl/operators/ClSub.cpp" |
| 821 | ] |
| 822 | } |
| 823 | }, |
| 824 | "Tile": { |
| 825 | "files": { |
| 826 | "common": [ |
| 827 | "src/core/CL/kernels/CLTileKernel.cpp", |
| 828 | "src/runtime/CL/functions/CLTile.cpp" |
| 829 | ] |
| 830 | } |
| 831 | }, |
| 832 | "Transpose": { |
| 833 | "files": { |
| 834 | "common": [ |
| 835 | "src/gpu/cl/kernels/ClTransposeKernel.cpp", |
| 836 | "src/gpu/cl/operators/ClTranspose.cpp", |
| 837 | "src/runtime/CL/functions/CLTranspose.cpp" |
| 838 | ] |
| 839 | } |
| 840 | }, |
| 841 | "Unstack": { |
| 842 | "deps": [ "StridedSlice" ], |
| 843 | "files": { |
| 844 | "common": [ "src/runtime/CL/functions/CLUnstack.cpp" ] |
| 845 | } |
| 846 | }, |
| 847 | "YUVNormalize": { |
| 848 | "files": { |
| 849 | "common": [ |
| 850 | "src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp", |
| 851 | "src/runtime/CL/functions/CLNormalizePlanarYUVLayer.cpp" |
| 852 | ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 853 | } |
Georgios Pinitas | bdcdc39 | 2021-04-22 16:42:03 +0100 | [diff] [blame] | 854 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 855 | } |
| 856 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 857 | "cpu": { |
Georgios Pinitas | 13ef176 | 2021-07-14 17:14:43 +0100 | [diff] [blame] | 858 | "common": [ |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 859 | "src/cpu/CpuContext.cpp", |
| 860 | "src/cpu/CpuQueue.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 861 | "src/cpu/CpuTensor.cpp", |
| 862 | "src/core/NEON/kernels/NEFillBorderKernel.cpp", |
| 863 | "src/runtime/NEON/INEOperator.cpp", |
| 864 | "src/runtime/NEON/INESimpleFunction.cpp", |
| 865 | "src/runtime/NEON/INESimpleFunctionNoBorder.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 866 | ], |
| 867 | "operators": { |
| 868 | "Activation": { |
Viet-Hoa Do | fd472f0 | 2023-03-15 14:05:06 +0000 | [diff] [blame] | 869 | "deps": [ "LUT" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 870 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 871 | "common": [ |
| 872 | "src/cpu/operators/CpuActivation.cpp", |
| 873 | "src/cpu/kernels/CpuActivationKernel.cpp", |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 874 | "src/runtime/NEON/functions/NEActivationLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 875 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 876 | "neon": { |
Dana Zlotnik | 3229171 | 2021-11-25 09:58:27 +0200 | [diff] [blame] | 877 | "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ], |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 878 | "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ], |
Viet-Hoa Do | 29db3d2 | 2022-08-10 11:56:49 +0100 | [diff] [blame] | 879 | "qasymm8": [ |
| 880 | "src/cpu/kernels/activation/generic/neon/qasymm8.cpp", |
| 881 | "src/cpu/kernels/activation/generic/neon/lut.cpp" |
| 882 | ], |
| 883 | "qasymm8_signed": [ |
| 884 | "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp", |
| 885 | "src/cpu/kernels/activation/generic/neon/lut.cpp" |
| 886 | ], |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 887 | "qsymm16": [ "src/cpu/kernels/activation/generic/neon/qsymm16.cpp" ] |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 888 | }, |
| 889 | "sve": { |
Dana Zlotnik | 3229171 | 2021-11-25 09:58:27 +0200 | [diff] [blame] | 890 | "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ], |
Pablo Marquez Tello | d158609 | 2022-06-10 14:37:10 +0100 | [diff] [blame] | 891 | "fp32": [ "src/cpu/kernels/activation/generic/sve/fp32.cpp" ], |
| 892 | "qasymm8": ["src/cpu/kernels/activation/generic/sve/lut.cpp"] |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 893 | }, |
alerah01 | c9e519d | 2022-01-31 19:04:10 +0200 | [diff] [blame] | 894 | "sve2":{ |
Dana Zlotnik | 3229171 | 2021-11-25 09:58:27 +0200 | [diff] [blame] | 895 | "qasymm8": [ "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp" ], |
| 896 | "qasymm8_signed": [ "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp" ], |
| 897 | "qsymm16": [ "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 898 | } |
| 899 | } |
| 900 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 901 | "ArgMinMax": { |
| 902 | "deps": [ "Reduction" ], |
| 903 | "files": { |
| 904 | "common": [ "src/runtime/NEON/functions/NEArgMinMaxLayer.cpp" ] |
| 905 | } |
| 906 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 907 | "Add": { |
| 908 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 909 | "common": [ |
| 910 | "src/cpu/operators/CpuAdd.cpp", |
| 911 | "src/cpu/kernels/CpuAddKernel.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 912 | "src/runtime/NEON/functions/NEArithmeticAddition.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 913 | ], |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 914 | "neon": { |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 915 | "common": ["src/cpu/kernels/add/generic/neon/impl.cpp"], |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 916 | "fp32":["src/cpu/kernels/add/generic/neon/fp32.cpp"], |
| 917 | "fp16":["src/cpu/kernels/add/generic/neon/fp16.cpp"], |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 918 | "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"], |
| 919 | "qasymm8": ["src/cpu/kernels/add/generic/neon/qasymm8.cpp"], |
| 920 | "qasymm8_signed": ["src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp"], |
| 921 | "qsymm16": ["src/cpu/kernels/add/generic/neon/qsymm16.cpp"] |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 922 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 923 | "sve": { |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 924 | "common": [ "src/cpu/kernels/add/generic/sve/impl.cpp" ], |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 925 | "integer":["src/cpu/kernels/add/generic/sve/integer.cpp"], |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 926 | "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"], |
| 927 | "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"] |
| 928 | }, |
| 929 | "sve2": { |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 930 | "qasymm8": [ "src/cpu/kernels/add/generic/sve2/qasymm8.cpp" ], |
| 931 | "qasymm8_signed": [ "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp" ], |
| 932 | "qsymm16": [ "src/cpu/kernels/add/generic/sve2/qsymm16.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 933 | } |
| 934 | } |
| 935 | }, |
Gunes Bayir | ae72a46 | 2023-01-29 13:24:24 +0000 | [diff] [blame] | 936 | "AddMulAdd": { |
| 937 | "files": { |
| 938 | "common": [ |
| 939 | "src/cpu/operators/CpuAddMulAdd.cpp", |
| 940 | "src/cpu/kernels/CpuAddMulAddKernel.cpp", |
| 941 | "src/runtime/NEON/functions/NEAddMulAdd.cpp" |
| 942 | ], |
| 943 | "neon": { |
| 944 | "fp32":["src/cpu/kernels/addmuladd/generic/neon/fp32.cpp"], |
| 945 | "fp16":["src/cpu/kernels/addmuladd/generic/neon/fp16.cpp"], |
| 946 | "qasymm8": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp"], |
| 947 | "qasymm8_signed": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8_signed.cpp"] |
| 948 | } |
| 949 | } |
| 950 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 951 | "BatchNormalize": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 952 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 953 | "common": [ |
| 954 | "src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp", |
| 955 | "src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp", |
| 956 | "src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp", |
| 957 | "src/runtime/NEON/functions/NEFuseBatchNormalization.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 958 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 959 | "neon": { |
Yair Schwarzbaum | 41a729e | 2021-11-15 20:42:47 +0200 | [diff] [blame] | 960 | "common": [ |
| 961 | "src/cpu/kernels/fuse_batch_normalization/generic/impl.cpp", |
| 962 | "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp", |
| 963 | "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp" |
| 964 | ], |
| 965 | "fp16": [ |
| 966 | "src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp", |
| 967 | "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp", |
| 968 | "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp" |
| 969 | ], |
| 970 | "fp32": [ |
| 971 | "src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp", |
| 972 | "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp", |
| 973 | "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp" |
| 974 | ] |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 975 | }, |
| 976 | "sve": { |
| 977 | "fp16": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp" ], |
| 978 | "fp32": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 979 | } |
| 980 | } |
| 981 | }, |
| 982 | "BatchToSpace": { |
| 983 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 984 | "common": [ |
| 985 | "src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp", |
| 986 | "src/runtime/NEON/functions/NEBatchToSpaceLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 987 | ] |
| 988 | } |
| 989 | }, |
| 990 | "BitwiseAnd": { |
| 991 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 992 | "common": [ |
| 993 | "src/core/NEON/kernels/NEBitwiseAndKernel.cpp", |
| 994 | "src/runtime/NEON/functions/NEBitwiseAnd.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 995 | ] |
| 996 | } |
| 997 | }, |
| 998 | "BitwiseNot": { |
| 999 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1000 | "common": [ |
| 1001 | "src/core/NEON/kernels/NEBitwiseNotKernel.cpp", |
| 1002 | "src/runtime/NEON/functions/NEBitwiseNot.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1003 | ] |
| 1004 | } |
| 1005 | }, |
| 1006 | "BitwiseOr": { |
| 1007 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1008 | "common": [ |
| 1009 | "src/core/NEON/kernels/NEBitwiseOrKernel.cpp", |
| 1010 | "src/runtime/NEON/functions/NEBitwiseOr.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1011 | ] |
| 1012 | } |
| 1013 | }, |
| 1014 | "BitwiseXor": { |
| 1015 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1016 | "common": [ |
| 1017 | "src/core/NEON/kernels/NEBitwiseXorKernel.cpp", |
| 1018 | "src/runtime/NEON/functions/NEBitwiseXor.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1019 | ] |
| 1020 | } |
| 1021 | }, |
| 1022 | "BoundingBoxTransform": { |
| 1023 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1024 | "common": [ |
| 1025 | "src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp", |
| 1026 | "src/runtime/NEON/functions/NEBoundingBoxTransform.cpp" |
Dana Zlotnik | 3475ffe | 2022-01-03 14:37:10 +0200 | [diff] [blame] | 1027 | ], |
| 1028 | "neon":{ |
| 1029 | "common":["src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp"], |
| 1030 | "fp32":["src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp"], |
| 1031 | "fp16":["src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp"], |
| 1032 | "qsymm16":["src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp"] |
Dana Zlotnik | 3475ffe | 2022-01-03 14:37:10 +0200 | [diff] [blame] | 1033 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1034 | } |
| 1035 | }, |
| 1036 | "Cast": { |
| 1037 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1038 | "common": [ |
| 1039 | "src/cpu/operators/CpuCast.cpp", |
| 1040 | "src/cpu/kernels/CpuCastKernel.cpp", |
Yair Schwarzbaum | 298b2c0 | 2022-02-01 08:55:56 +0200 | [diff] [blame] | 1041 | "src/runtime/NEON/functions/NECast.cpp", |
| 1042 | "src/cpu/kernels/cast/generic/neon/bfloat16.cpp" |
| 1043 | ], |
| 1044 | "neon":{ |
| 1045 | "fp16":["src/cpu/kernels/cast/generic/neon/fp16.cpp"] |
| 1046 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1047 | } |
| 1048 | }, |
| 1049 | "ChannelShuffle": { |
| 1050 | "files": { |
| 1051 | "common": [ |
| 1052 | "src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp", |
| 1053 | "src/runtime/NEON/functions/NEChannelShuffleLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1054 | ] |
| 1055 | } |
| 1056 | }, |
| 1057 | "Concatenate": { |
| 1058 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1059 | "common": [ |
| 1060 | "src/cpu/operators/CpuConcatenate.cpp", |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 1061 | "src/cpu/kernels/CpuConcatenateWidthKernel.cpp", |
| 1062 | "src/cpu/kernels/CpuConcatenateBatchKernel.cpp", |
| 1063 | "src/cpu/kernels/CpuConcatenateDepthKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1064 | "src/cpu/kernels/CpuConcatenateHeightKernel.cpp", |
| 1065 | "src/runtime/NEON/functions/NEConcatenateLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1066 | ] |
| 1067 | } |
| 1068 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1069 | "Conv2d": { |
| 1070 | "deps": [ |
| 1071 | "Activation", |
| 1072 | "ElementwiseBinary", |
| 1073 | "FFT2D", |
| 1074 | "Gemm", |
| 1075 | "Mul", |
| 1076 | "Pad", |
| 1077 | "Permute", |
| 1078 | "Reshape", |
| 1079 | "Reverse", |
| 1080 | "Slice" |
| 1081 | ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1082 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1083 | "common": [ |
| 1084 | "src/cpu/operators/CpuConv2d.cpp", |
| 1085 | "src/cpu/operators/CpuDirectConv2d.cpp", |
| 1086 | "src/cpu/operators/CpuGemmDirectConv2d.cpp", |
| 1087 | "src/cpu/operators/CpuGemmConv2d.cpp", |
| 1088 | "src/cpu/operators/CpuWinogradConv2d.cpp", |
| 1089 | "src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp", |
| 1090 | "src/cpu/kernels/CpuDirectConv2dKernel.cpp", |
| 1091 | "src/cpu/kernels/CpuDirectConv2dOutputStageKernel.cpp", |
| 1092 | "src/cpu/kernels/CpuWinogradConv2dKernel.cpp", |
| 1093 | "src/cpu/kernels/CpuCol2ImKernel.cpp", |
| 1094 | "src/cpu/kernels/CpuIm2ColKernel.cpp", |
| 1095 | "src/cpu/kernels/CpuWeightsReshapeKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1096 | "src/runtime/NEON/functions/NEConvolutionLayer.cpp", |
| 1097 | "src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp", |
| 1098 | "src/runtime/NEON/functions/NEFFTConvolutionLayer.cpp", |
| 1099 | "src/runtime/NEON/functions/NEGEMMConv2d.cpp", |
| 1100 | "src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp", |
| 1101 | "src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp" |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1102 | ], |
| 1103 | "neon": { |
| 1104 | "common": [ |
| 1105 | "src/core/NEON/kernels/convolution/common/padding.cpp", |
| 1106 | "src/core/NEON/kernels/convolution/common/qasymm8.cpp", |
| 1107 | "src/core/NEON/kernels/convolution/common/qsymm8.cpp", |
| 1108 | "src/core/NEON/kernels/convolution/common/utils.cpp", |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1109 | "src/core/NEON/kernels/convolution/winograd/input_transforms_fp16.cpp", |
| 1110 | "src/core/NEON/kernels/convolution/winograd/input_transforms_fp32.cpp", |
| 1111 | "src/core/NEON/kernels/convolution/winograd/output_transforms_fp16.cpp", |
| 1112 | "src/core/NEON/kernels/convolution/winograd/output_transforms_fp32.cpp", |
| 1113 | "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp16.cpp", |
| 1114 | "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp32.cpp", |
| 1115 | "src/core/NEON/kernels/convolution/winograd/winograd_fp16.cpp", |
| 1116 | "src/core/NEON/kernels/convolution/winograd/winograd_fp32.cpp", |
| 1117 | "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp", |
| 1118 | "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp", |
| 1119 | "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp", |
| 1120 | "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp", |
| 1121 | "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp", |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1122 | "src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp", |
| 1123 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp", |
| 1124 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp", |
| 1125 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp", |
| 1126 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp", |
| 1127 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp", |
| 1128 | "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp", |
| 1129 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/a64_fp16_4x4_3x3.cpp", |
| 1130 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_3x3.cpp", |
| 1131 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_5x5.cpp", |
| 1132 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_4x4_3x3.cpp", |
| 1133 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x2_1x7.cpp", |
| 1134 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp", |
| 1135 | "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x6_1x3.cpp", |
alerah01 | c9e519d | 2022-01-31 19:04:10 +0200 | [diff] [blame] | 1136 | "src/cpu/kernels/directconv2d/nhwc/neon/impl.cpp", |
| 1137 | "src/cpu/kernels/directconv2d/nchw/all.cpp" |
| 1138 | ], |
| 1139 | "fp32": [ |
| 1140 | "src/cpu/kernels/directconv2d/nhwc/neon/fp32.cpp" |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1141 | ] |
ramelg01 | 99c4647 | 2022-07-15 13:08:39 +0100 | [diff] [blame] | 1142 | }, |
| 1143 | "sve": { |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1144 | "common": [ |
| 1145 | "src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp", |
| 1146 | "src/core/NEON/kernels/convolution/winograd/input_transforms/sme_fp32_mla_6x6.cpp", |
| 1147 | "src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp" |
| 1148 | ] |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1149 | } |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 1150 | } |
| 1151 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1152 | "Copy": { |
| 1153 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1154 | "common": [ |
| 1155 | "src/cpu/operators/CpuCopy.cpp", |
| 1156 | "src/cpu/kernels/CpuCopyKernel.cpp", |
| 1157 | "src/runtime/NEON/functions/NECopy.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1158 | ] |
| 1159 | } |
| 1160 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1161 | "CropResize": { |
| 1162 | "deps": [ "Scale" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1163 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1164 | "common": [ |
| 1165 | "src/core/NEON/kernels/NECropKernel.cpp", |
| 1166 | "src/runtime/NEON/functions/NECropResize.cpp" |
alerah01 | 2761993 | 2021-11-24 17:21:43 +0200 | [diff] [blame] | 1167 | ], |
| 1168 | "neon": { |
| 1169 | "common": [ "src/cpu/kernels/crop/generic/neon/impl.cpp" ], |
| 1170 | "fp32": [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ], |
| 1171 | "fp16": [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ], |
| 1172 | "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp" ] |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1173 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1174 | } |
| 1175 | }, |
| 1176 | "Deconv2d": { |
| 1177 | "deps": [ "Conv2d", "Reverse", "Transpose"], |
| 1178 | "files": { |
| 1179 | "common": [ |
| 1180 | "src/runtime/NEON/functions/NEDeconvolutionLayer.cpp" |
| 1181 | ] |
| 1182 | } |
| 1183 | }, |
| 1184 | "DepthConvert": { |
| 1185 | "deps": [ "Cast"], |
| 1186 | "files": { |
| 1187 | "common": [ |
| 1188 | "src/runtime/NEON/functions/NEDepthConvertLayer.cpp" |
| 1189 | ] |
| 1190 | } |
| 1191 | }, |
| 1192 | "DepthToSpace": { |
| 1193 | "files": { |
| 1194 | "common": [ |
| 1195 | "src/core/NEON/kernels/NEDepthToSpaceLayerKernel.cpp", |
| 1196 | "src/runtime/NEON/functions/NEDepthToSpaceLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1197 | ] |
| 1198 | } |
| 1199 | }, |
| 1200 | "DepthwiseConv2d": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1201 | "deps": [ "Activation", "Permute" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1202 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1203 | "common": [ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 1204 | "src/cpu/operators/CpuDepthwiseConv2d.cpp", |
| 1205 | "src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1206 | "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.cpp", |
| 1207 | "src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1208 | "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1209 | ], |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1210 | "neon": { |
| 1211 | "common": [ |
| 1212 | "src/core/NEON/kernels/convolution/common/padding.cpp", |
| 1213 | "src/core/NEON/kernels/convolution/common/qasymm8.cpp", |
| 1214 | "src/core/NEON/kernels/convolution/common/qsymm8.cpp", |
| 1215 | "src/core/NEON/kernels/convolution/common/utils.cpp", |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 1216 | "src/core/NEON/kernels/arm_conv/addressing.cpp", |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 1217 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_common.cpp", |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 1218 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_strategies_common.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1219 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp", |
| 1220 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp", |
| 1221 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp", |
| 1222 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp", |
| 1223 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp", |
| 1224 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
| 1225 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp", |
| 1226 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp", |
| 1227 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
| 1228 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", |
| 1229 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", |
| 1230 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
| 1231 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", |
| 1232 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", |
| 1233 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
| 1234 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 1235 | "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1236 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1237 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1238 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", |
| 1239 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", |
| 1240 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", |
| 1241 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", |
| 1242 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1243 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1244 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1245 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1246 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
| 1247 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1248 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1249 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", |
| 1250 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", |
| 1251 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", |
| 1252 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", |
| 1253 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1254 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1255 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1256 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1257 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
| 1258 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1259 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1260 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1261 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1262 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
| 1263 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1264 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1265 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1266 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1267 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1268 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
| 1269 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1270 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1271 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1272 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 1273 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1274 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1275 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1276 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp", |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 1277 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp", |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 1278 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp", |
| 1279 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp", |
Dana Zlotnik | ebbae94 | 2022-02-03 12:52:15 +0200 | [diff] [blame] | 1280 | "src/cpu/kernels/depthwiseconv2d/generic/neon/impl.cpp" |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1281 | ], |
Dana Zlotnik | ebbae94 | 2022-02-03 12:52:15 +0200 | [diff] [blame] | 1282 | "fp16":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp16.cpp"], |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1283 | "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"], |
| 1284 | "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"], |
Dana Zlotnik | ebbae94 | 2022-02-03 12:52:15 +0200 | [diff] [blame] | 1285 | "qasymm8_signed":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8_signed.cpp"] |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1286 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1287 | "sve": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1288 | "common": [ |
Freddie Liardet | d216f57 | 2021-08-03 15:57:32 +0100 | [diff] [blame] | 1289 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_8b_mla.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1290 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp", |
| 1291 | "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp", |
| 1292 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1293 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1294 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", |
| 1295 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", |
| 1296 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", |
| 1297 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", |
| 1298 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1299 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1300 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1301 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1302 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1303 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1304 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", |
| 1305 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", |
| 1306 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", |
| 1307 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", |
| 1308 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1309 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1310 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1311 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1312 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp", |
| 1313 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp", |
| 1314 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp", |
| 1315 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", |
| 1316 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1317 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1318 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1319 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1320 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", |
| 1321 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", |
| 1322 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1323 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", |
| 1324 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1325 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
| 1326 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1327 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", |
| 1328 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", |
| 1329 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1330 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1331 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", |
| 1332 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1333 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1334 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", |
| 1335 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", |
| 1336 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", |
| 1337 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", |
| 1338 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", |
| 1339 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", |
| 1340 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s1_4rows_mla_za/generic.cpp", |
| 1341 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp", |
| 1342 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp", |
| 1343 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s2_4rows_mla_za/generic.cpp", |
| 1344 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp", |
| 1345 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s2_4rows_dot_za/generic.cpp", |
| 1346 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp", |
| 1347 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp", |
| 1348 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s1_4rows_dot_za/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1349 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp", |
| 1350 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s1_4rows_dot_za/generic.cpp", |
| 1351 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s2_4rows_dot_za/generic.cpp", |
| 1352 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s1_4rows_dot_za/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1353 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_4rows_dot_za/generic.cpp", |
| 1354 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s1_4rows_dot_za/generic.cpp", |
| 1355 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s2_4rows_dot_za/generic.cpp", |
| 1356 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1357 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_4rows_dot_za/generic.cpp", |
| 1358 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s1_4rows_dot_za/generic.cpp", |
| 1359 | "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1360 | ] |
| 1361 | } |
| 1362 | } |
| 1363 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1364 | "Dequantize": { |
| 1365 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1366 | "common": [ |
| 1367 | "src/cpu/operators/CpuDequantize.cpp", |
| 1368 | "src/cpu/kernels/CpuDequantizeKernel.cpp", |
| 1369 | "src/runtime/NEON/functions/NEDequantizationLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1370 | ] |
| 1371 | } |
| 1372 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1373 | "DetectionPostProcess": { |
| 1374 | "deps": [ "Dequantize" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1375 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1376 | "common" : [ "src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1377 | } |
| 1378 | }, |
Sheri Zhang | 6d9c982 | 2021-09-24 16:02:57 +0100 | [diff] [blame] | 1379 | "Conv3d": { |
| 1380 | "deps": [ |
| 1381 | "Activation" |
| 1382 | ], |
| 1383 | "files": { |
| 1384 | "common": [ |
| 1385 | "src/cpu/operators/CpuDirectConv3d.cpp", |
| 1386 | "src/cpu/kernels/CpuDirectConv3dKernel.cpp", |
| 1387 | "src/runtime/NEON/functions/NEConv3D.cpp" |
| 1388 | ] |
| 1389 | } |
| 1390 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1391 | "ElementwiseBinary": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1392 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1393 | "common": [ |
| 1394 | "src/cpu/operators/CpuElementwise.cpp", |
| 1395 | "src/cpu/kernels/CpuElementwiseKernel.cpp", |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 1396 | "src/runtime/NEON/functions/NEElementwiseOperations.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1397 | ], |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1398 | "neon":{ |
| 1399 | "fp32": ["src/cpu/kernels/elementwise_binary/generic/neon/fp32.cpp"], |
| 1400 | "fp16": ["src/cpu/kernels/elementwise_binary/generic/neon/fp16.cpp"], |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 1401 | "integer": ["src/cpu/kernels/elementwise_binary/generic/neon/integer.cpp"], |
| 1402 | "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8.cpp"], |
| 1403 | "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp"] |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1404 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1405 | "sve": { |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1406 | "common": ["src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp" ], |
| 1407 | "integer": ["src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp"], |
| 1408 | "fp32": ["src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp"], |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 1409 | "fp16": ["src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp"] |
| 1410 | |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1411 | }, |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 1412 | "sve2":{ |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1413 | "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8.cpp"], |
| 1414 | "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8_signed.cpp"] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1415 | } |
| 1416 | } |
| 1417 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1418 | "ElementwiseUnary":{ |
Viet-Hoa Do | fd472f0 | 2023-03-15 14:05:06 +0000 | [diff] [blame] | 1419 | "deps": [ "LUT" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1420 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1421 | "common": [ |
| 1422 | "src/cpu/operators/CpuElementwiseUnary.cpp", |
| 1423 | "src/cpu/kernels/CpuElementwiseUnaryKernel.cpp", |
| 1424 | "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1425 | ], |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1426 | "neon": { |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 1427 | "common":["src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp"], |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1428 | "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"], |
| 1429 | "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"], |
Viet-Hoa Do | fd472f0 | 2023-03-15 14:05:06 +0000 | [diff] [blame] | 1430 | "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"], |
Ramy Elgammal | 8b7f42a | 2023-03-31 16:16:15 +0100 | [diff] [blame] | 1431 | "qasymm8": [ |
| 1432 | "src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp", |
| 1433 | "src/cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp" |
| 1434 | ], |
| 1435 | "qasymm8_signed": [ |
| 1436 | "src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp", |
| 1437 | "src/cpu/kernels/elementwise_unary/generic/neon/qasymm8_signed.cpp" |
| 1438 | ] |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1439 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1440 | "sve": { |
Dana Zlotnik | d5c496d | 2021-11-28 14:46:12 +0200 | [diff] [blame] | 1441 | "common": ["src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp" ], |
| 1442 | "integer": ["src/cpu/kernels/elementwise_unary/generic/sve/integer.cpp"], |
| 1443 | "fp32": ["src/cpu/kernels/elementwise_unary/generic/sve/fp32.cpp"], |
Viet-Hoa Do | fd472f0 | 2023-03-15 14:05:06 +0000 | [diff] [blame] | 1444 | "fp16": ["src/cpu/kernels/elementwise_unary/generic/sve/fp16.cpp"], |
| 1445 | "qasymm8": ["src/cpu/kernels/elementwise_unary/generic/sve/q8.cpp"], |
| 1446 | "qasymm8_signed": ["src/cpu/kernels/elementwise_unary/generic/sve/q8.cpp"] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1447 | } |
| 1448 | } |
| 1449 | }, |
| 1450 | "FFT1D": { |
Freddie Liardet | 487d390 | 2021-09-21 12:36:43 +0100 | [diff] [blame] | 1451 | "deps": [ "Reduction" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1452 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1453 | "common": [ |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1454 | "src/core/NEON/kernels/NEFFTDigitReverseKernel.cpp", |
| 1455 | "src/core/NEON/kernels/NEFFTRadixStageKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1456 | "src/core/NEON/kernels/NEFFTScaleKernel.cpp", |
| 1457 | "src/runtime/NEON/functions/NEFFT1D.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1458 | ] |
| 1459 | } |
| 1460 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1461 | "FFT2D": { |
| 1462 | "deps": [ "FFT1D" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1463 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1464 | "common": [ |
| 1465 | "src/runtime/NEON/functions/NEFFT2D.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1466 | ] |
| 1467 | } |
| 1468 | }, |
| 1469 | "Fill": { |
| 1470 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1471 | "common": [ |
| 1472 | "src/cpu/operators/CpuFill.cpp", |
| 1473 | "src/cpu/kernels/CpuFillKernel.cpp", |
| 1474 | "src/runtime/NEON/functions/NEFill.cpp" |
| 1475 | ] |
| 1476 | } |
| 1477 | }, |
| 1478 | "Flatten": { |
| 1479 | "deps: ": [ "Reshape" ], |
| 1480 | "files": { |
| 1481 | "common": [ |
| 1482 | "src/cpu/operators/CpuFlatten.cpp", |
| 1483 | "src/runtime/NEON/functions/NEFlattenLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1484 | ] |
| 1485 | } |
| 1486 | }, |
| 1487 | "Floor": { |
| 1488 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1489 | "common": [ |
| 1490 | "src/cpu/operators/CpuFloor.cpp", |
| 1491 | "src/cpu/kernels/CpuFloorKernel.cpp", |
| 1492 | "src/runtime/NEON/functions/NEFloor.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1493 | ], |
| 1494 | "neon": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1495 | "fp32": [ "src/cpu/kernels/floor/neon/fp32.cpp" ], |
| 1496 | "fp16": [ "src/cpu/kernels/floor/neon/fp16.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1497 | } |
| 1498 | } |
| 1499 | }, |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 1500 | "FullyConnected": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1501 | "deps": [ "Flatten", "Gemm", "Transpose"], |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 1502 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1503 | "common": [ |
| 1504 | "src/cpu/kernels/CpuConvertFullyConnectedWeightsKernel.cpp", |
| 1505 | "src/cpu/operators/CpuConvertFullyConnectedWeights.cpp", |
| 1506 | "src/cpu/operators/CpuFullyConnected.cpp", |
| 1507 | "src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp", |
| 1508 | "src/runtime/NEON/functions/NEFullyConnectedLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1509 | ] |
| 1510 | } |
| 1511 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1512 | "Gather": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1513 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1514 | "common": [ |
| 1515 | "src/core/NEON/kernels/NEGatherKernel.cpp", |
| 1516 | "src/runtime/NEON/functions/NEGather.cpp" |
| 1517 | ] |
| 1518 | } |
| 1519 | }, |
| 1520 | "Gemm": { |
Freddie Liardet | 487d390 | 2021-09-21 12:36:43 +0100 | [diff] [blame] | 1521 | "deps": [ "Quantize", "Add"], |
Francesco.Petrogalli@arm.com | 5fcf22d | 2022-04-05 10:31:08 +0000 | [diff] [blame] | 1522 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1523 | "common": [ |
| 1524 | "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp", |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 1525 | "src/cpu/kernels/CpuGemmMatrixAdditionKernel.cpp", |
| 1526 | "src/cpu/kernels/CpuGemmMatrixMultiplyKernel.cpp", |
| 1527 | "src/cpu/kernels/CpuGemmTranspose1xWKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1528 | "src/cpu/kernels/CpuGemmInterleave4x4Kernel.cpp", |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 1529 | "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp", |
| 1530 | "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp", |
| 1531 | "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp", |
| 1532 | "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp", |
| 1533 | "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp", |
| 1534 | "src/cpu/kernels/CpuGemmLowpMatrixReductionKernel.cpp", |
| 1535 | "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1536 | "src/cpu/kernels/CpuGemmLowpOffsetContributionKernel.cpp", |
| 1537 | "src/cpu/operators/CpuGemm.cpp", |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 1538 | "src/cpu/operators/CpuMatMul.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1539 | "src/cpu/operators/CpuGemmLowpOutputStage.cpp", |
| 1540 | "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1541 | "src/runtime/NEON/functions/NEGEMM.cpp", |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 1542 | "src/runtime/NEON/functions/NEMatMul.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1543 | "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1544 | "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1545 | ], |
| 1546 | "neon": { |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1547 | "common": [ |
| 1548 | "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp", |
| 1549 | "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp", |
| 1550 | "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp", |
| 1551 | "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp", |
| 1552 | "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp", |
| 1553 | "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp", |
| 1554 | "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp", |
| 1555 | "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp", |
| 1556 | "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp", |
| 1557 | "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp", |
| 1558 | "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp", |
| 1559 | "src/core/NEON/kernels/arm_gemm/mergeresults.cpp", |
| 1560 | "src/core/NEON/kernels/arm_gemm/misc.cpp", |
| 1561 | "src/core/NEON/kernels/arm_gemm/quantized.cpp", |
| 1562 | "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp", |
| 1563 | "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp", |
| 1564 | "src/core/NEON/kernels/arm_gemm/transform.cpp", |
| 1565 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp", |
| 1566 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp", |
| 1567 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp", |
| 1568 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp", |
| 1569 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp", |
| 1570 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp", |
| 1571 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp", |
| 1572 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp", |
| 1573 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp", |
| 1574 | "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp", |
| 1575 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp", |
| 1576 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp", |
| 1577 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp", |
| 1578 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp", |
| 1579 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp", |
| 1580 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp", |
| 1581 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp", |
| 1582 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp", |
| 1583 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp", |
| 1584 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp", |
| 1585 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp", |
| 1586 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp", |
| 1587 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp", |
| 1588 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp", |
| 1589 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp", |
| 1590 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp", |
| 1591 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp", |
| 1592 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp", |
| 1593 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp", |
| 1594 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp", |
| 1595 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp", |
| 1596 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp", |
| 1597 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp", |
| 1598 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp", |
| 1599 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp", |
| 1600 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp", |
| 1601 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp", |
| 1602 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp", |
| 1603 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp", |
| 1604 | "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp", |
| 1605 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp", |
| 1606 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp", |
| 1607 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp", |
| 1608 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp", |
| 1609 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp", |
| 1610 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp", |
| 1611 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp", |
| 1612 | "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp", |
| 1613 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp", |
| 1614 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp", |
| 1615 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp", |
| 1616 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp", |
| 1617 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp", |
| 1618 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp", |
| 1619 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp", |
| 1620 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp", |
| 1621 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp", |
| 1622 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp", |
| 1623 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp", |
| 1624 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp", |
| 1625 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp", |
| 1626 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp", |
| 1627 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp", |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1628 | "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp", |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 1629 | "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp", |
| 1630 | "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp" |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1631 | ], |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 1632 | "fp32":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp", |
| 1633 | "src/cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp"], |
| 1634 | "fp16":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp", |
| 1635 | "src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp"], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1636 | "estate32": [ |
| 1637 | "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp", |
| 1638 | "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp", |
| 1639 | "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp" |
| 1640 | ], |
| 1641 | "estate64": [ |
Freddie Liardet | 487d390 | 2021-09-21 12:36:43 +0100 | [diff] [blame] | 1642 | "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp" |
Francesco.Petrogalli@arm.com | 5fcf22d | 2022-04-05 10:31:08 +0000 | [diff] [blame] | 1643 | ], |
| 1644 | "experimental_fixed_format_kernels": [ |
| 1645 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp", |
| 1646 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp", |
| 1647 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp", |
| 1648 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp", |
| 1649 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp", |
| 1650 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp", |
| 1651 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp", |
| 1652 | "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1653 | ] |
| 1654 | }, |
| 1655 | "sve": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1656 | "common": [ |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1657 | "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp", |
| 1658 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp", |
| 1659 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp", |
| 1660 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp", |
| 1661 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp", |
| 1662 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp", |
| 1663 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp", |
| 1664 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp", |
| 1665 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp", |
David Mansell | aaa9da1 | 2023-03-10 13:48:50 +0000 | [diff] [blame] | 1666 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp", |
| 1667 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp", |
| 1668 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1669 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp", |
| 1670 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp", |
| 1671 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp", |
| 1672 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp", |
| 1673 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp", |
| 1674 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp", |
| 1675 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp", |
| 1676 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp", |
| 1677 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp", |
| 1678 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp", |
| 1679 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp", |
| 1680 | "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1681 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1682 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp", |
| 1683 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1684 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1685 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1686 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1687 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1688 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1689 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp", |
| 1690 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1691 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1692 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1693 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1694 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp", |
| 1695 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1696 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1697 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1698 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1699 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp", |
| 1700 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1701 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1702 | "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1703 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp", |
| 1704 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1705 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1706 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1707 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1708 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp", |
| 1709 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1710 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1711 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp", |
| 1712 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1713 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1714 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp", |
| 1715 | "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp", |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 1716 | "src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1717 | "src/core/NEON/kernels/arm_gemm/transform-sve.cpp", |
| 1718 | "src/core/NEON/kernels/arm_gemm/misc-sve.cpp" |
Francesco.Petrogalli@arm.com | 5fcf22d | 2022-04-05 10:31:08 +0000 | [diff] [blame] | 1719 | ], |
| 1720 | "experimental_fixed_format_kernels": [ |
| 1721 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp", |
| 1722 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp", |
| 1723 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp", |
| 1724 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp", |
| 1725 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp", |
| 1726 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp", |
| 1727 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp", |
| 1728 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp", |
| 1729 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp", |
| 1730 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp", |
| 1731 | "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp" |
| 1732 | ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1733 | } |
| 1734 | } |
| 1735 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1736 | "GenerateProposals": { |
| 1737 | "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1738 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1739 | "common": [ |
| 1740 | "src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp", |
| 1741 | "src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp" |
Dana Zlotnik | 50ce277 | 2021-12-29 13:55:56 +0200 | [diff] [blame] | 1742 | ], |
| 1743 | "neon":{ |
| 1744 | "common":["src/cpu/kernels/genproposals/generic/neon/impl.cpp"], |
| 1745 | "fp16":["src/cpu/kernels/genproposals/generic/neon/fp16.cpp"], |
| 1746 | "fp32":["src/cpu/kernels/genproposals/generic/neon/fp32.cpp"], |
| 1747 | "qsymm16":["src/cpu/kernels/genproposals/generic/neon/qsymm16.cpp"] |
| 1748 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1749 | } |
| 1750 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1751 | "InstanceNormalize": { |
| 1752 | "deps": [ "Permute", "Reduction" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1753 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1754 | "common": [ |
| 1755 | "src/core/NEON/kernels/NEInstanceNormalizationLayerKernel.cpp", |
| 1756 | "src/runtime/NEON/functions/NEInstanceNormalizationLayer.cpp" |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1757 | ], |
Dana Zlotnik | d7e2ec5 | 2022-01-03 10:59:41 +0200 | [diff] [blame] | 1758 | "neon":{ |
| 1759 | "common":["src/cpu/kernels/instancenorm/generic/neon/impl.cpp"], |
| 1760 | "fp16":["src/cpu/kernels/instancenorm/generic/neon/fp16.cpp"], |
| 1761 | "fp32":["src/cpu/kernels/instancenorm/generic/neon/fp32.cpp"] |
| 1762 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1763 | } |
| 1764 | }, |
| 1765 | "L2Normalize": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1766 | "deps": [ "Reduction" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1767 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1768 | "common": [ |
| 1769 | "src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp", |
| 1770 | "src/runtime/NEON/functions/NEL2NormalizeLayer.cpp" |
Yair Schwarzbaum | 5e99318 | 2022-01-10 15:11:07 +0200 | [diff] [blame] | 1771 | ], |
| 1772 | "neon":{ |
| 1773 | "common":["src/cpu/kernels/l2normlayer/generic/neon/impl.cpp"], |
| 1774 | "fp32":["src/cpu/kernels/l2normlayer/generic/neon/fp32.cpp"], |
| 1775 | "fp16":["src/cpu/kernels/l2normlayer/generic/neon/fp16.cpp"] |
| 1776 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1777 | } |
| 1778 | }, |
| 1779 | "Logical": { |
| 1780 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1781 | "common": [ |
| 1782 | "src/core/NEON/kernels/NELogicalKernel.cpp", |
| 1783 | "src/runtime/NEON/functions/NELogical.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1784 | ] |
| 1785 | } |
| 1786 | }, |
Viet-Hoa Do | fd472f0 | 2023-03-15 14:05:06 +0000 | [diff] [blame] | 1787 | "LUT": { |
| 1788 | "files": { |
| 1789 | "common": [], |
| 1790 | "neon":{ |
| 1791 | "qasymm8": ["src/cpu/kernels/lut/generic/neon/u8.cpp"], |
| 1792 | "qasymm8_signed": ["src/cpu/kernels/lut/generic/neon/u8.cpp"] |
| 1793 | }, |
| 1794 | "sve": { |
| 1795 | "qasymm8": ["src/cpu/kernels/lut/generic/sve/u8.cpp"], |
| 1796 | "qasymm8_signed": ["src/cpu/kernels/lut/generic/sve/u8.cpp"] |
| 1797 | } |
| 1798 | } |
| 1799 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1800 | "LSTM": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1801 | "deps": [ |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1802 | "Activation", |
| 1803 | "Concatenate", |
| 1804 | "Copy", |
| 1805 | "Dequantize", |
| 1806 | "ElementwiseBinary", |
| 1807 | "Fill", |
| 1808 | "FullyConnected", |
| 1809 | "Gemm", |
| 1810 | "MeanStdDevNormalize", |
| 1811 | "Mul", |
| 1812 | "Quantize", |
| 1813 | "Slice", |
| 1814 | "Transpose" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1815 | ], |
| 1816 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1817 | "common": [ |
| 1818 | "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp", |
| 1819 | "src/runtime/NEON/functions/NELSTMLayer.cpp", |
| 1820 | "src/runtime/NEON/functions/NELSTMLayerQuantized.cpp", |
| 1821 | "src/runtime/NEON/functions/NEQLSTMLayer.cpp" |
| 1822 | ] |
| 1823 | } |
| 1824 | }, |
| 1825 | "MaxUnpool2d": { |
| 1826 | "deps": [ "Fill" ], |
| 1827 | "files": { |
| 1828 | "common": [ |
Dana Zlotnik | 149203b | 2022-01-26 12:38:03 +0200 | [diff] [blame] | 1829 | "src/cpu/kernels/CpuMaxUnpoolingLayerKernel.cpp", |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1830 | "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp", |
Dana Zlotnik | 149203b | 2022-01-26 12:38:03 +0200 | [diff] [blame] | 1831 | "src/cpu/operators/CpuMaxUnpooling.cpp" |
Dana Zlotnik | 143310c | 2021-12-30 13:24:52 +0200 | [diff] [blame] | 1832 | ], |
| 1833 | "neon":{ |
Dana Zlotnik | 143310c | 2021-12-30 13:24:52 +0200 | [diff] [blame] | 1834 | "fp32":["src/cpu/kernels/maxunpool/generic/neon/fp32.cpp"], |
| 1835 | "fp16":["src/cpu/kernels/maxunpool/generic/neon/fp16.cpp"], |
| 1836 | "qasymm8":["src/cpu/kernels/maxunpool/generic/neon/qasymm8.cpp"], |
Dana Zlotnik | 149203b | 2022-01-26 12:38:03 +0200 | [diff] [blame] | 1837 | "qasymm8_signed":[ "src/cpu/kernels/maxunpool/generic/neon/qasymm8_signed.cpp"], |
| 1838 | "common":["src/cpu/kernels/maxunpool/generic/neon/impl.cpp"] |
Dana Zlotnik | 143310c | 2021-12-30 13:24:52 +0200 | [diff] [blame] | 1839 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1840 | } |
| 1841 | }, |
| 1842 | "Mean": { |
| 1843 | "deps" : [ "Reduction" ], |
| 1844 | "files": { |
| 1845 | "common": [ "src/runtime/NEON/functions/NEReduceMean.cpp" ] |
| 1846 | } |
| 1847 | }, |
| 1848 | "MeanStdDevNormalize": { |
| 1849 | "files": { |
| 1850 | "common": [ |
| 1851 | "src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp", |
| 1852 | "src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp" |
Dana Zlotnik | 027bcef | 2021-12-27 17:35:00 +0200 | [diff] [blame] | 1853 | ], |
| 1854 | "neon":{ |
| 1855 | "common":["src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp"], |
| 1856 | "fp32":["src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp"], |
Murray Kornelsen | 6e09e14 | 2022-07-13 21:40:26 -0400 | [diff] [blame] | 1857 | "fp16":["src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp"], |
| 1858 | "qasymm8":["src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp"] |
Dana Zlotnik | 027bcef | 2021-12-27 17:35:00 +0200 | [diff] [blame] | 1859 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1860 | } |
| 1861 | }, |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 1862 | "MatMul" : { |
| 1863 | "files": { |
| 1864 | "common": [ |
| 1865 | "src/cpu/operators/CpuMatMul.cpp", |
| 1866 | "src/runtime/NEON/functions/NEMatMul.cpp" |
| 1867 | ] |
| 1868 | } |
| 1869 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1870 | "Mul": { |
| 1871 | "files": { |
| 1872 | "common": [ |
| 1873 | "src/cpu/operators/CpuMul.cpp", |
| 1874 | "src/cpu/kernels/CpuMulKernel.cpp", |
| 1875 | "src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp" |
| 1876 | ] |
| 1877 | } |
| 1878 | }, |
| 1879 | "Normalize": { |
| 1880 | "deps": [ "Mul" ], |
| 1881 | "files": { |
| 1882 | "common": [ |
| 1883 | "src/core/NEON/kernels/NENormalizationLayerKernel.cpp", |
| 1884 | "src/runtime/NEON/functions/NENormalizationLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1885 | ] |
| 1886 | } |
| 1887 | }, |
| 1888 | "Pad": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1889 | "deps": [ "Concatenate", "Copy", "StridedSlice" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1890 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1891 | "common": [ |
| 1892 | "src/core/NEON/kernels/NEPadLayerKernel.cpp", |
| 1893 | "src/runtime/NEON/functions/NEPadLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1894 | ] |
| 1895 | } |
| 1896 | }, |
| 1897 | "Permute": { |
| 1898 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1899 | "common": [ |
| 1900 | "src/cpu/operators/CpuPermute.cpp", |
| 1901 | "src/cpu/kernels/CpuPermuteKernel.cpp", |
| 1902 | "src/runtime/NEON/functions/NEPermute.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1903 | ] |
| 1904 | } |
| 1905 | }, |
| 1906 | "Pool2d": { |
| 1907 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1908 | "common": [ |
| 1909 | "src/cpu/operators/CpuPool2d.cpp", |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 1910 | "src/cpu/kernels/CpuPool2dKernel.cpp", |
| 1911 | "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp", |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1912 | "src/runtime/NEON/functions/NEPoolingLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1913 | ], |
| 1914 | "neon": { |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1915 | "common": [ |
| 1916 | "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp", |
| 1917 | "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp", |
| 1918 | "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp", |
| 1919 | "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp", |
| 1920 | "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp", |
| 1921 | "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp", |
| 1922 | "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp", |
| 1923 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1924 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1925 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1926 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp", |
| 1927 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1928 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1929 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1930 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp", |
| 1931 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1932 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1933 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1934 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1935 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp", |
| 1936 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1937 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1938 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1939 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp", |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1940 | "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp" |
Motti Gondabi | 9d9ad33 | 2022-01-23 12:42:24 +0200 | [diff] [blame] | 1941 | ], |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1942 | "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ], |
| 1943 | "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ], |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 1944 | "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ], |
| 1945 | "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ], |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 1946 | "qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1947 | }, |
| 1948 | "sve": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1949 | "common": [ |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1950 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1951 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1952 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1953 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp", |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 1954 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1955 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1956 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1957 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_generic_depthfirst/generic.cpp", |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1958 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1959 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1960 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1961 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1962 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp", |
| 1963 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1964 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1965 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1966 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp", |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 1967 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp", |
| 1968 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1969 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1970 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1971 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_generic_depthfirst/generic.cpp", |
| 1972 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", |
| 1973 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1974 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1975 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp", |
| 1976 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1977 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1978 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1979 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1980 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp", |
| 1981 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1982 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", |
| 1983 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_generic_depthfirst/generic.cpp", |
| 1984 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_avg_generic_depthfirst/generic.cpp", |
| 1985 | "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 1986 | ] |
| 1987 | } |
| 1988 | } |
| 1989 | }, |
Adnan AlSinan | 171fc3d | 2022-03-15 18:46:42 +0000 | [diff] [blame] | 1990 | "Pool3d": { |
| 1991 | "files": { |
| 1992 | "common": [ |
| 1993 | "src/cpu/operators/CpuPool3d.cpp", |
| 1994 | "src/cpu/kernels/CpuPool3dKernel.cpp", |
| 1995 | "src/runtime/NEON/functions/NEPooling3dLayer.cpp" |
| 1996 | ], |
| 1997 | "neon": { |
Adnan AlSinan | 9104cd5 | 2022-04-06 16:19:31 +0100 | [diff] [blame] | 1998 | "common": [ "src/cpu/kernels/pool3d/neon/impl.cpp" ], |
| 1999 | "fp16": [ "src/cpu/kernels/pool3d/neon/fp16.cpp" ], |
| 2000 | "fp32": [ "src/cpu/kernels/pool3d/neon/fp32.cpp" ], |
| 2001 | "qasymm8": [ "src/cpu/kernels/pool3d/neon/qasymm8.cpp" ], |
| 2002 | "qasymm8_signed": [ "src/cpu/kernels/pool3d/neon/qasymm8_signed.cpp" ] |
Adnan AlSinan | 171fc3d | 2022-03-15 18:46:42 +0000 | [diff] [blame] | 2003 | } |
| 2004 | } |
| 2005 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2006 | "PRelu": { |
| 2007 | "deps": [ "ElementwiseBinary" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2008 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2009 | "common": [ |
| 2010 | "src/runtime/NEON/functions/NEPReluLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2011 | ] |
| 2012 | } |
| 2013 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2014 | "PriorBox": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2015 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2016 | "common": [ |
| 2017 | "src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp", |
| 2018 | "src/runtime/NEON/functions/NEPriorBoxLayer.cpp" |
| 2019 | ] |
| 2020 | } |
| 2021 | }, |
| 2022 | "Quantize": { |
| 2023 | "files": { |
| 2024 | "common": [ |
| 2025 | "src/cpu/operators/CpuQuantize.cpp", |
| 2026 | "src/cpu/kernels/CpuQuantizeKernel.cpp", |
| 2027 | "src/runtime/NEON/functions/NEQuantizationLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2028 | ] |
| 2029 | } |
| 2030 | }, |
| 2031 | "Range": { |
| 2032 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2033 | "common": [ |
| 2034 | "src/core/NEON/kernels/NERangeKernel.cpp", |
| 2035 | "src/runtime/NEON/functions/NERange.cpp" |
Yair Schwarzbaum | 9e8a770 | 2021-11-08 10:58:06 +0200 | [diff] [blame] | 2036 | ], |
| 2037 | "neon": { |
| 2038 | "common": [ "src/cpu/kernels/range/generic/neon/impl.cpp" ], |
| 2039 | "fp32": [ "src/cpu/kernels/range/generic/neon/fp32.cpp" ], |
| 2040 | "fp16": [ "src/cpu/kernels/range/generic/neon/fp16.cpp" ], |
| 2041 | "integer": [ "src/cpu/kernels/range/generic/neon/integer.cpp" ] |
| 2042 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2043 | } |
| 2044 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2045 | "Reduction":{ |
| 2046 | "deps": [ "Reshape" ], |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2047 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2048 | "common": [ |
| 2049 | "src/core/NEON/kernels/NEReductionOperationKernel.cpp", |
| 2050 | "src/runtime/NEON/functions/NEReductionOperation.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2051 | ] |
| 2052 | } |
| 2053 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2054 | "Reorg": { |
| 2055 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2056 | "common": [ |
| 2057 | "src/core/NEON/kernels/NEReorgLayerKernel.cpp", |
| 2058 | "src/runtime/NEON/functions/NEReorgLayer.cpp" |
| 2059 | ] |
| 2060 | } |
| 2061 | }, |
| 2062 | "Reshape": { |
| 2063 | "files": { |
| 2064 | "common": [ |
| 2065 | "src/cpu/operators/CpuReshape.cpp", |
| 2066 | "src/cpu/kernels/CpuReshapeKernel.cpp", |
| 2067 | "src/runtime/NEON/functions/NEReshapeLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2068 | ] |
| 2069 | } |
| 2070 | }, |
| 2071 | "Reverse": { |
| 2072 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2073 | "common": [ |
| 2074 | "src/core/NEON/kernels/NEReverseKernel.cpp", |
| 2075 | "src/runtime/NEON/functions/NEReverse.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2076 | ] |
| 2077 | } |
| 2078 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2079 | "RNN": { |
| 2080 | "deps": [ "Activation", "Add", "FullyConnected", "Gemm"], |
| 2081 | "files": { |
| 2082 | "common": [ "src/runtime/NEON/functions/NERNNLayer.cpp" ] |
| 2083 | } |
| 2084 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2085 | "ROIAlign": { |
| 2086 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2087 | "common": [ |
| 2088 | "src/core/NEON/kernels/NEROIAlignLayerKernel.cpp", |
| 2089 | "src/runtime/NEON/functions/NEROIAlignLayer.cpp" |
Dana Zlotnik | 9ae5a4e | 2022-01-03 16:22:47 +0200 | [diff] [blame] | 2090 | ], |
| 2091 | "neon":{ |
| 2092 | "common":["src/cpu/kernels/roialign/generic/neon/impl.cpp"], |
| 2093 | "fp32":["src/cpu/kernels/roialign/generic/neon/fp32.cpp"], |
| 2094 | "fp16":["src/cpu/kernels/roialign/generic/neon/fp16.cpp"], |
| 2095 | "qasymm8":["src/cpu/kernels/roialign/generic/neon/qasymm8.cpp"], |
| 2096 | "qasymm8_signed":["src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp"] |
| 2097 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2098 | } |
| 2099 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2100 | "ROIPool2d": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2101 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2102 | "common": [ |
| 2103 | "src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp", |
| 2104 | "src/runtime/NEON/functions/NEROIPoolingLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2105 | ] |
| 2106 | } |
| 2107 | }, |
| 2108 | "Scale": { |
| 2109 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2110 | "common": [ |
| 2111 | "src/cpu/operators/CpuScale.cpp", |
| 2112 | "src/cpu/kernels/CpuScaleKernel.cpp", |
| 2113 | "src/runtime/NEON/functions/NEScale.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2114 | ], |
| 2115 | "sve": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2116 | "fp16": [ "src/cpu/kernels/scale/sve/fp16.cpp" ], |
| 2117 | "fp32": [ "src/cpu/kernels/scale/sve/fp32.cpp" ], |
| 2118 | "integer": [ "src/cpu/kernels/scale/sve/integer.cpp" ], |
| 2119 | "qasymm8": [ "src/cpu/kernels/scale/sve/qasymm8.cpp" ], |
| 2120 | "qasymm8_signed": [ "src/cpu/kernels/scale/sve/qasymm8_signed.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2121 | }, |
| 2122 | "neon": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2123 | "fp16": [ "src/cpu/kernels/scale/neon/fp16.cpp" ], |
| 2124 | "integer": [ "src/cpu/kernels/scale/neon/integer.cpp" ], |
Gunes Bayir | c4f2743 | 2022-09-11 15:59:19 +0100 | [diff] [blame] | 2125 | "qasymm8": [ "src/cpu/kernels/scale/neon/qasymm8.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ], |
| 2126 | "qasymm8_signed": [ "src/cpu/kernels/scale/neon/qasymm8_signed.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2127 | } |
| 2128 | } |
| 2129 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2130 | "Select": { |
| 2131 | "files": { |
| 2132 | "common": [ |
| 2133 | "src/core/NEON/kernels/NESelectKernel.cpp", |
| 2134 | "src/runtime/NEON/functions/NESelect.cpp" |
Anton Vainer | 8a9a0fb | 2022-01-09 14:37:12 +0200 | [diff] [blame] | 2135 | ], |
| 2136 | "neon": { |
| 2137 | "common": [ "src/cpu/kernels/select/generic/neon/impl.cpp" ], |
| 2138 | "fp32": [ "src/cpu/kernels/select/generic/neon/fp32.cpp" ], |
| 2139 | "fp16": [ "src/cpu/kernels/select/generic/neon/fp16.cpp" ], |
| 2140 | "integer": [ "src/cpu/kernels/select/generic/neon/integer.cpp" ] |
| 2141 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2142 | } |
| 2143 | }, |
| 2144 | "Slice": { |
| 2145 | "deps": [ "StridedSlice" ], |
| 2146 | "files": { |
| 2147 | "common": [ "src/runtime/NEON/functions/NESlice.cpp" ] |
| 2148 | } |
| 2149 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2150 | "Softmax": { |
| 2151 | "deps": [ |
| 2152 | "Permute" |
| 2153 | ], |
| 2154 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2155 | "common": [ |
| 2156 | "src/cpu/operators/CpuSoftmax.cpp", |
| 2157 | "src/cpu/kernels/CpuSoftmaxKernel.cpp", |
| 2158 | "src/runtime/NEON/functions/NESoftmaxLayer.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2159 | ], |
Dana Zlotnik | c48a3e5 | 2021-12-21 13:34:42 +0200 | [diff] [blame] | 2160 | "neon":{ |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 2161 | "common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"], |
Dana Zlotnik | c48a3e5 | 2021-12-21 13:34:42 +0200 | [diff] [blame] | 2162 | "fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"], |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 2163 | "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"], |
| 2164 | "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"], |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 2165 | "qasymm8_signed":["src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp"] |
Dana Zlotnik | c48a3e5 | 2021-12-21 13:34:42 +0200 | [diff] [blame] | 2166 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2167 | "sve": { |
Dana Zlotnik | c48a3e5 | 2021-12-21 13:34:42 +0200 | [diff] [blame] | 2168 | "common": [ "src/cpu/kernels/softmax/generic/sve/impl.cpp" ], |
| 2169 | "fp32": ["src/cpu/kernels/softmax/generic/sve/fp32.cpp"], |
| 2170 | "fp16": ["src/cpu/kernels/softmax/generic/sve/fp16.cpp"], |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 2171 | "qasymm8": ["src/cpu/kernels/softmax/generic/sve/qasymm8.cpp" ], |
| 2172 | "qasymm8_signed": ["src/cpu/kernels/softmax/generic/sve/qasymm8_signed.cpp"] |
| 2173 | }, |
| 2174 | "sve2":{ |
Dana Zlotnik | a538ae5 | 2022-02-21 13:12:41 +0200 | [diff] [blame] | 2175 | "common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"], |
ramelg01 | a1f7851 | 2022-06-29 16:28:10 +0100 | [diff] [blame] | 2176 | "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"], |
Dana Zlotnik | 6a2df88 | 2022-01-17 09:54:26 +0200 | [diff] [blame] | 2177 | "qasymm8_signed":["src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp"] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2178 | } |
| 2179 | } |
| 2180 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2181 | "SpaceToBatch": { |
| 2182 | "files": { |
| 2183 | "common": [ |
| 2184 | "src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp", |
| 2185 | "src/runtime/NEON/functions/NESpaceToBatchLayer.cpp" |
| 2186 | ] |
| 2187 | } |
| 2188 | }, |
| 2189 | "SpaceToDepth": { |
| 2190 | "files": { |
| 2191 | "common": [ |
| 2192 | "src/core/NEON/kernels/NESpaceToDepthLayerKernel.cpp", |
| 2193 | "src/runtime/NEON/functions/NESpaceToDepthLayer.cpp" |
| 2194 | ] |
| 2195 | } |
| 2196 | }, |
| 2197 | "Split": { |
| 2198 | "deps": [ "StridedSlice" ], |
| 2199 | "files": { |
| 2200 | "common": [ |
| 2201 | "src/runtime/NEON/functions/NESplit.cpp" |
| 2202 | ] |
| 2203 | } |
| 2204 | }, |
| 2205 | "Stack": { |
| 2206 | "files": { |
| 2207 | "common": [ |
| 2208 | "src/core/NEON/kernels/NEStackLayerKernel.cpp", |
| 2209 | "src/runtime/NEON/functions/NEStackLayer.cpp" |
| 2210 | ] |
| 2211 | } |
| 2212 | }, |
| 2213 | "StridedSlice": { |
| 2214 | "files": { |
| 2215 | "common": [ |
| 2216 | "src/core/NEON/kernels/NEStridedSliceKernel.cpp", |
| 2217 | "src/runtime/NEON/functions/NEStridedSlice.cpp" |
| 2218 | ] |
| 2219 | } |
| 2220 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2221 | "Sub": { |
| 2222 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2223 | "common": [ |
| 2224 | "src/cpu/operators/CpuSub.cpp", |
| 2225 | "src/cpu/kernels/CpuSubKernel.cpp", |
Freddie Liardet | 487d390 | 2021-09-21 12:36:43 +0100 | [diff] [blame] | 2226 | "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp", |
| 2227 | "src/cpu/kernels/sub/neon/qasymm8.cpp", |
| 2228 | "src/cpu/kernels/sub/neon/qasymm8_signed.cpp", |
| 2229 | "src/cpu/kernels/sub/neon/qsymm16.cpp" |
| 2230 | ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2231 | } |
| 2232 | }, |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2233 | "Tile": { |
| 2234 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2235 | "common": [ |
| 2236 | "src/core/NEON/kernels/NETileKernel.cpp", |
| 2237 | "src/runtime/NEON/functions/NETile.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2238 | ] |
| 2239 | } |
| 2240 | }, |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2241 | "Transpose": { |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2242 | "files": { |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2243 | "common": [ |
| 2244 | "src/cpu/kernels/CpuTransposeKernel.cpp", |
| 2245 | "src/cpu/operators/CpuTranspose.cpp", |
| 2246 | "src/runtime/NEON/functions/NETranspose.cpp" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2247 | ] |
| 2248 | } |
Georgios Pinitas | b6af482 | 2021-09-14 12:33:34 +0100 | [diff] [blame] | 2249 | }, |
| 2250 | "Unstack": { |
| 2251 | "deps": [ "StridedSlice" ], |
| 2252 | "files": { |
| 2253 | "common": [ "src/runtime/NEON/functions/NEUnstack.cpp" ] |
| 2254 | } |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2255 | } |
| 2256 | } |
Giorgio Arena | 232c452 | 2022-03-03 10:09:01 +0000 | [diff] [blame] | 2257 | }, |
| 2258 | "experimental": { |
| 2259 | "dynamic_fusion": [ |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2260 | "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp", |
| 2261 | "src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp", |
Gunes Bayir | 1dc6ff1 | 2022-12-06 20:48:31 +0000 | [diff] [blame] | 2262 | "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp", |
Jakub Sujak | 3274172 | 2022-11-25 16:43:18 +0000 | [diff] [blame] | 2263 | "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp", |
SiCong Li | 5a63d1e | 2023-01-06 16:28:57 +0000 | [diff] [blame] | 2264 | "src/dynamic_fusion/sketch/attributes/Conv2dAttributes.cpp", |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 2265 | "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp", |
Mohammed Suhail Munshi | a18d85c | 2023-01-03 10:16:16 +0000 | [diff] [blame] | 2266 | "src/dynamic_fusion/sketch/attributes/Pool2dAttributes.cpp", |
Jakub Sujak | 8ae5714 | 2022-12-02 16:09:06 +0000 | [diff] [blame] | 2267 | "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp", |
Gunes Bayir | aecb5d9 | 2022-12-18 21:31:29 +0000 | [diff] [blame] | 2268 | "src/dynamic_fusion/sketch/attributes/SoftmaxAttributes.cpp", |
Ramy Elgammal | f800adf | 2022-12-14 15:39:29 +0000 | [diff] [blame] | 2269 | "src/dynamic_fusion/sketch/attributes/ReshapeAttributes.cpp", |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2270 | "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.cpp", |
| 2271 | "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp", |
| 2272 | "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp", |
| 2273 | "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.cpp", |
| 2274 | "src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp", |
| 2275 | "src/dynamic_fusion/sketch/gpu/GpuOperatorGroup.cpp", |
| 2276 | "src/dynamic_fusion/sketch/gpu/GpuWorkloadContext.cpp", |
| 2277 | "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.cpp", |
Jakub Sujak | 3274172 | 2022-11-25 16:43:18 +0000 | [diff] [blame] | 2278 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.cpp", |
Gunes Bayir | 1dc6ff1 | 2022-12-06 20:48:31 +0000 | [diff] [blame] | 2279 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentCast.cpp", |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 2280 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp", |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2281 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp", |
Mohammed Suhail Munshi | a18d85c | 2023-01-03 10:16:16 +0000 | [diff] [blame] | 2282 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentPool2d.cpp", |
Ramy Elgammal | 404462a | 2022-11-08 02:14:46 +0000 | [diff] [blame] | 2283 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp", |
Gunes Bayir | aecb5d9 | 2022-12-18 21:31:29 +0000 | [diff] [blame] | 2284 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp", |
Ramy Elgammal | 002e653 | 2023-01-11 18:48:04 +0000 | [diff] [blame] | 2285 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp", |
Ramy Elgammal | f800adf | 2022-12-14 15:39:29 +0000 | [diff] [blame] | 2286 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentReshape.cpp", |
Jakub Sujak | 8ae5714 | 2022-12-02 16:09:06 +0000 | [diff] [blame] | 2287 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp", |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2288 | "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp", |
Ramy Elgammal | 404462a | 2022-11-08 02:14:46 +0000 | [diff] [blame] | 2289 | "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp", |
Gunes Bayir | 1dc6ff1 | 2022-12-06 20:48:31 +0000 | [diff] [blame] | 2290 | "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp", |
Jakub Sujak | 3274172 | 2022-11-25 16:43:18 +0000 | [diff] [blame] | 2291 | "src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp", |
Gunes Bayir | 1dc6ff1 | 2022-12-06 20:48:31 +0000 | [diff] [blame] | 2292 | "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp", |
Jakub Sujak | 7359a87 | 2023-01-05 14:24:13 +0000 | [diff] [blame] | 2293 | "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp", |
| 2294 | "src/dynamic_fusion/sketch/gpu/operators/GpuMul.cpp", |
Ramy Elgammal | f800adf | 2022-12-14 15:39:29 +0000 | [diff] [blame] | 2295 | "src/dynamic_fusion/sketch/gpu/operators/GpuReshape.cpp", |
Mohammed Suhail Munshi | a18d85c | 2023-01-03 10:16:16 +0000 | [diff] [blame] | 2296 | "src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp", |
Viet-Hoa Do | b84e253 | 2022-12-13 13:09:10 +0000 | [diff] [blame] | 2297 | "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp", |
Jakub Sujak | 8ae5714 | 2022-12-02 16:09:06 +0000 | [diff] [blame] | 2298 | "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", |
Viet-Hoa Do | 98aca0f | 2023-03-02 17:43:45 +0000 | [diff] [blame] | 2299 | "src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp", |
Ramy Elgammal | 002e653 | 2023-01-11 18:48:04 +0000 | [diff] [blame] | 2300 | "src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp", |
Ramy Elgammal | ec320d9 | 2022-12-14 09:20:09 +0000 | [diff] [blame] | 2301 | "src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp", |
Viet-Hoa Do | 98aca0f | 2023-03-02 17:43:45 +0000 | [diff] [blame] | 2302 | "src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp", |
Ramy Elgammal | 404462a | 2022-11-08 02:14:46 +0000 | [diff] [blame] | 2303 | "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp", |
Jakub Sujak | 3274172 | 2022-11-25 16:43:18 +0000 | [diff] [blame] | 2304 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp", |
Gunes Bayir | 1dc6ff1 | 2022-12-06 20:48:31 +0000 | [diff] [blame] | 2305 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateCast.cpp", |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 2306 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp", |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2307 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp", |
Mohammed Suhail Munshi | a18d85c | 2023-01-03 10:16:16 +0000 | [diff] [blame] | 2308 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplatePool2d.cpp", |
Ramy Elgammal | 404462a | 2022-11-08 02:14:46 +0000 | [diff] [blame] | 2309 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp", |
Gunes Bayir | aecb5d9 | 2022-12-18 21:31:29 +0000 | [diff] [blame] | 2310 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DMaxShiftExpSum.cpp", |
Ramy Elgammal | 002e653 | 2023-01-11 18:48:04 +0000 | [diff] [blame] | 2311 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DNorm.cpp", |
Ramy Elgammal | f800adf | 2022-12-14 15:39:29 +0000 | [diff] [blame] | 2312 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateReshape.cpp", |
Jakub Sujak | 8ae5714 | 2022-12-02 16:09:06 +0000 | [diff] [blame] | 2313 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp", |
SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 2314 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp", |
| 2315 | "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp", |
| 2316 | "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp" |
Giorgio Arena | 232c452 | 2022-03-03 10:09:01 +0000 | [diff] [blame] | 2317 | ] |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 2318 | } |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 2319 | } |