1{ 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", 9 "src/common/TensorPack.cpp", 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" 72 ], 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 }, 84 "c_api": { 85 "common": [ 86 "src/c/AclContext.cpp", 87 "src/c/AclOperator.cpp", 88 "src/c/AclQueue.cpp", 89 "src/c/AclTensor.cpp", 90 "src/c/AclTensorPack.cpp", 91 "src/c/AclVersion.cpp" 92 ], 93 "cpu": [ 94 ], 95 "gpu": [ 96 "src/c/cl/AclOpenClExt.cpp" 97 ], 98 "operators": 99 [ 100 "src/c/operators/AclActivation.cpp" 101 ] 102 }, 103 "high_priority": [ 104 "Activation", 105 "DepthwiseConv2d", 106 "Conv2d", 107 "Permute", 108 "Pool2d", 109 "Reshape" 110 ], 111 "gpu": { 112 "common": [ 113 "src/core/CL/CLCompileContext.cpp", 114 "src/core/CL/DefaultLWSHeuristics.cpp", 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", 124 "src/gpu/cl/ClContext.cpp", 125 "src/gpu/cl/ClKernelLibrary.cpp", 126 "src/gpu/cl/ClQueue.cpp", 127 "src/gpu/cl/ClTensor.cpp", 128 "src/core/CL/kernels/CLFillBorderKernel.cpp", 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", 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 ], 149 "operators": { 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", 291 "src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp", 292 "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.cpp", 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", 297 "src/gpu/cl/operators/ClIndirectConv2d.cpp", 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", 304 "src/runtime/CL/functions/CLIndirectConvolutionLayer.cpp", 305 "src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp" 306 ] 307 } 308 }, 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 }, 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", 344 "src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp", 345 "src/gpu/cl/operators/ClTransposedConvolution.cpp", 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", 480 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp", 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", 486 "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp", 487 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp", 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", 502 "src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp", 503 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp", 504 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp", 505 "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigBifrost.cpp", 506 "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigValhall.cpp", 507 "src/runtime/heuristics/dwc_native/ClDWCNativeHeuristicsHelpers.cpp", 508 "src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp" 509 ] 510 } 511 }, 512 "GenerateProposals": { 513 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ], 514 "files": { 515 "common": [ 516 "src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp", 517 "src/runtime/CL/functions/CLGenerateProposalsLayer.cpp" 518 ] 519 } 520 }, 521 "InstanceNormalize": { 522 "files": { 523 "common": [ 524 "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp", 525 "src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp" 526 ] 527 } 528 }, 529 "L2Normalize": { 530 "deps": [ "Reduction" ], 531 "files": { 532 "common": [ 533 "src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp", 534 "src/runtime/CL/functions/CLL2NormalizeLayer.cpp" 535 ] 536 } 537 }, 538 "Logical": { 539 "files": { 540 "common": [ 541 "src/gpu/cl/operators/ClLogicalNot.cpp", 542 "src/runtime/CL/functions/CLLogicalAnd.cpp", 543 "src/runtime/CL/functions/CLLogicalNot.cpp", 544 "src/runtime/CL/functions/CLLogicalOr.cpp" 545 ] 546 } 547 }, 548 "LSTM": { 549 "deps": [ 550 "Activation", 551 "Concatenate", 552 "Copy", 553 "Dequantize", 554 "ElementwiseBinary", 555 "Fill", 556 "FullyConnected", 557 "Gemm", 558 "MeanStdDevNormalize", 559 "Mul", 560 "Quantize", 561 "Slice", 562 "Transpose" 563 ], 564 "files": { 565 "common": [ 566 "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp", 567 "src/runtime/CL/functions/CLQLSTMLayer.cpp", 568 "src/runtime/CL/functions/CLLSTMLayer.cpp", 569 "src/runtime/CL/functions/CLLSTMLayerQuantized.cpp" 570 ] 571 } 572 }, 573 "MaxUnpool2d": { 574 "deps": [ "Fill" ], 575 "files": { 576 "common": [ 577 "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.cpp", 578 "src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp" 579 ] 580 } 581 }, 582 "MeanStdDevNormalize": { 583 "deps": [ "Reduction" ], 584 "files": { 585 "common": [ 586 "src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp", 587 "src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp", 588 "src/runtime/CL/functions/CLReduceMean.cpp" 589 ] 590 } 591 }, 592 "Mul": { 593 "files": { 594 "common": [ 595 "src/gpu/cl/kernels/ClMulKernel.cpp", 596 "src/gpu/cl/operators/ClMul.cpp", 597 "src/runtime/CL/functions/CLPixelWiseMultiplication.cpp" 598 ] 599 } 600 }, 601 "Normalize": { 602 "files": { 603 "common": [ 604 "src/core/CL/kernels/CLNormalizationLayerKernel.cpp", 605 "src/runtime/CL/functions/CLNormalizationLayer.cpp" 606 ] 607 } 608 }, 609 "Pad": { 610 "deps": [ "Copy" ], 611 "files": { 612 "common": [ 613 "src/core/CL/kernels/CLPadLayerKernel.cpp", 614 "src/runtime/CL/functions/CLPadLayer.cpp" 615 ] 616 } 617 }, 618 "Permute": { 619 "files": { 620 "common": [ 621 "src/gpu/cl/kernels/ClPermuteKernel.cpp", 622 "src/gpu/cl/operators/ClPermute.cpp", 623 "src/runtime/CL/functions/CLPermute.cpp" 624 ] 625 } 626 }, 627 "Pool2d": { 628 "files": { 629 "common": [ 630 "src/gpu/cl/kernels/ClPool2dKernel.cpp", 631 "src/gpu/cl/operators/ClPool2d.cpp", 632 "src/runtime/CL/functions/CLPoolingLayer.cpp" 633 ] 634 } 635 }, 636 "Pool3d": { 637 "files": { 638 "common": [ 639 "src/gpu/cl/kernels/ClPool3dKernel.cpp", 640 "src/gpu/cl/operators/ClPool3d.cpp", 641 "src/runtime/CL/functions/CLPooling3dLayer.cpp" 642 ] 643 } 644 }, 645 "PRelu": { 646 "deps": [ "ElementwiseBinary" ], 647 "files": { 648 "common": [ 649 "src/gpu/cl/operators/ClPRelu.cpp", 650 "src/runtime/CL/functions/CLPReluLayer.cpp" 651 ] 652 } 653 }, 654 "PriorBox": { 655 "files": { 656 "common": [ 657 "src/core/CL/kernels/CLPriorBoxLayerKernel.cpp", 658 "src/runtime/CL/functions/CLPriorBoxLayer.cpp" 659 ] 660 } 661 }, 662 "Quantize": { 663 "files": { 664 "common": [ 665 "src/gpu/cl/kernels/ClQuantizeKernel.cpp", 666 "src/gpu/cl/operators/ClQuantize.cpp", 667 "src/runtime/CL/functions/CLQuantizationLayer.cpp" 668 ] 669 } 670 }, 671 "Range": { 672 "files": { 673 "common": [ 674 "src/core/CL/kernels/CLRangeKernel.cpp", 675 "src/runtime/CL/functions/CLRange.cpp" 676 ] 677 } 678 }, 679 "Reduction": { 680 "deps": [ "Reshape" ], 681 "files": { 682 "common": [ 683 "src/core/CL/kernels/CLReductionOperationKernel.cpp", 684 "src/runtime/CL/functions/CLReductionOperation.cpp" 685 ] 686 } 687 }, 688 "Reorg": { 689 "files": { 690 "common": [ 691 "src/core/CL/kernels/CLReorgLayerKernel.cpp", 692 "src/runtime/CL/functions/CLReorgLayer.cpp" 693 ] 694 } 695 }, 696 "Reshape": { 697 "files": { 698 "common": [ 699 "src/gpu/cl/kernels/ClReshapeKernel.cpp", 700 "src/gpu/cl/operators/ClReshape.cpp", 701 "src/runtime/CL/functions/CLReshapeLayer.cpp" 702 ] 703 } 704 }, 705 "Reverse": { 706 "files": { 707 "common": [ 708 "src/core/CL/kernels/CLReverseKernel.cpp", 709 "src/runtime/CL/functions/CLReverse.cpp" 710 ] 711 } 712 }, 713 "RNN": { 714 "deps": [ "Activation", "Cast", "ElementwiseBinary", "FullyConnected", "Gemm"], 715 "files": { 716 "common": [ "src/runtime/CL/functions/CLRNNLayer.cpp" ] 717 } 718 }, 719 "ROIAlign": { 720 "files": { 721 "common": [ 722 "src/core/CL/kernels/CLROIAlignLayerKernel.cpp", 723 "src/runtime/CL/functions/CLROIAlignLayer.cpp" 724 ] 725 } 726 }, 727 "ROIPool2d": { 728 "files": { 729 "common": [ 730 "src/core/CL/kernels/CLROIPoolingLayerKernel.cpp", 731 "src/runtime/CL/functions/CLROIPoolingLayer.cpp" 732 ] 733 } 734 }, 735 "Scale": { 736 "files": { 737 "common": [ 738 "src/gpu/cl/kernels/ClScaleKernel.cpp", 739 "src/gpu/cl/operators/ClScale.cpp", 740 "src/runtime/CL/functions/CLScale.cpp" 741 ] 742 } 743 }, 744 "Select": { 745 "files": { 746 "common": [ 747 "src/core/CL/kernels/CLSelectKernel.cpp", 748 "src/runtime/CL/functions/CLSelect.cpp" 749 ] 750 } 751 }, 752 "Slice": { 753 "deps": [ "StridedSlice" ], 754 "files": { 755 "common": [ "src/runtime/CL/functions/CLSlice.cpp" ] 756 } 757 }, 758 "Softmax": { 759 "deps": [ "Permute" ], 760 "files": { 761 "common": [ 762 "src/gpu/cl/kernels/ClSoftmaxKernel.cpp", 763 "src/gpu/cl/operators/ClSoftmax.cpp", 764 "src/runtime/CL/functions/CLSoftmaxLayer.cpp" 765 ] 766 } 767 }, 768 "SpaceToBatch": { 769 "files": { 770 "common": [ 771 "src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp", 772 "src/runtime/CL/functions/CLSpaceToBatchLayer.cpp" 773 ] 774 } 775 }, 776 "SpaceToDepth": { 777 "files": { 778 "common": [ 779 "src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp", 780 "src/runtime/CL/functions/CLSpaceToDepthLayer.cpp" 781 ] 782 } 783 }, 784 "Split": { 785 "deps": [ "StridedSlice" ], 786 "files": { 787 "common": [ "src/runtime/CL/functions/CLSplit.cpp" ] 788 } 789 }, 790 "Stack": { 791 "files": { 792 "common": [ 793 "src/core/CL/kernels/CLStackLayerKernel.cpp", 794 "src/runtime/CL/functions/CLStackLayer.cpp" 795 ] 796 } 797 }, 798 "StridedSlice": { 799 "files": { 800 "common": [ 801 "src/core/CL/kernels/CLStridedSliceKernel.cpp", 802 "src/runtime/CL/functions/CLStridedSlice.cpp" 803 ] 804 } 805 }, 806 "Sub": { 807 "files": { 808 "common": [ 809 "src/gpu/cl/kernels/ClElementwiseKernel.cpp", 810 "src/gpu/cl/operators/ClSub.cpp" 811 ] 812 } 813 }, 814 "Tile": { 815 "files": { 816 "common": [ 817 "src/core/CL/kernels/CLTileKernel.cpp", 818 "src/runtime/CL/functions/CLTile.cpp" 819 ] 820 } 821 }, 822 "Transpose": { 823 "files": { 824 "common": [ 825 "src/gpu/cl/kernels/ClTransposeKernel.cpp", 826 "src/gpu/cl/operators/ClTranspose.cpp", 827 "src/runtime/CL/functions/CLTranspose.cpp" 828 ] 829 } 830 }, 831 "Unstack": { 832 "deps": [ "StridedSlice" ], 833 "files": { 834 "common": [ "src/runtime/CL/functions/CLUnstack.cpp" ] 835 } 836 }, 837 "YUVNormalize": { 838 "files": { 839 "common": [ 840 "src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp", 841 "src/runtime/CL/functions/CLNormalizePlanarYUVLayer.cpp" 842 ] 843 } 844 } 845 } 846}, 847 "cpu": { 848 "common": [ 849 "src/cpu/CpuContext.cpp", 850 "src/cpu/CpuQueue.cpp", 851 "src/cpu/CpuTensor.cpp", 852 "src/core/NEON/kernels/NEFillBorderKernel.cpp", 853 "src/runtime/NEON/INEOperator.cpp", 854 "src/runtime/NEON/INESimpleFunction.cpp", 855 "src/runtime/NEON/INESimpleFunctionNoBorder.cpp" 856 ], 857 "operators": { 858 "Activation": { 859 "files": { 860 "common": [ 861 "src/cpu/operators/CpuActivation.cpp", 862 "src/cpu/kernels/CpuActivationKernel.cpp", 863 "src/runtime/NEON/functions/NEActivationLayer.cpp" 864 ], 865 "neon": { 866 "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ], 867 "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ], 868 "qasymm8": [ 869 "src/cpu/kernels/activation/generic/neon/qasymm8.cpp", 870 "src/cpu/kernels/activation/generic/neon/lut.cpp" 871 ], 872 "qasymm8_signed": [ 873 "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp", 874 "src/cpu/kernels/activation/generic/neon/lut.cpp" 875 ], 876 "qsymm16": [ "src/cpu/kernels/activation/generic/neon/qsymm16.cpp" ] 877 }, 878 "sve": { 879 "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ], 880 "fp32": [ "src/cpu/kernels/activation/generic/sve/fp32.cpp" ], 881 "qasymm8": ["src/cpu/kernels/activation/generic/sve/lut.cpp"] 882 }, 883 "sve2":{ 884 "qasymm8": [ "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp" ], 885 "qasymm8_signed": [ "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp" ], 886 "qsymm16": [ "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp" ] 887 } 888 } 889 }, 890 "ArgMinMax": { 891 "deps": [ "Reduction" ], 892 "files": { 893 "common": [ "src/runtime/NEON/functions/NEArgMinMaxLayer.cpp" ] 894 } 895 }, 896 "Add": { 897 "files": { 898 "common": [ 899 "src/cpu/operators/CpuAdd.cpp", 900 "src/cpu/kernels/CpuAddKernel.cpp", 901 "src/runtime/NEON/functions/NEArithmeticAddition.cpp" 902 ], 903 "neon": { 904 "common": ["src/cpu/kernels/add/generic/neon/impl.cpp"], 905 "fp32":["src/cpu/kernels/add/generic/neon/fp32.cpp"], 906 "fp16":["src/cpu/kernels/add/generic/neon/fp16.cpp"], 907 "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"], 908 "qasymm8": ["src/cpu/kernels/add/generic/neon/qasymm8.cpp"], 909 "qasymm8_signed": ["src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp"], 910 "qsymm16": ["src/cpu/kernels/add/generic/neon/qsymm16.cpp"] 911 }, 912 "sve": { 913 "common": [ "src/cpu/kernels/add/generic/sve/impl.cpp" ], 914 "integer":["src/cpu/kernels/add/generic/sve/integer.cpp"], 915 "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"], 916 "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"] 917 }, 918 "sve2": { 919 "qasymm8": [ "src/cpu/kernels/add/generic/sve2/qasymm8.cpp" ], 920 "qasymm8_signed": [ "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp" ], 921 "qsymm16": [ "src/cpu/kernels/add/generic/sve2/qsymm16.cpp" ] 922 } 923 } 924 }, 925 "AddMulAdd": { 926 "files": { 927 "common": [ 928 "src/cpu/operators/CpuAddMulAdd.cpp", 929 "src/cpu/kernels/CpuAddMulAddKernel.cpp", 930 "src/runtime/NEON/functions/NEAddMulAdd.cpp" 931 ], 932 "neon": { 933 "fp32":["src/cpu/kernels/addmuladd/generic/neon/fp32.cpp"], 934 "fp16":["src/cpu/kernels/addmuladd/generic/neon/fp16.cpp"], 935 "qasymm8": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp"], 936 "qasymm8_signed": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8_signed.cpp"] 937 } 938 } 939 }, 940 "BatchNormalize": { 941 "files": { 942 "common": [ 943 "src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp", 944 "src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp", 945 "src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp", 946 "src/runtime/NEON/functions/NEFuseBatchNormalization.cpp" 947 ], 948 "neon": { 949 "common": [ 950 "src/cpu/kernels/fuse_batch_normalization/generic/impl.cpp", 951 "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp", 952 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp" 953 ], 954 "fp16": [ 955 "src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp", 956 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp", 957 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp" 958 ], 959 "fp32": [ 960 "src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp", 961 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp", 962 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp" 963 ] 964 }, 965 "sve": { 966 "fp16": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp" ], 967 "fp32": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp" ] 968 } 969 } 970 }, 971 "BatchToSpace": { 972 "files": { 973 "common": [ 974 "src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp", 975 "src/runtime/NEON/functions/NEBatchToSpaceLayer.cpp" 976 ] 977 } 978 }, 979 "BitwiseAnd": { 980 "files": { 981 "common": [ 982 "src/core/NEON/kernels/NEBitwiseAndKernel.cpp", 983 "src/runtime/NEON/functions/NEBitwiseAnd.cpp" 984 ] 985 } 986 }, 987 "BitwiseNot": { 988 "files": { 989 "common": [ 990 "src/core/NEON/kernels/NEBitwiseNotKernel.cpp", 991 "src/runtime/NEON/functions/NEBitwiseNot.cpp" 992 ] 993 } 994 }, 995 "BitwiseOr": { 996 "files": { 997 "common": [ 998 "src/core/NEON/kernels/NEBitwiseOrKernel.cpp", 999 "src/runtime/NEON/functions/NEBitwiseOr.cpp" 1000 ] 1001 } 1002 }, 1003 "BitwiseXor": { 1004 "files": { 1005 "common": [ 1006 "src/core/NEON/kernels/NEBitwiseXorKernel.cpp", 1007 "src/runtime/NEON/functions/NEBitwiseXor.cpp" 1008 ] 1009 } 1010 }, 1011 "BoundingBoxTransform": { 1012 "files": { 1013 "common": [ 1014 "src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp", 1015 "src/runtime/NEON/functions/NEBoundingBoxTransform.cpp" 1016 ], 1017 "neon":{ 1018 "common":["src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp"], 1019 "fp32":["src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp"], 1020 "fp16":["src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp"], 1021 "qsymm16":["src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp"] 1022 } 1023 } 1024 }, 1025 "Cast": { 1026 "files": { 1027 "common": [ 1028 "src/cpu/operators/CpuCast.cpp", 1029 "src/cpu/kernels/CpuCastKernel.cpp", 1030 "src/runtime/NEON/functions/NECast.cpp", 1031 "src/cpu/kernels/cast/generic/neon/bfloat16.cpp" 1032 ], 1033 "neon":{ 1034 "fp16":["src/cpu/kernels/cast/generic/neon/fp16.cpp"] 1035 } 1036 } 1037 }, 1038 "ChannelShuffle": { 1039 "files": { 1040 "common": [ 1041 "src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp", 1042 "src/runtime/NEON/functions/NEChannelShuffleLayer.cpp" 1043 ] 1044 } 1045 }, 1046 "Concatenate": { 1047 "files": { 1048 "common": [ 1049 "src/cpu/operators/CpuConcatenate.cpp", 1050 "src/cpu/kernels/CpuConcatenateWidthKernel.cpp", 1051 "src/cpu/kernels/CpuConcatenateBatchKernel.cpp", 1052 "src/cpu/kernels/CpuConcatenateDepthKernel.cpp", 1053 "src/cpu/kernels/CpuConcatenateHeightKernel.cpp", 1054 "src/runtime/NEON/functions/NEConcatenateLayer.cpp" 1055 ] 1056 } 1057 }, 1058 "Conv2d": { 1059 "deps": [ 1060 "Activation", 1061 "ElementwiseBinary", 1062 "FFT2D", 1063 "Gemm", 1064 "Mul", 1065 "Pad", 1066 "Permute", 1067 "Reshape", 1068 "Reverse", 1069 "Slice" 1070 ], 1071 "files": { 1072 "common": [ 1073 "src/cpu/operators/CpuConv2d.cpp", 1074 "src/cpu/operators/CpuDirectConv2d.cpp", 1075 "src/cpu/operators/CpuGemmDirectConv2d.cpp", 1076 "src/cpu/operators/CpuGemmConv2d.cpp", 1077 "src/cpu/operators/CpuWinogradConv2d.cpp", 1078 "src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp", 1079 "src/cpu/kernels/CpuDirectConv2dKernel.cpp", 1080 "src/cpu/kernels/CpuDirectConv2dOutputStageKernel.cpp", 1081 "src/cpu/kernels/CpuWinogradConv2dKernel.cpp", 1082 "src/cpu/kernels/CpuCol2ImKernel.cpp", 1083 "src/cpu/kernels/CpuIm2ColKernel.cpp", 1084 "src/cpu/kernels/CpuWeightsReshapeKernel.cpp", 1085 "src/runtime/NEON/functions/NEConvolutionLayer.cpp", 1086 "src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp", 1087 "src/runtime/NEON/functions/NEFFTConvolutionLayer.cpp", 1088 "src/runtime/NEON/functions/NEGEMMConv2d.cpp", 1089 "src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp", 1090 "src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp" 1091 ], 1092 "neon": { 1093 "common": [ 1094 "src/core/NEON/kernels/convolution/common/padding.cpp", 1095 "src/core/NEON/kernels/convolution/common/qasymm8.cpp", 1096 "src/core/NEON/kernels/convolution/common/qsymm8.cpp", 1097 "src/core/NEON/kernels/convolution/common/utils.cpp", 1098 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp16.cpp", 1099 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp32.cpp", 1100 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp16.cpp", 1101 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp32.cpp", 1102 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp16.cpp", 1103 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp32.cpp", 1104 "src/core/NEON/kernels/convolution/winograd/winograd_fp16.cpp", 1105 "src/core/NEON/kernels/convolution/winograd/winograd_fp32.cpp", 1106 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp", 1107 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp", 1108 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp", 1109 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp", 1110 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp", 1111 "src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp", 1112 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp", 1113 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp", 1114 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp", 1115 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp", 1116 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp", 1117 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp", 1118 "src/core/NEON/kernels/convolution/winograd/weight_transforms/a64_fp16_4x4_3x3.cpp", 1119 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_3x3.cpp", 1120 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_5x5.cpp", 1121 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_4x4_3x3.cpp", 1122 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x2_1x7.cpp", 1123 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp", 1124 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x6_1x3.cpp", 1125 "src/cpu/kernels/directconv2d/nhwc/neon/impl.cpp", 1126 "src/cpu/kernels/directconv2d/nchw/all.cpp" 1127 ], 1128 "fp32": [ 1129 "src/cpu/kernels/directconv2d/nhwc/neon/fp32.cpp" 1130 ] 1131 }, 1132 "sve": { 1133 "common": [ 1134 "src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp", 1135 "src/core/NEON/kernels/convolution/winograd/input_transforms/sme_fp32_mla_6x6.cpp", 1136 "src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp" 1137 ] 1138 } 1139 } 1140 }, 1141 "Copy": { 1142 "files": { 1143 "common": [ 1144 "src/cpu/operators/CpuCopy.cpp", 1145 "src/cpu/kernels/CpuCopyKernel.cpp", 1146 "src/runtime/NEON/functions/NECopy.cpp" 1147 ] 1148 } 1149 }, 1150 "CropResize": { 1151 "deps": [ "Scale" ], 1152 "files": { 1153 "common": [ 1154 "src/core/NEON/kernels/NECropKernel.cpp", 1155 "src/runtime/NEON/functions/NECropResize.cpp" 1156 ], 1157 "neon": { 1158 "common": [ "src/cpu/kernels/crop/generic/neon/impl.cpp" ], 1159 "fp32": [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ], 1160 "fp16": [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ], 1161 "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp" ] 1162 } 1163 } 1164 }, 1165 "Deconv2d": { 1166 "deps": [ "Conv2d", "Reverse", "Transpose"], 1167 "files": { 1168 "common": [ 1169 "src/runtime/NEON/functions/NEDeconvolutionLayer.cpp" 1170 ] 1171 } 1172 }, 1173 "DepthConvert": { 1174 "deps": [ "Cast"], 1175 "files": { 1176 "common": [ 1177 "src/runtime/NEON/functions/NEDepthConvertLayer.cpp" 1178 ] 1179 } 1180 }, 1181 "DepthToSpace": { 1182 "files": { 1183 "common": [ 1184 "src/core/NEON/kernels/NEDepthToSpaceLayerKernel.cpp", 1185 "src/runtime/NEON/functions/NEDepthToSpaceLayer.cpp" 1186 ] 1187 } 1188 }, 1189 "DepthwiseConv2d": { 1190 "deps": [ "Activation", "Permute" ], 1191 "files": { 1192 "common": [ 1193 "src/cpu/operators/CpuDepthwiseConv2d.cpp", 1194 "src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp", 1195 "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.cpp", 1196 "src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp", 1197 "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp" 1198 ], 1199 "neon": { 1200 "common": [ 1201 "src/core/NEON/kernels/convolution/common/padding.cpp", 1202 "src/core/NEON/kernels/convolution/common/qasymm8.cpp", 1203 "src/core/NEON/kernels/convolution/common/qsymm8.cpp", 1204 "src/core/NEON/kernels/convolution/common/utils.cpp", 1205 "src/core/NEON/kernels/arm_conv/addressing.cpp", 1206 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_strategies_common.cpp", 1207 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp", 1208 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp", 1209 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp", 1210 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp", 1211 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp", 1212 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1213 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp", 1214 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp", 1215 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1216 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", 1217 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", 1218 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1219 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", 1220 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", 1221 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1222 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1223 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp", 1224 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1225 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1226 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", 1227 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", 1228 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", 1229 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", 1230 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", 1231 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", 1232 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1233 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1234 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1235 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1236 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1237 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", 1238 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", 1239 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", 1240 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", 1241 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", 1242 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", 1243 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1244 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1245 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1246 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1247 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1248 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1249 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1250 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1251 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1252 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1253 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1254 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1255 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1256 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1257 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1258 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1259 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1260 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1261 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1262 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1263 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1264 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp", 1265 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp", 1266 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp", 1267 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp", 1268 "src/cpu/kernels/depthwiseconv2d/generic/neon/impl.cpp" 1269 ], 1270 "fp16":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp16.cpp"], 1271 "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"], 1272 "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"], 1273 "qasymm8_signed":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8_signed.cpp"] 1274 }, 1275 "sve": { 1276 "common": [ 1277 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_8b_mla.cpp", 1278 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp", 1279 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp", 1280 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1281 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1282 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", 1283 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", 1284 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", 1285 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", 1286 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", 1287 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", 1288 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1289 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1290 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1291 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1292 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1293 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst_strided/generic.cpp", 1294 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", 1295 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", 1296 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", 1297 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", 1298 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", 1299 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", 1300 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1301 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1302 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp", 1303 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp", 1304 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp", 1305 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp", 1306 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1307 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1308 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1309 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1310 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", 1311 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", 1312 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1313 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp", 1314 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1315 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1316 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1317 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp", 1318 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp", 1319 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp", 1320 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp", 1321 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp", 1322 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp", 1323 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp", 1324 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp", 1325 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp", 1326 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp", 1327 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp", 1328 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp", 1329 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp", 1330 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s1_4rows_mla_za/generic.cpp", 1331 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp", 1332 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp", 1333 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s2_4rows_mla_za/generic.cpp", 1334 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp", 1335 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s2_4rows_dot_za/generic.cpp", 1336 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp", 1337 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp", 1338 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s1_4rows_dot_za/generic.cpp", 1339 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_2rows_dot_za/generic.cpp", 1340 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp", 1341 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s1_4rows_dot_za/generic.cpp", 1342 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s2_4rows_dot_za/generic.cpp", 1343 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s1_4rows_dot_za/generic.cpp", 1344 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_2rows_dot_za/generic.cpp", 1345 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_4rows_dot_za/generic.cpp", 1346 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s1_4rows_dot_za/generic.cpp", 1347 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s2_4rows_dot_za/generic.cpp", 1348 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp", 1349 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_2rows_dot_za/generic.cpp", 1350 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_4rows_dot_za/generic.cpp", 1351 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s1_4rows_dot_za/generic.cpp", 1352 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp" 1353 ] 1354 } 1355 } 1356 }, 1357 "Dequantize": { 1358 "files": { 1359 "common": [ 1360 "src/cpu/operators/CpuDequantize.cpp", 1361 "src/cpu/kernels/CpuDequantizeKernel.cpp", 1362 "src/runtime/NEON/functions/NEDequantizationLayer.cpp" 1363 ] 1364 } 1365 }, 1366 "DetectionPostProcess": { 1367 "deps": [ "Dequantize" ], 1368 "files": { 1369 "common" : [ "src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp" ] 1370 } 1371 }, 1372 "Conv3d": { 1373 "deps": [ 1374 "Activation" 1375 ], 1376 "files": { 1377 "common": [ 1378 "src/cpu/operators/CpuDirectConv3d.cpp", 1379 "src/cpu/kernels/CpuDirectConv3dKernel.cpp", 1380 "src/runtime/NEON/functions/NEConv3D.cpp" 1381 ] 1382 } 1383 }, 1384 "ElementwiseBinary": { 1385 "files": { 1386 "common": [ 1387 "src/cpu/operators/CpuElementwise.cpp", 1388 "src/cpu/kernels/CpuElementwiseKernel.cpp", 1389 "src/runtime/NEON/functions/NEElementwiseOperations.cpp" 1390 ], 1391 "neon":{ 1392 "fp32": ["src/cpu/kernels/elementwise_binary/generic/neon/fp32.cpp"], 1393 "fp16": ["src/cpu/kernels/elementwise_binary/generic/neon/fp16.cpp"], 1394 "integer": ["src/cpu/kernels/elementwise_binary/generic/neon/integer.cpp"], 1395 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8.cpp"], 1396 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp"] 1397 }, 1398 "sve": { 1399 "common": ["src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp" ], 1400 "integer": ["src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp"], 1401 "fp32": ["src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp"], 1402 "fp16": ["src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp"] 1403 1404 }, 1405 "sve2":{ 1406 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8.cpp"], 1407 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8_signed.cpp"] 1408 } 1409 } 1410 }, 1411 "ElementwiseUnary":{ 1412 "files": { 1413 "common": [ 1414 "src/cpu/operators/CpuElementwiseUnary.cpp", 1415 "src/cpu/kernels/CpuElementwiseUnaryKernel.cpp", 1416 "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp" 1417 ], 1418 "neon": { 1419 "common":["src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp"], 1420 "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"], 1421 "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"], 1422 "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"] 1423 }, 1424 "sve": { 1425 "common": ["src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp" ], 1426 "integer": ["src/cpu/kernels/elementwise_unary/generic/sve/integer.cpp"], 1427 "fp32": ["src/cpu/kernels/elementwise_unary/generic/sve/fp32.cpp"], 1428 "fp16": ["src/cpu/kernels/elementwise_unary/generic/sve/fp16.cpp"] 1429 } 1430 } 1431 }, 1432 "FFT1D": { 1433 "deps": [ "Reduction" ], 1434 "files": { 1435 "common": [ 1436 "src/core/NEON/kernels/NEFFTDigitReverseKernel.cpp", 1437 "src/core/NEON/kernels/NEFFTRadixStageKernel.cpp", 1438 "src/core/NEON/kernels/NEFFTScaleKernel.cpp", 1439 "src/runtime/NEON/functions/NEFFT1D.cpp" 1440 ] 1441 } 1442 }, 1443 "FFT2D": { 1444 "deps": [ "FFT1D" ], 1445 "files": { 1446 "common": [ 1447 "src/runtime/NEON/functions/NEFFT2D.cpp" 1448 ] 1449 } 1450 }, 1451 "Fill": { 1452 "files": { 1453 "common": [ 1454 "src/cpu/operators/CpuFill.cpp", 1455 "src/cpu/kernels/CpuFillKernel.cpp", 1456 "src/runtime/NEON/functions/NEFill.cpp" 1457 ] 1458 } 1459 }, 1460 "Flatten": { 1461 "deps: ": [ "Reshape" ], 1462 "files": { 1463 "common": [ 1464 "src/cpu/operators/CpuFlatten.cpp", 1465 "src/runtime/NEON/functions/NEFlattenLayer.cpp" 1466 ] 1467 } 1468 }, 1469 "Floor": { 1470 "files": { 1471 "common": [ 1472 "src/cpu/operators/CpuFloor.cpp", 1473 "src/cpu/kernels/CpuFloorKernel.cpp", 1474 "src/runtime/NEON/functions/NEFloor.cpp" 1475 ], 1476 "neon": { 1477 "fp32": [ "src/cpu/kernels/floor/neon/fp32.cpp" ], 1478 "fp16": [ "src/cpu/kernels/floor/neon/fp16.cpp" ] 1479 } 1480 } 1481 }, 1482 "FullyConnected": { 1483 "deps": [ "Flatten", "Gemm", "Transpose"], 1484 "files": { 1485 "common": [ 1486 "src/cpu/kernels/CpuConvertFullyConnectedWeightsKernel.cpp", 1487 "src/cpu/operators/CpuConvertFullyConnectedWeights.cpp", 1488 "src/cpu/operators/CpuFullyConnected.cpp", 1489 "src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp", 1490 "src/runtime/NEON/functions/NEFullyConnectedLayer.cpp" 1491 ] 1492 } 1493 }, 1494 "Gather": { 1495 "files": { 1496 "common": [ 1497 "src/core/NEON/kernels/NEGatherKernel.cpp", 1498 "src/runtime/NEON/functions/NEGather.cpp" 1499 ] 1500 } 1501 }, 1502 "Gemm": { 1503 "deps": [ "Quantize", "Add"], 1504 "files": { 1505 "common": [ 1506 "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp", 1507 "src/cpu/kernels/CpuGemmMatrixAdditionKernel.cpp", 1508 "src/cpu/kernels/CpuGemmMatrixMultiplyKernel.cpp", 1509 "src/cpu/kernels/CpuGemmTranspose1xWKernel.cpp", 1510 "src/cpu/kernels/CpuGemmInterleave4x4Kernel.cpp", 1511 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp", 1512 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp", 1513 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp", 1514 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp", 1515 "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp", 1516 "src/cpu/kernels/CpuGemmLowpMatrixReductionKernel.cpp", 1517 "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.cpp", 1518 "src/cpu/kernels/CpuGemmLowpOffsetContributionKernel.cpp", 1519 "src/cpu/operators/CpuGemm.cpp", 1520 "src/cpu/operators/CpuGemmLowpOutputStage.cpp", 1521 "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp", 1522 "src/runtime/NEON/functions/NEGEMM.cpp", 1523 "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp", 1524 "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp" 1525 ], 1526 "neon": { 1527 "common": [ 1528 "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp", 1529 "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp", 1530 "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp", 1531 "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp", 1532 "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp", 1533 "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp", 1534 "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp", 1535 "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp", 1536 "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp", 1537 "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp", 1538 "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp", 1539 "src/core/NEON/kernels/arm_gemm/mergeresults.cpp", 1540 "src/core/NEON/kernels/arm_gemm/misc.cpp", 1541 "src/core/NEON/kernels/arm_gemm/quantized.cpp", 1542 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp", 1543 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp", 1544 "src/core/NEON/kernels/arm_gemm/transform.cpp", 1545 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp", 1546 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp", 1547 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp", 1548 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp", 1549 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp", 1550 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp", 1551 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp", 1552 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp", 1553 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp", 1554 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp", 1555 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp", 1556 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp", 1557 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp", 1558 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp", 1559 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp", 1560 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp", 1561 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp", 1562 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp", 1563 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp", 1564 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp", 1565 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp", 1566 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp", 1567 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp", 1568 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp", 1569 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp", 1570 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp", 1571 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp", 1572 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp", 1573 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp", 1574 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp", 1575 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp", 1576 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp", 1577 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp", 1578 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp", 1579 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp", 1580 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp", 1581 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp", 1582 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp", 1583 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp", 1584 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp", 1585 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp", 1586 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp", 1587 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp", 1588 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp", 1589 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp", 1590 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp", 1591 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp", 1592 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp", 1593 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp", 1594 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp", 1595 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp", 1596 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp", 1597 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp", 1598 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp", 1599 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp", 1600 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp", 1601 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp", 1602 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp", 1603 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp", 1604 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp", 1605 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp", 1606 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp", 1607 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp", 1608 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp", 1609 "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp", 1610 "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp" 1611 ], 1612 "fp32":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp", 1613 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp"], 1614 "fp16":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp", 1615 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp"], 1616 "estate32": [ 1617 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp", 1618 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp", 1619 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp" 1620 ], 1621 "estate64": [ 1622 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp" 1623 ], 1624 "experimental_fixed_format_kernels": [ 1625 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp", 1626 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp", 1627 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp", 1628 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp", 1629 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp", 1630 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp", 1631 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp", 1632 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp" 1633 ] 1634 }, 1635 "sve": { 1636 "common": [ 1637 "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp", 1638 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp", 1639 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp", 1640 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp", 1641 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp", 1642 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp", 1643 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp", 1644 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp", 1645 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp", 1646 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp", 1647 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp", 1648 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp", 1649 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp", 1650 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp", 1651 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp", 1652 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp", 1653 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp", 1654 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp", 1655 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp", 1656 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp", 1657 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp", 1658 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp", 1659 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp", 1660 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp", 1661 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp", 1662 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp", 1663 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp", 1664 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp", 1665 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp", 1666 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp", 1667 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp", 1668 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp", 1669 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp", 1670 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp", 1671 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp", 1672 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp", 1673 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp", 1674 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp", 1675 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp", 1676 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp", 1677 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp", 1678 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp", 1679 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp", 1680 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp", 1681 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp", 1682 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp", 1683 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp", 1684 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp", 1685 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp", 1686 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL/generic.cpp", 1687 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp", 1688 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp", 1689 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp", 1690 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp", 1691 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp", 1692 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp", 1693 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_fp32_mla_8x1VL/generic.cpp", 1694 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_s8s32_dot_8x1VL/generic.cpp", 1695 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_u8u32_dot_8x1VL/generic.cpp", 1696 "src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp", 1697 "src/core/NEON/kernels/arm_gemm/transform-sve.cpp", 1698 "src/core/NEON/kernels/arm_gemm/misc-sve.cpp" 1699 ], 1700 "experimental_fixed_format_kernels": [ 1701 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp", 1702 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp", 1703 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp", 1704 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp", 1705 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp", 1706 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp", 1707 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp", 1708 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp", 1709 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp", 1710 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp", 1711 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp" 1712 ] 1713 } 1714 } 1715 }, 1716 "GenerateProposals": { 1717 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ], 1718 "files": { 1719 "common": [ 1720 "src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp", 1721 "src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp" 1722 ], 1723 "neon":{ 1724 "common":["src/cpu/kernels/genproposals/generic/neon/impl.cpp"], 1725 "fp16":["src/cpu/kernels/genproposals/generic/neon/fp16.cpp"], 1726 "fp32":["src/cpu/kernels/genproposals/generic/neon/fp32.cpp"], 1727 "qsymm16":["src/cpu/kernels/genproposals/generic/neon/qsymm16.cpp"] 1728 } 1729 } 1730 }, 1731 "InstanceNormalize": { 1732 "deps": [ "Permute", "Reduction" ], 1733 "files": { 1734 "common": [ 1735 "src/core/NEON/kernels/NEInstanceNormalizationLayerKernel.cpp", 1736 "src/runtime/NEON/functions/NEInstanceNormalizationLayer.cpp" 1737 ], 1738 "neon":{ 1739 "common":["src/cpu/kernels/instancenorm/generic/neon/impl.cpp"], 1740 "fp16":["src/cpu/kernels/instancenorm/generic/neon/fp16.cpp"], 1741 "fp32":["src/cpu/kernels/instancenorm/generic/neon/fp32.cpp"] 1742 } 1743 } 1744 }, 1745 "L2Normalize": { 1746 "deps": [ "Reduction" ], 1747 "files": { 1748 "common": [ 1749 "src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp", 1750 "src/runtime/NEON/functions/NEL2NormalizeLayer.cpp" 1751 ], 1752 "neon":{ 1753 "common":["src/cpu/kernels/l2normlayer/generic/neon/impl.cpp"], 1754 "fp32":["src/cpu/kernels/l2normlayer/generic/neon/fp32.cpp"], 1755 "fp16":["src/cpu/kernels/l2normlayer/generic/neon/fp16.cpp"] 1756 } 1757 } 1758 }, 1759 "Logical": { 1760 "files": { 1761 "common": [ 1762 "src/core/NEON/kernels/NELogicalKernel.cpp", 1763 "src/runtime/NEON/functions/NELogical.cpp" 1764 ] 1765 } 1766 }, 1767 "LSTM": { 1768 "deps": [ 1769 "Activation", 1770 "Concatenate", 1771 "Copy", 1772 "Dequantize", 1773 "ElementwiseBinary", 1774 "Fill", 1775 "FullyConnected", 1776 "Gemm", 1777 "MeanStdDevNormalize", 1778 "Mul", 1779 "Quantize", 1780 "Slice", 1781 "Transpose" 1782 ], 1783 "files": { 1784 "common": [ 1785 "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp", 1786 "src/runtime/NEON/functions/NELSTMLayer.cpp", 1787 "src/runtime/NEON/functions/NELSTMLayerQuantized.cpp", 1788 "src/runtime/NEON/functions/NEQLSTMLayer.cpp" 1789 ] 1790 } 1791 }, 1792 "MaxUnpool2d": { 1793 "deps": [ "Fill" ], 1794 "files": { 1795 "common": [ 1796 "src/cpu/kernels/CpuMaxUnpoolingLayerKernel.cpp", 1797 "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp", 1798 "src/cpu/operators/CpuMaxUnpooling.cpp" 1799 ], 1800 "neon":{ 1801 "fp32":["src/cpu/kernels/maxunpool/generic/neon/fp32.cpp"], 1802 "fp16":["src/cpu/kernels/maxunpool/generic/neon/fp16.cpp"], 1803 "qasymm8":["src/cpu/kernels/maxunpool/generic/neon/qasymm8.cpp"], 1804 "qasymm8_signed":[ "src/cpu/kernels/maxunpool/generic/neon/qasymm8_signed.cpp"], 1805 "common":["src/cpu/kernels/maxunpool/generic/neon/impl.cpp"] 1806 } 1807 } 1808 }, 1809 "Mean": { 1810 "deps" : [ "Reduction" ], 1811 "files": { 1812 "common": [ "src/runtime/NEON/functions/NEReduceMean.cpp" ] 1813 } 1814 }, 1815 "MeanStdDevNormalize": { 1816 "files": { 1817 "common": [ 1818 "src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp", 1819 "src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp" 1820 ], 1821 "neon":{ 1822 "common":["src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp"], 1823 "fp32":["src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp"], 1824 "fp16":["src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp"], 1825 "qasymm8":["src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp"] 1826 } 1827 } 1828 }, 1829 "Mul": { 1830 "files": { 1831 "common": [ 1832 "src/cpu/operators/CpuMul.cpp", 1833 "src/cpu/kernels/CpuMulKernel.cpp", 1834 "src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp" 1835 ] 1836 } 1837 }, 1838 "Normalize": { 1839 "deps": [ "Mul" ], 1840 "files": { 1841 "common": [ 1842 "src/core/NEON/kernels/NENormalizationLayerKernel.cpp", 1843 "src/runtime/NEON/functions/NENormalizationLayer.cpp" 1844 ] 1845 } 1846 }, 1847 "Pad": { 1848 "deps": [ "Concatenate", "Copy", "StridedSlice" ], 1849 "files": { 1850 "common": [ 1851 "src/core/NEON/kernels/NEPadLayerKernel.cpp", 1852 "src/runtime/NEON/functions/NEPadLayer.cpp" 1853 ] 1854 } 1855 }, 1856 "Permute": { 1857 "files": { 1858 "common": [ 1859 "src/cpu/operators/CpuPermute.cpp", 1860 "src/cpu/kernels/CpuPermuteKernel.cpp", 1861 "src/runtime/NEON/functions/NEPermute.cpp" 1862 ] 1863 } 1864 }, 1865 "Pool2d": { 1866 "files": { 1867 "common": [ 1868 "src/cpu/operators/CpuPool2d.cpp", 1869 "src/cpu/kernels/CpuPool2dKernel.cpp", 1870 "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp", 1871 "src/runtime/NEON/functions/NEPoolingLayer.cpp" 1872 ], 1873 "neon": { 1874 "common": [ 1875 "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp", 1876 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp", 1877 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp", 1878 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp", 1879 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp", 1880 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp", 1881 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp", 1882 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1883 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp", 1884 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1885 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp", 1886 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1887 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp", 1888 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1889 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp", 1890 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp", 1891 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1892 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp", 1893 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp", 1894 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp", 1895 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp", 1896 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1897 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp", 1898 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp", 1899 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp" 1900 ], 1901 "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ], 1902 "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ], 1903 "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ], 1904 "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ], 1905 "qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"] 1906 }, 1907 "sve": { 1908 "common": [ 1909 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1910 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp", 1911 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1912 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp", 1913 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1914 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_generic_depthfirst/generic.cpp", 1915 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1916 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_generic_depthfirst/generic.cpp", 1917 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_avg_generic_depthfirst/generic.cpp", 1918 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1919 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp", 1920 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp", 1921 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp", 1922 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp", 1923 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1924 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp", 1925 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp", 1926 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp", 1927 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1928 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp", 1929 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1930 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_generic_depthfirst/generic.cpp", 1931 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp", 1932 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_generic_depthfirst/generic.cpp", 1933 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1934 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp", 1935 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_avg_generic_depthfirst/generic.cpp", 1936 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1937 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_generic_depthfirst/generic.cpp", 1938 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_avg_generic_depthfirst/generic.cpp", 1939 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp", 1940 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_avg_generic_depthfirst/generic.cpp", 1941 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp", 1942 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_generic_depthfirst/generic.cpp", 1943 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_avg_generic_depthfirst/generic.cpp", 1944 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp" 1945 ] 1946 } 1947 } 1948 }, 1949 "Pool3d": { 1950 "files": { 1951 "common": [ 1952 "src/cpu/operators/CpuPool3d.cpp", 1953 "src/cpu/kernels/CpuPool3dKernel.cpp", 1954 "src/runtime/NEON/functions/NEPooling3dLayer.cpp" 1955 ], 1956 "neon": { 1957 "common": [ "src/cpu/kernels/pool3d/neon/impl.cpp" ], 1958 "fp16": [ "src/cpu/kernels/pool3d/neon/fp16.cpp" ], 1959 "fp32": [ "src/cpu/kernels/pool3d/neon/fp32.cpp" ], 1960 "qasymm8": [ "src/cpu/kernels/pool3d/neon/qasymm8.cpp" ], 1961 "qasymm8_signed": [ "src/cpu/kernels/pool3d/neon/qasymm8_signed.cpp" ] 1962 } 1963 } 1964 }, 1965 "PRelu": { 1966 "deps": [ "ElementwiseBinary" ], 1967 "files": { 1968 "common": [ 1969 "src/runtime/NEON/functions/NEPReluLayer.cpp" 1970 ] 1971 } 1972 }, 1973 "PriorBox": { 1974 "files": { 1975 "common": [ 1976 "src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp", 1977 "src/runtime/NEON/functions/NEPriorBoxLayer.cpp" 1978 ] 1979 } 1980 }, 1981 "Quantize": { 1982 "files": { 1983 "common": [ 1984 "src/cpu/operators/CpuQuantize.cpp", 1985 "src/cpu/kernels/CpuQuantizeKernel.cpp", 1986 "src/runtime/NEON/functions/NEQuantizationLayer.cpp" 1987 ] 1988 } 1989 }, 1990 "Range": { 1991 "files": { 1992 "common": [ 1993 "src/core/NEON/kernels/NERangeKernel.cpp", 1994 "src/runtime/NEON/functions/NERange.cpp" 1995 ], 1996 "neon": { 1997 "common": [ "src/cpu/kernels/range/generic/neon/impl.cpp" ], 1998 "fp32": [ "src/cpu/kernels/range/generic/neon/fp32.cpp" ], 1999 "fp16": [ "src/cpu/kernels/range/generic/neon/fp16.cpp" ], 2000 "integer": [ "src/cpu/kernels/range/generic/neon/integer.cpp" ] 2001 } 2002 } 2003 }, 2004 "Reduction":{ 2005 "deps": [ "Reshape" ], 2006 "files": { 2007 "common": [ 2008 "src/core/NEON/kernels/NEReductionOperationKernel.cpp", 2009 "src/runtime/NEON/functions/NEReductionOperation.cpp" 2010 ] 2011 } 2012 }, 2013 "Reorg": { 2014 "files": { 2015 "common": [ 2016 "src/core/NEON/kernels/NEReorgLayerKernel.cpp", 2017 "src/runtime/NEON/functions/NEReorgLayer.cpp" 2018 ] 2019 } 2020 }, 2021 "Reshape": { 2022 "files": { 2023 "common": [ 2024 "src/cpu/operators/CpuReshape.cpp", 2025 "src/cpu/kernels/CpuReshapeKernel.cpp", 2026 "src/runtime/NEON/functions/NEReshapeLayer.cpp" 2027 ] 2028 } 2029 }, 2030 "Reverse": { 2031 "files": { 2032 "common": [ 2033 "src/core/NEON/kernels/NEReverseKernel.cpp", 2034 "src/runtime/NEON/functions/NEReverse.cpp" 2035 ] 2036 } 2037 }, 2038 "RNN": { 2039 "deps": [ "Activation", "Add", "FullyConnected", "Gemm"], 2040 "files": { 2041 "common": [ "src/runtime/NEON/functions/NERNNLayer.cpp" ] 2042 } 2043 }, 2044 "ROIAlign": { 2045 "files": { 2046 "common": [ 2047 "src/core/NEON/kernels/NEROIAlignLayerKernel.cpp", 2048 "src/runtime/NEON/functions/NEROIAlignLayer.cpp" 2049 ], 2050 "neon":{ 2051 "common":["src/cpu/kernels/roialign/generic/neon/impl.cpp"], 2052 "fp32":["src/cpu/kernels/roialign/generic/neon/fp32.cpp"], 2053 "fp16":["src/cpu/kernels/roialign/generic/neon/fp16.cpp"], 2054 "qasymm8":["src/cpu/kernels/roialign/generic/neon/qasymm8.cpp"], 2055 "qasymm8_signed":["src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp"] 2056 } 2057 } 2058 }, 2059 "ROIPool2d": { 2060 "files": { 2061 "common": [ 2062 "src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp", 2063 "src/runtime/NEON/functions/NEROIPoolingLayer.cpp" 2064 ] 2065 } 2066 }, 2067 "Scale": { 2068 "files": { 2069 "common": [ 2070 "src/cpu/operators/CpuScale.cpp", 2071 "src/cpu/kernels/CpuScaleKernel.cpp", 2072 "src/runtime/NEON/functions/NEScale.cpp" 2073 ], 2074 "sve": { 2075 "fp16": [ "src/cpu/kernels/scale/sve/fp16.cpp" ], 2076 "fp32": [ "src/cpu/kernels/scale/sve/fp32.cpp" ], 2077 "integer": [ "src/cpu/kernels/scale/sve/integer.cpp" ], 2078 "qasymm8": [ "src/cpu/kernels/scale/sve/qasymm8.cpp" ], 2079 "qasymm8_signed": [ "src/cpu/kernels/scale/sve/qasymm8_signed.cpp" ] 2080 }, 2081 "neon": { 2082 "fp16": [ "src/cpu/kernels/scale/neon/fp16.cpp" ], 2083 "integer": [ "src/cpu/kernels/scale/neon/integer.cpp" ], 2084 "qasymm8": [ "src/cpu/kernels/scale/neon/qasymm8.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ], 2085 "qasymm8_signed": [ "src/cpu/kernels/scale/neon/qasymm8_signed.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ] 2086 } 2087 } 2088 }, 2089 "Select": { 2090 "files": { 2091 "common": [ 2092 "src/core/NEON/kernels/NESelectKernel.cpp", 2093 "src/runtime/NEON/functions/NESelect.cpp" 2094 ], 2095 "neon": { 2096 "common": [ "src/cpu/kernels/select/generic/neon/impl.cpp" ], 2097 "fp32": [ "src/cpu/kernels/select/generic/neon/fp32.cpp" ], 2098 "fp16": [ "src/cpu/kernels/select/generic/neon/fp16.cpp" ], 2099 "integer": [ "src/cpu/kernels/select/generic/neon/integer.cpp" ] 2100 } 2101 } 2102 }, 2103 "Slice": { 2104 "deps": [ "StridedSlice" ], 2105 "files": { 2106 "common": [ "src/runtime/NEON/functions/NESlice.cpp" ] 2107 } 2108 }, 2109 "Softmax": { 2110 "deps": [ 2111 "Permute" 2112 ], 2113 "files": { 2114 "common": [ 2115 "src/cpu/operators/CpuSoftmax.cpp", 2116 "src/cpu/kernels/CpuSoftmaxKernel.cpp", 2117 "src/runtime/NEON/functions/NESoftmaxLayer.cpp" 2118 ], 2119 "neon":{ 2120 "common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"], 2121 "fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"], 2122 "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"], 2123 "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"], 2124 "qasymm8_signed":["src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp"] 2125 }, 2126 "sve": { 2127 "common": [ "src/cpu/kernels/softmax/generic/sve/impl.cpp" ], 2128 "fp32": ["src/cpu/kernels/softmax/generic/sve/fp32.cpp"], 2129 "fp16": ["src/cpu/kernels/softmax/generic/sve/fp16.cpp"], 2130 "qasymm8": ["src/cpu/kernels/softmax/generic/sve/qasymm8.cpp" ], 2131 "qasymm8_signed": ["src/cpu/kernels/softmax/generic/sve/qasymm8_signed.cpp"] 2132 }, 2133 "sve2":{ 2134 "common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"], 2135 "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"], 2136 "qasymm8_signed":["src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp"] 2137 } 2138 } 2139 }, 2140 "SpaceToBatch": { 2141 "files": { 2142 "common": [ 2143 "src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp", 2144 "src/runtime/NEON/functions/NESpaceToBatchLayer.cpp" 2145 ] 2146 } 2147 }, 2148 "SpaceToDepth": { 2149 "files": { 2150 "common": [ 2151 "src/core/NEON/kernels/NESpaceToDepthLayerKernel.cpp", 2152 "src/runtime/NEON/functions/NESpaceToDepthLayer.cpp" 2153 ] 2154 } 2155 }, 2156 "Split": { 2157 "deps": [ "StridedSlice" ], 2158 "files": { 2159 "common": [ 2160 "src/runtime/NEON/functions/NESplit.cpp" 2161 ] 2162 } 2163 }, 2164 "Stack": { 2165 "files": { 2166 "common": [ 2167 "src/core/NEON/kernels/NEStackLayerKernel.cpp", 2168 "src/runtime/NEON/functions/NEStackLayer.cpp" 2169 ] 2170 } 2171 }, 2172 "StridedSlice": { 2173 "files": { 2174 "common": [ 2175 "src/core/NEON/kernels/NEStridedSliceKernel.cpp", 2176 "src/runtime/NEON/functions/NEStridedSlice.cpp" 2177 ] 2178 } 2179 }, 2180 "Sub": { 2181 "files": { 2182 "common": [ 2183 "src/cpu/operators/CpuSub.cpp", 2184 "src/cpu/kernels/CpuSubKernel.cpp", 2185 "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp", 2186 "src/cpu/kernels/sub/neon/qasymm8.cpp", 2187 "src/cpu/kernels/sub/neon/qasymm8_signed.cpp", 2188 "src/cpu/kernels/sub/neon/qsymm16.cpp" 2189 ] 2190 } 2191 }, 2192 "Tile": { 2193 "files": { 2194 "common": [ 2195 "src/core/NEON/kernels/NETileKernel.cpp", 2196 "src/runtime/NEON/functions/NETile.cpp" 2197 ] 2198 } 2199 }, 2200 "Transpose": { 2201 "files": { 2202 "common": [ 2203 "src/cpu/kernels/CpuTransposeKernel.cpp", 2204 "src/cpu/operators/CpuTranspose.cpp", 2205 "src/runtime/NEON/functions/NETranspose.cpp" 2206 ] 2207 } 2208 }, 2209 "Unstack": { 2210 "deps": [ "StridedSlice" ], 2211 "files": { 2212 "common": [ "src/runtime/NEON/functions/NEUnstack.cpp" ] 2213 } 2214 } 2215 } 2216 }, 2217 "experimental": { 2218 "dynamic_fusion": [ 2219 "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp", 2220 "src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp", 2221 "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp", 2222 "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp", 2223 "src/dynamic_fusion/sketch/attributes/Conv2dAttributes.cpp", 2224 "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp", 2225 "src/dynamic_fusion/sketch/attributes/Pool2dAttributes.cpp", 2226 "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp", 2227 "src/dynamic_fusion/sketch/attributes/SoftmaxAttributes.cpp", 2228 "src/dynamic_fusion/sketch/attributes/ReshapeAttributes.cpp", 2229 "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.cpp", 2230 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp", 2231 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp", 2232 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.cpp", 2233 "src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp", 2234 "src/dynamic_fusion/sketch/gpu/GpuOperatorGroup.cpp", 2235 "src/dynamic_fusion/sketch/gpu/GpuWorkloadContext.cpp", 2236 "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.cpp", 2237 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.cpp", 2238 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentCast.cpp", 2239 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp", 2240 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp", 2241 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentPool2d.cpp", 2242 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp", 2243 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp", 2244 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DNorm.cpp", 2245 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentReshape.cpp", 2246 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp", 2247 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp", 2248 "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp", 2249 "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp", 2250 "src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp", 2251 "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp", 2252 "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp", 2253 "src/dynamic_fusion/sketch/gpu/operators/GpuMul.cpp", 2254 "src/dynamic_fusion/sketch/gpu/operators/GpuReshape.cpp", 2255 "src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp", 2256 "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp", 2257 "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", 2258 "src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp", 2259 "src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp", 2260 "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp", 2261 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp", 2262 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateCast.cpp", 2263 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp", 2264 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp", 2265 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplatePool2d.cpp", 2266 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp", 2267 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DMaxShiftExpSum.cpp", 2268 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DNorm.cpp", 2269 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateReshape.cpp", 2270 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp", 2271 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp", 2272 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp", 2273 "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp" 2274 ] 2275 } 2276} 2277