blob: e833de9fc71cf09b7a6459726a6e34a9bf5f0ffe [file] [log] [blame]
Georgios Pinitasbdcdc392021-04-22 16:42:03 +01001{
Michalis Spyrou20fca522021-06-07 14:23:57 +01002 "common": [
3 "src/common/cpuinfo/CpuInfo.cpp",
4 "src/common/cpuinfo/CpuModel.cpp",
5 "src/common/cpuinfo/CpuIsaInfo.cpp",
6 "src/common/utils/LegacySupport.cpp",
7 "src/common/AllocatorWrapper.cpp",
8 "src/common/ITensorV2.cpp",
Georgios Pinitas41648142021-08-03 08:24:00 +01009 "src/common/TensorPack.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010010 "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",
Mohammed Suhail Munshi7467ba82023-12-05 14:27:31 +000017 "src/core/helpers/LUTManager.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010018 "src/core/IAccessWindow.cpp",
19 "src/core/IKernel.cpp",
20 "src/core/ITensor.cpp",
21 "src/core/ITensorPack.cpp",
22 "src/core/Rounding.cpp",
23 "src/core/Size2D.cpp",
24 "src/core/SubTensorInfo.cpp",
25 "src/core/TensorInfo.cpp",
26 "src/core/Utils.cpp",
27 "src/core/Validate.cpp",
28 "src/core/Version.cpp",
29 "src/core/helpers/SoftmaxHelpers.cpp",
Anitha Rajb566b6e2023-08-23 11:40:06 +010030 "src/core/helpers/Utils.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010031 "src/core/helpers/WindowHelpers.cpp",
Matthew Bentham314d3e22023-06-23 10:53:52 +000032 "src/core/utils/ActivationFunctionUtils.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010033 "src/core/utils/AssemblyUtils.cpp",
Matthew Bentham314d3e22023-06-23 10:53:52 +000034 "src/core/utils/DataTypeUtils.cpp",
35 "src/core/utils/DataLayoutUtils.cpp",
36 "src/core/utils/InterpolationPolicyUtils.cpp",
37 "src/core/utils/FormatUtils.cpp",
Viet-Hoa Doef9da002023-09-27 16:39:05 +010038 "src/core/utils/Math.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010039 "src/core/utils/ScaleUtils.cpp",
Matthew Bentham314d3e22023-06-23 10:53:52 +000040 "src/core/utils/StringUtils.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +010041 "src/core/utils/helpers/fft.cpp",
42 "src/core/utils/helpers/tensor_transform.cpp",
43 "src/core/utils/io/FileHandler.cpp",
44 "src/core/utils/misc/MMappedFile.cpp",
45 "src/core/utils/quantization/AsymmHelpers.cpp",
46 "src/core/CPP/CPPTypes.cpp",
47 "src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp",
48 "src/core/CPP/kernels/CPPNonMaximumSuppressionKernel.cpp",
49 "src/core/CPP/kernels/CPPPermuteKernel.cpp",
50 "src/core/CPP/kernels/CPPTopKVKernel.cpp",
51 "src/core/CPP/kernels/CPPUpsampleKernel.cpp",
52 "src/runtime/Allocator.cpp",
53 "src/runtime/BlobLifetimeManager.cpp",
54 "src/runtime/BlobMemoryPool.cpp",
55 "src/runtime/ISimpleLifetimeManager.cpp",
56 "src/runtime/ITensorAllocator.cpp",
57 "src/runtime/IWeightsManager.cpp",
58 "src/runtime/IScheduler.cpp",
59 "src/runtime/Memory.cpp",
60 "src/runtime/MemoryManagerOnDemand.cpp",
61 "src/runtime/OffsetLifetimeManager.cpp",
62 "src/runtime/OffsetMemoryPool.cpp",
63 "src/runtime/OperatorTensor.cpp",
64 "src/runtime/PoolManager.cpp",
65 "src/runtime/RuntimeContext.cpp",
66 "src/runtime/Scheduler.cpp",
67 "src/runtime/SchedulerFactory.cpp",
68 "src/runtime/SchedulerUtils.cpp",
69 "src/runtime/SubTensor.cpp",
70 "src/runtime/Tensor.cpp",
71 "src/runtime/TensorAllocator.cpp",
72 "src/runtime/Utils.cpp",
73 "src/runtime/CPP/ICPPSimpleFunction.cpp",
74 "src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp",
75 "src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp",
76 "src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp",
77 "src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp",
78 "src/runtime/CPP/functions/CPPPermute.cpp",
79 "src/runtime/CPP/functions/CPPTopKV.cpp",
80 "src/runtime/CPP/functions/CPPUpsample.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +010081 ],
Georgios Pinitasb6af4822021-09-14 12:33:34 +010082 "logging": [
83 "src/core/utils/logging/FilePrinter.cpp",
84 "src/core/utils/logging/Helpers.cpp",
85 "src/core/utils/logging/Logger.cpp",
86 "src/core/utils/logging/LoggerRegistry.cpp"
87 ],
88 "scheduler": {
89 "single": [ "src/runtime/CPP/SingleThreadScheduler.cpp" ],
90 "threads": [ "src/runtime/CPP/CPPScheduler.cpp" ],
91 "omp": [ "src/runtime/OMP/OMPScheduler.cpp"]
92 },
Michalis Spyrou20fca522021-06-07 14:23:57 +010093 "c_api": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +010094 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +010095 "src/c/AclContext.cpp",
Georgios Pinitas06ac6e42021-07-05 08:08:52 +010096 "src/c/AclOperator.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +010097 "src/c/AclQueue.cpp",
98 "src/c/AclTensor.cpp",
99 "src/c/AclTensorPack.cpp",
100 "src/c/AclVersion.cpp"
Georgios Pinitasbdcdc392021-04-22 16:42:03 +0100101 ],
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100102 "cpu": [
103 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100104 "gpu": [
105 "src/c/cl/AclOpenClExt.cpp"
Georgios Pinitas41648142021-08-03 08:24:00 +0100106 ],
107 "operators":
108 [
109 "src/c/operators/AclActivation.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100110 ]
111 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100112 "high_priority": [
113 "Activation",
114 "DepthwiseConv2d",
115 "Conv2d",
116 "Permute",
117 "Pool2d",
SiCong Li83d06742023-05-11 11:54:22 +0100118 "Reshape",
119 "MatMul"
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100120 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100121 "gpu": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100122 "common": [
Viet-Hoa Do500e10b2023-09-12 17:49:38 +0100123 "src/core/CL/CLCommandBuffer.cpp",
124 "src/core/CL/CLCompatCommandBuffer.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100125 "src/core/CL/CLCompileContext.cpp",
Viet-Hoa Do500e10b2023-09-12 17:49:38 +0100126 "src/core/CL/CLMutableCommandBuffer.cpp",
Giorgio Arenaba2dd822021-07-28 16:10:03 +0100127 "src/core/CL/DefaultLWSHeuristics.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100128 "src/core/CL/CLHelpers.cpp",
129 "src/core/CL/CLKernelLibrary.cpp",
130 "src/core/CL/CLUtils.cpp",
131 "src/core/CL/ICLKernel.cpp",
132 "src/core/CL/ICLSimple2DKernel.cpp",
133 "src/core/CL/ICLSimple3DKernel.cpp",
134 "src/core/CL/ICLSimpleKernel.cpp",
135 "src/core/CL/ICLTensor.cpp",
136 "src/core/CL/OpenCL.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100137 "src/gpu/cl/ClContext.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100138 "src/gpu/cl/ClKernelLibrary.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100139 "src/gpu/cl/ClQueue.cpp",
140 "src/gpu/cl/ClTensor.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100141 "src/core/CL/kernels/CLFillBorderKernel.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100142 "src/runtime/CL/CLBufferAllocator.cpp",
143 "src/runtime/CL/CLGEMMHeuristicsHandle.cpp",
144 "src/runtime/CL/CLHelpers.cpp",
145 "src/runtime/CL/CLMemory.cpp",
146 "src/runtime/CL/CLMemoryRegion.cpp",
147 "src/runtime/CL/CLOperator.cpp",
148 "src/runtime/CL/CLRuntimeContext.cpp",
149 "src/runtime/CL/CLScheduler.cpp",
150 "src/runtime/CL/CLSubTensor.cpp",
151 "src/runtime/CL/CLTensor.cpp",
152 "src/runtime/CL/CLTensorAllocator.cpp",
153 "src/runtime/CL/CLTuner.cpp",
154 "src/runtime/CL/ICLSimpleFunction.cpp",
155 "src/runtime/CL/Utils.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100156 "src/runtime/CL/mlgo/HeuristicTree.cpp",
157 "src/runtime/CL/mlgo/MLGOHeuristics.cpp",
158 "src/runtime/CL/mlgo/MLGOParser.cpp",
159 "src/runtime/CL/mlgo/Utils.cpp",
160 "src/runtime/CL/tuners/CLTuningParametersList.cpp"
161 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100162 "operators": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100163 "Activation":{
164 "files": {
165 "common": [
166 "src/gpu/cl/kernels/ClActivationKernel.cpp",
167 "src/gpu/cl/operators/ClActivation.cpp",
168 "src/runtime/CL/functions/CLActivationLayer.cpp"
169 ]
170 }
171 },
172 "ArgMinMax": {
173 "deps": [ "Reshape" ],
174 "files": {
175 "common": [
176 "src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp",
177 "src/runtime/CL/functions/CLArgMinMaxLayer.cpp"
178 ]
179 }
180 },
181 "Add": {
182 "files": {
183 "common": [
184 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
185 "src/gpu/cl/operators/ClAdd.cpp"
186 ]
187 }
188 },
189 "BatchNormalization": {
190 "files": {
191 "common": [
192 "src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp",
193 "src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp",
194 "src/runtime/CL/functions/CLBatchNormalizationLayer.cpp",
195 "src/runtime/CL/functions/CLFuseBatchNormalization.cpp"
196 ]
197 }
198 },
199 "BatchToSpace": {
200 "files": {
201 "common": [
202 "src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp",
203 "src/runtime/CL/functions/CLBatchToSpaceLayer.cpp"
204 ]
205 }
206 },
207 "Bitwise": {
208 "files": {
209 "common": [ "src/core/CL/kernels/CLBitwiseKernel.cpp" ]
210 }
211 },
212 "BitwiseAnd": {
213 "deps": [ "Bitwise" ],
214 "files": {
215 "common": [ "src/runtime/CL/functions/CLBitwiseAnd.cpp" ]
216 }
217 },
218 "BitwiseNot": {
219 "deps": [ "Bitwise" ],
220 "files": {
221 "common": [ "src/runtime/CL/functions/CLBitwiseNot.cpp" ]
222 }
223 },
224 "BitwiseOr": {
225 "deps": [ "Bitwise" ],
226 "files": {
227 "common": [ "src/runtime/CL/functions/CLBitwiseOr.cpp" ]
228 }
229 },
230 "BitwiseXor": {
231 "deps": [ "Bitwise" ],
232 "files": {
233 "common": [ "src/runtime/CL/functions/CLBitwiseXor.cpp" ]
234 }
235 },
236 "BoundingBoxTransform": {
237 "files": {
238 "common": [
239 "src/core/CL/kernels/CLBoundingBoxTransformKernel.cpp",
240 "src/runtime/CL/functions/CLBoundingBoxTransform.cpp"
241 ]
242 }
243 },
244 "Cast": {
245 "files": {
246 "common": [
247 "src/gpu/cl/kernels/ClCastKernel.cpp",
248 "src/gpu/cl/operators/ClCast.cpp",
249 "src/runtime/CL/functions/CLCast.cpp"
250 ]
251 }
252 },
253 "ChannelShuffle": {
254 "files": {
255 "common": [
256 "src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp",
257 "src/runtime/CL/functions/CLChannelShuffleLayer.cpp"
258 ]
259 }
260 },
261 "Comparison": {
262 "files": {
263 "common": [
264 "src/core/CL/kernels/CLComparisonKernel.cpp",
265 "src/runtime/CL/functions/CLComparison.cpp"
266 ]
267 }
268 },
269 "Concatenate": {
270 "files": {
271 "common": [
272 "src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp",
273 "src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp",
274 "src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp",
275 "src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp",
276 "src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp",
277 "src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp",
278 "src/gpu/cl/operators/ClConcatenate.cpp",
279 "src/runtime/CL/functions/CLConcatenateLayer.cpp"
280 ]
281 }
282 },
283 "Conv2d": {
284 "deps": [
285 "Activation",
286 "ElementwiseBinary",
287 "FFT2D",
288 "Gemm",
289 "Mul",
290 "Pad",
291 "Permute",
292 "Reduction",
293 "Reshape",
294 "Reverse",
295 "Slice"
296 ],
297 "files": {
298 "common": [
299 "src/gpu/cl/kernels/ClDirectConv2dKernel.cpp",
300 "src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp",
301 "src/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp",
302 "src/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp",
303 "src/gpu/cl/kernels/ClIm2ColKernel.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000304 "src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp",
Gian Marco Iodice5d016812022-11-17 11:03:39 +0000305 "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100306 "src/gpu/cl/kernels/ClCol2ImKernel.cpp",
307 "src/gpu/cl/operators/ClConv2d.cpp",
308 "src/gpu/cl/operators/ClDirectConv2d.cpp",
309 "src/gpu/cl/operators/ClGemmConv2d.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000310 "src/gpu/cl/operators/ClIndirectConv2d.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100311 "src/gpu/cl/operators/ClWinogradConv2d.cpp",
312 "src/gpu/cl/kernels/ClWeightsReshapeKernel.cpp",
313 "src/runtime/CL/functions/CLConvolutionLayer.cpp",
314 "src/runtime/CL/functions/CLDirectConvolutionLayer.cpp",
315 "src/runtime/CL/functions/CLFFTConvolutionLayer.cpp",
316 "src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000317 "src/runtime/CL/functions/CLIndirectConvolutionLayer.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100318 "src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp"
319 ]
320 }
321 },
Giorgio Arena273c28c2021-10-14 15:59:15 +0100322 "Conv3d": {
323 "files": {
324 "common": [
325 "src/gpu/cl/kernels/ClDirectConv3dKernel.cpp",
326 "src/gpu/cl/operators/ClDirectConv3d.cpp",
327 "src/runtime/CL/functions/CLConv3D.cpp"
328 ]
329 }
330 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100331 "Copy": {
332 "files": {
333 "common": [
334 "src/gpu/cl/kernels/ClCopyKernel.cpp",
335 "src/gpu/cl/operators/ClCopy.cpp",
336 "src/runtime/CL/functions/CLCopy.cpp"
337 ]
338 }
339 },
340 "CropResize": {
341 "deps": [ "Copy", "Fill", "Scale" ],
342 "files": {
343 "common": [
344 "src/gpu/cl/kernels/ClCropKernel.cpp",
345 "src/gpu/cl/operators/ClCrop.cpp",
346 "src/runtime/CL/functions/CLCrop.cpp",
347 "src/runtime/CL/functions/CLCropResize.cpp"
348 ]
349 }
350 },
351 "Deconv2d": {
352 "deps": [ "Conv2d", "Reverse", "Transpose"],
353 "files": {
354 "common": [
355 "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp",
356 "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp",
Gunes Bayirec0113d2022-11-09 09:26:27 +0000357 "src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp",
358 "src/gpu/cl/operators/ClTransposedConvolution.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100359 "src/runtime/CL/functions/CLDeconvolutionLayer.cpp",
360 "src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp",
361 "src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp",
362 "src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp"
363 ]
364 }
365 },
366 "DepthConvert": {
367 "deps": [ "Cast"],
368 "files": {
369 "common": [ "src/runtime/CL/functions/CLDepthConvertLayer.cpp" ]
370 }
371 },
372 "DepthToSpace": {
373 "files": {
374 "common": [
375 "src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp",
376 "src/runtime/CL/functions/CLDepthToSpaceLayer.cpp"
377 ]
378 }
379 },
380 "DepthwiseConv2d": {
381 "deps": [ "Permute" ],
382 "files": {
383 "common": [
384 "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp",
385 "src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp"
386 ]
387 }
388 },
389 "Dequantize": {
390 "files": {
391 "common": [
392 "src/gpu/cl/kernels/ClDequantizeKernel.cpp",
393 "src/gpu/cl/operators/ClDequantize.cpp",
394 "src/runtime/CL/functions/CLDequantizationLayer.cpp"
395 ]
396 }
397 },
398 "ElementwiseBinary": {
399 "deps": ["Add", "Sub"],
400 "files": {
401 "common": [
402 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
403 "src/gpu/cl/operators/ClElementwiseOperations.cpp",
404 "src/runtime/CL/functions/CLElementwiseOperations.cpp"
405 ]
406 }
407 },
408 "ElementwiseUnary":{
409 "files": {
410 "common": [
411 "src/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp",
412 "src/gpu/cl/operators/ClElementwiseUnary.cpp",
413 "src/runtime/CL/functions/CLElementwiseUnaryLayer.cpp"
414 ]
415 }
416 },
417 "FFT1D": {
418 "files": {
419 "common": [
420 "src/core/CL/kernels/CLFFTDigitReverseKernel.cpp",
421 "src/core/CL/kernels/CLFFTRadixStageKernel.cpp",
422 "src/core/CL/kernels/CLFFTScaleKernel.cpp",
423 "src/runtime/CL/functions/CLFFT1D.cpp"
424 ]
425 }
426 },
427 "FFT2D": {
428 "deps": [ "FFT1D" ],
429 "files": {
430 "common": [ "src/runtime/CL/functions/CLFFT2D.cpp" ]
431 }
432 },
433 "Fill": {
434 "files": {
435 "common": [
436 "src/gpu/cl/kernels/ClFillKernel.cpp",
437 "src/gpu/cl/operators/ClFill.cpp",
438 "src/runtime/CL/functions/CLFill.cpp"
439 ]
440 }
441 },
442 "Flatten": {
443 "files": {
444 "common": [
445 "src/gpu/cl/operators/ClFlatten.cpp",
446 "src/runtime/CL/functions/CLFlattenLayer.cpp"
447 ]
448 }
449 },
450 "Floor": {
451 "files": {
452 "common": [
453 "src/gpu/cl/kernels/ClFloorKernel.cpp",
454 "src/gpu/cl/operators/ClFloor.cpp",
455 "src/runtime/CL/functions/CLFloor.cpp"
456 ]
457 }
458 },
459 "FullyConnected": {
460 "deps": [ "Flatten", "Gemm", "Transpose"],
461 "files": {
462 "common": [
463 "src/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp",
464 "src/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp",
465 "src/gpu/cl/operators/ClFullyConnected.cpp",
466 "src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp",
467 "src/runtime/CL/functions/CLFullyConnectedLayer.cpp"
468 ]
469 }
470 },
471 "Gather": {
472 "files": {
473 "common": [
474 "src/core/CL/kernels/CLGatherKernel.cpp",
475 "src/runtime/CL/functions/CLGather.cpp"]
476 }
477 },
478 "Gemm": {
479 "deps": [ "Cast" ],
480 "files": {
481 "common": [
482 "src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp",
483 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp",
484 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeMidgard.cpp",
485 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeValhall.cpp",
486 "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedBifrost.cpp",
487 "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedValhall.cpp",
488 "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyBifrost.cpp",
489 "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp",
490 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp",
491 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp",
492 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.cpp",
Freddie Liardete572dff2022-05-16 14:09:10 +0100493 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100494 "src/gpu/cl/kernels/ClGemmLowpOffsetContributionKernel.cpp",
495 "src/gpu/cl/kernels/ClGemmLowpOffsetContributionOutputStageKernel.cpp",
496 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.cpp",
497 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp",
498 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100499 "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp",
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000500 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100501 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp",
502 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp",
503 "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.cpp",
504 "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.cpp",
505 "src/gpu/cl/kernels/ClGemmLowpReductionKernel.cpp",
506 "src/gpu/cl/operators/ClGemm.cpp",
507 "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp",
508 "src/gpu/cl/operators/ClGemmLowpOutputStage.cpp",
509 "src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp",
510 "src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp",
511 "src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp",
512 "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp",
513 "src/runtime/CL/functions/CLGEMM.cpp",
514 "src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp",
Ramy Elgammaldf6a3b02022-11-30 16:23:10 +0000515 "src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp",
516 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp",
Gian Marco Iodicea5cb79f2022-12-28 13:53:51 +0000517 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp",
Gian Marco Iodice9d3bd412022-12-30 09:45:00 +0000518 "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigBifrost.cpp",
519 "src/runtime/heuristics/dwc_native/ClDWCNativeDefaultConfigValhall.cpp",
520 "src/runtime/heuristics/dwc_native/ClDWCNativeHeuristicsHelpers.cpp",
Ramy Elgammal34a95d12023-05-11 18:23:44 +0100521 "src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp"
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100522 ]
523 }
524 },
Ramy Elgammal2b6ebfe2023-03-09 21:15:37 +0000525 "MatMul": {
526 "files": {
527 "common": [
Gunes Bayir9d0c4de2023-04-13 18:22:58 +0100528 "src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp",
Gunes Bayire87fa662023-09-07 12:20:33 +0100529 "src/gpu/cl/kernels/ClMatMulLowpNativeMMULKernel.cpp",
Jakub Sujak1ed6a142023-04-13 21:14:42 +0100530 "src/gpu/cl/kernels/ClMatMulNativeKernel.cpp",
SiCong Lia8d80582023-05-19 14:23:37 +0100531 "src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp",
Gunes Bayire87fa662023-09-07 12:20:33 +0100532 "src/gpu/cl/kernels/helpers/MatMulKernelHelpers.cpp",
Ramy Elgammalf26ea2f2023-03-24 11:42:03 +0000533 "src/gpu/cl/operators/ClMatMul.cpp",
Ramy Elgammal34a95d12023-05-11 18:23:44 +0100534 "src/runtime/CL/functions/CLMatMul.cpp",
535 "src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp",
Gunes Bayir85cafff2023-12-18 13:29:31 +0000536 "src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp",
537 "src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultVariantValhall.cpp"
Ramy Elgammal2b6ebfe2023-03-09 21:15:37 +0000538 ]
539 }
540 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100541 "GenerateProposals": {
542 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ],
543 "files": {
544 "common": [
545 "src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp",
546 "src/runtime/CL/functions/CLGenerateProposalsLayer.cpp"
547 ]
548 }
549 },
550 "InstanceNormalize": {
551 "files": {
552 "common": [
553 "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp",
554 "src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp"
555 ]
556 }
557 },
558 "L2Normalize": {
559 "deps": [ "Reduction" ],
560 "files": {
561 "common": [
562 "src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp",
563 "src/runtime/CL/functions/CLL2NormalizeLayer.cpp"
564 ]
565 }
566 },
567 "Logical": {
568 "files": {
569 "common": [
570 "src/gpu/cl/operators/ClLogicalNot.cpp",
571 "src/runtime/CL/functions/CLLogicalAnd.cpp",
572 "src/runtime/CL/functions/CLLogicalNot.cpp",
573 "src/runtime/CL/functions/CLLogicalOr.cpp"
574 ]
575 }
576 },
577 "LSTM": {
578 "deps": [
579 "Activation",
580 "Concatenate",
581 "Copy",
582 "Dequantize",
583 "ElementwiseBinary",
584 "Fill",
585 "FullyConnected",
586 "Gemm",
587 "MeanStdDevNormalize",
588 "Mul",
589 "Quantize",
590 "Slice",
591 "Transpose"
592 ],
593 "files": {
594 "common": [
595 "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp",
596 "src/runtime/CL/functions/CLQLSTMLayer.cpp",
597 "src/runtime/CL/functions/CLLSTMLayer.cpp",
598 "src/runtime/CL/functions/CLLSTMLayerQuantized.cpp"
599 ]
600 }
601 },
602 "MaxUnpool2d": {
603 "deps": [ "Fill" ],
604 "files": {
605 "common": [
606 "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.cpp",
607 "src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp"
608 ]
609 }
610 },
611 "MeanStdDevNormalize": {
612 "deps": [ "Reduction" ],
613 "files": {
614 "common": [
615 "src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp",
616 "src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp",
617 "src/runtime/CL/functions/CLReduceMean.cpp"
618 ]
619 }
620 },
621 "Mul": {
622 "files": {
623 "common": [
624 "src/gpu/cl/kernels/ClMulKernel.cpp",
625 "src/gpu/cl/operators/ClMul.cpp",
626 "src/runtime/CL/functions/CLPixelWiseMultiplication.cpp"
627 ]
628 }
629 },
630 "Normalize": {
631 "files": {
632 "common": [
633 "src/core/CL/kernels/CLNormalizationLayerKernel.cpp",
634 "src/runtime/CL/functions/CLNormalizationLayer.cpp"
635 ]
636 }
637 },
638 "Pad": {
639 "deps": [ "Copy" ],
640 "files": {
641 "common": [
642 "src/core/CL/kernels/CLPadLayerKernel.cpp",
643 "src/runtime/CL/functions/CLPadLayer.cpp"
644 ]
645 }
646 },
647 "Permute": {
648 "files": {
649 "common": [
650 "src/gpu/cl/kernels/ClPermuteKernel.cpp",
651 "src/gpu/cl/operators/ClPermute.cpp",
652 "src/runtime/CL/functions/CLPermute.cpp"
653 ]
654 }
655 },
656 "Pool2d": {
657 "files": {
658 "common": [
659 "src/gpu/cl/kernels/ClPool2dKernel.cpp",
660 "src/gpu/cl/operators/ClPool2d.cpp",
661 "src/runtime/CL/functions/CLPoolingLayer.cpp"
662 ]
663 }
664 },
ramelg0137515692022-02-26 22:06:20 +0000665 "Pool3d": {
666 "files": {
667 "common": [
668 "src/gpu/cl/kernels/ClPool3dKernel.cpp",
669 "src/gpu/cl/operators/ClPool3d.cpp",
670 "src/runtime/CL/functions/CLPooling3dLayer.cpp"
671 ]
672 }
673 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100674 "PRelu": {
675 "deps": [ "ElementwiseBinary" ],
676 "files": {
677 "common": [
678 "src/gpu/cl/operators/ClPRelu.cpp",
679 "src/runtime/CL/functions/CLPReluLayer.cpp"
680 ]
681 }
682 },
683 "PriorBox": {
684 "files": {
685 "common": [
686 "src/core/CL/kernels/CLPriorBoxLayerKernel.cpp",
687 "src/runtime/CL/functions/CLPriorBoxLayer.cpp"
688 ]
689 }
690 },
691 "Quantize": {
692 "files": {
693 "common": [
694 "src/gpu/cl/kernels/ClQuantizeKernel.cpp",
695 "src/gpu/cl/operators/ClQuantize.cpp",
696 "src/runtime/CL/functions/CLQuantizationLayer.cpp"
697 ]
698 }
699 },
700 "Range": {
701 "files": {
702 "common": [
703 "src/core/CL/kernels/CLRangeKernel.cpp",
704 "src/runtime/CL/functions/CLRange.cpp"
705 ]
706 }
707 },
708 "Reduction": {
709 "deps": [ "Reshape" ],
710 "files": {
ramelg01a1f78512022-06-29 16:28:10 +0100711 "common": [
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100712 "src/core/CL/kernels/CLReductionOperationKernel.cpp",
713 "src/runtime/CL/functions/CLReductionOperation.cpp"
714 ]
715 }
716 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100717 "Reorg": {
718 "files": {
719 "common": [
720 "src/core/CL/kernels/CLReorgLayerKernel.cpp",
721 "src/runtime/CL/functions/CLReorgLayer.cpp"
722 ]
723 }
724 },
725 "Reshape": {
726 "files": {
727 "common": [
728 "src/gpu/cl/kernels/ClReshapeKernel.cpp",
729 "src/gpu/cl/operators/ClReshape.cpp",
730 "src/runtime/CL/functions/CLReshapeLayer.cpp"
731 ]
732 }
733 },
734 "Reverse": {
735 "files": {
736 "common": [
737 "src/core/CL/kernels/CLReverseKernel.cpp",
738 "src/runtime/CL/functions/CLReverse.cpp"
739 ]
740 }
741 },
742 "RNN": {
743 "deps": [ "Activation", "Cast", "ElementwiseBinary", "FullyConnected", "Gemm"],
744 "files": {
745 "common": [ "src/runtime/CL/functions/CLRNNLayer.cpp" ]
746 }
747 },
748 "ROIAlign": {
749 "files": {
750 "common": [
751 "src/core/CL/kernels/CLROIAlignLayerKernel.cpp",
752 "src/runtime/CL/functions/CLROIAlignLayer.cpp"
753 ]
754 }
755 },
756 "ROIPool2d": {
757 "files": {
758 "common": [
759 "src/core/CL/kernels/CLROIPoolingLayerKernel.cpp",
760 "src/runtime/CL/functions/CLROIPoolingLayer.cpp"
761 ]
762 }
763 },
764 "Scale": {
765 "files": {
766 "common": [
767 "src/gpu/cl/kernels/ClScaleKernel.cpp",
768 "src/gpu/cl/operators/ClScale.cpp",
769 "src/runtime/CL/functions/CLScale.cpp"
770 ]
771 }
772 },
Mohammed Suhail Munshi8609ca02024-02-29 17:00:07 +0000773 "Scatter": {
774 "files": {
775 "common": [
776 "src/gpu/cl/kernels/ClScatterKernel.cpp",
777 "src/gpu/cl/operators/ClScatter.cpp",
778 "src/runtime/CL/functions/CLScatter.cpp"
779 ]
780 }
781 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100782 "Select": {
783 "files": {
784 "common": [
785 "src/core/CL/kernels/CLSelectKernel.cpp",
786 "src/runtime/CL/functions/CLSelect.cpp"
787 ]
788 }
789 },
790 "Slice": {
791 "deps": [ "StridedSlice" ],
792 "files": {
793 "common": [ "src/runtime/CL/functions/CLSlice.cpp" ]
794 }
795 },
796 "Softmax": {
797 "deps": [ "Permute" ],
798 "files": {
799 "common": [
800 "src/gpu/cl/kernels/ClSoftmaxKernel.cpp",
801 "src/gpu/cl/operators/ClSoftmax.cpp",
802 "src/runtime/CL/functions/CLSoftmaxLayer.cpp"
803 ]
804 }
805 },
806 "SpaceToBatch": {
807 "files": {
808 "common": [
809 "src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp",
810 "src/runtime/CL/functions/CLSpaceToBatchLayer.cpp"
811 ]
812 }
813 },
814 "SpaceToDepth": {
815 "files": {
816 "common": [
817 "src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp",
818 "src/runtime/CL/functions/CLSpaceToDepthLayer.cpp"
819 ]
820 }
821 },
822 "Split": {
823 "deps": [ "StridedSlice" ],
824 "files": {
825 "common": [ "src/runtime/CL/functions/CLSplit.cpp" ]
826 }
827 },
828 "Stack": {
829 "files": {
830 "common": [
831 "src/core/CL/kernels/CLStackLayerKernel.cpp",
832 "src/runtime/CL/functions/CLStackLayer.cpp"
833 ]
834 }
835 },
836 "StridedSlice": {
837 "files": {
838 "common": [
839 "src/core/CL/kernels/CLStridedSliceKernel.cpp",
840 "src/runtime/CL/functions/CLStridedSlice.cpp"
841 ]
842 }
843 },
844 "Sub": {
845 "files": {
846 "common": [
847 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
848 "src/gpu/cl/operators/ClSub.cpp"
849 ]
850 }
851 },
852 "Tile": {
853 "files": {
854 "common": [
855 "src/core/CL/kernels/CLTileKernel.cpp",
856 "src/runtime/CL/functions/CLTile.cpp"
857 ]
858 }
859 },
860 "Transpose": {
861 "files": {
862 "common": [
863 "src/gpu/cl/kernels/ClTransposeKernel.cpp",
864 "src/gpu/cl/operators/ClTranspose.cpp",
865 "src/runtime/CL/functions/CLTranspose.cpp"
866 ]
867 }
868 },
869 "Unstack": {
870 "deps": [ "StridedSlice" ],
871 "files": {
872 "common": [ "src/runtime/CL/functions/CLUnstack.cpp" ]
873 }
874 },
875 "YUVNormalize": {
876 "files": {
877 "common": [
878 "src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp",
879 "src/runtime/CL/functions/CLNormalizePlanarYUVLayer.cpp"
880 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100881 }
Georgios Pinitasbdcdc392021-04-22 16:42:03 +0100882 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100883 }
884},
Michalis Spyrou20fca522021-06-07 14:23:57 +0100885 "cpu": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100886 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +0100887 "src/cpu/CpuContext.cpp",
888 "src/cpu/CpuQueue.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100889 "src/cpu/CpuTensor.cpp",
890 "src/core/NEON/kernels/NEFillBorderKernel.cpp",
891 "src/runtime/NEON/INEOperator.cpp",
892 "src/runtime/NEON/INESimpleFunction.cpp",
893 "src/runtime/NEON/INESimpleFunctionNoBorder.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100894 ],
895 "operators": {
896 "Activation": {
Viet-Hoa Dofd472f02023-03-15 14:05:06 +0000897 "deps": [ "LUT" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100898 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100899 "common": [
900 "src/cpu/operators/CpuActivation.cpp",
901 "src/cpu/kernels/CpuActivationKernel.cpp",
Dana Zlotnika538ae52022-02-21 13:12:41 +0200902 "src/runtime/NEON/functions/NEActivationLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100903 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100904 "neon": {
Dana Zlotnik32291712021-11-25 09:58:27 +0200905 "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ],
ramelg01c827e992022-04-08 03:52:28 +0100906 "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ],
Viet-Hoa Do29db3d22022-08-10 11:56:49 +0100907 "qasymm8": [
908 "src/cpu/kernels/activation/generic/neon/qasymm8.cpp",
909 "src/cpu/kernels/activation/generic/neon/lut.cpp"
910 ],
911 "qasymm8_signed": [
912 "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp",
913 "src/cpu/kernels/activation/generic/neon/lut.cpp"
914 ],
Dana Zlotnika538ae52022-02-21 13:12:41 +0200915 "qsymm16": [ "src/cpu/kernels/activation/generic/neon/qsymm16.cpp" ]
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100916 },
917 "sve": {
Dana Zlotnik32291712021-11-25 09:58:27 +0200918 "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ],
SiCong Lic0463a22023-05-17 13:46:13 +0100919 "fp32": [ "src/cpu/kernels/activation/generic/sve/fp32.cpp" ]
ramelg01c827e992022-04-08 03:52:28 +0100920 },
alerah01c9e519d2022-01-31 19:04:10 +0200921 "sve2":{
SiCong Lic0463a22023-05-17 13:46:13 +0100922 "qasymm8": [
923 "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp",
924 "src/cpu/kernels/activation/generic/sve2/lut.cpp"
925 ],
Dana Zlotnik32291712021-11-25 09:58:27 +0200926 "qasymm8_signed": [ "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp" ],
927 "qsymm16": [ "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100928 }
929 }
930 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100931 "ArgMinMax": {
932 "deps": [ "Reduction" ],
933 "files": {
934 "common": [ "src/runtime/NEON/functions/NEArgMinMaxLayer.cpp" ]
935 }
936 },
Michalis Spyrou20fca522021-06-07 14:23:57 +0100937 "Add": {
938 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100939 "common": [
940 "src/cpu/operators/CpuAdd.cpp",
941 "src/cpu/kernels/CpuAddKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200942 "src/runtime/NEON/functions/NEArithmeticAddition.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100943 ],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200944 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200945 "common": ["src/cpu/kernels/add/generic/neon/impl.cpp"],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200946 "fp32":["src/cpu/kernels/add/generic/neon/fp32.cpp"],
947 "fp16":["src/cpu/kernels/add/generic/neon/fp16.cpp"],
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200948 "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"],
949 "qasymm8": ["src/cpu/kernels/add/generic/neon/qasymm8.cpp"],
950 "qasymm8_signed": ["src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp"],
951 "qsymm16": ["src/cpu/kernels/add/generic/neon/qsymm16.cpp"]
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200952 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100953 "sve": {
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200954 "common": [ "src/cpu/kernels/add/generic/sve/impl.cpp" ],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200955 "integer":["src/cpu/kernels/add/generic/sve/integer.cpp"],
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200956 "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"],
957 "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"]
958 },
959 "sve2": {
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200960 "qasymm8": [ "src/cpu/kernels/add/generic/sve2/qasymm8.cpp" ],
961 "qasymm8_signed": [ "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp" ],
962 "qsymm16": [ "src/cpu/kernels/add/generic/sve2/qsymm16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100963 }
964 }
965 },
Gunes Bayirae72a462023-01-29 13:24:24 +0000966 "AddMulAdd": {
967 "files": {
968 "common": [
969 "src/cpu/operators/CpuAddMulAdd.cpp",
970 "src/cpu/kernels/CpuAddMulAddKernel.cpp",
971 "src/runtime/NEON/functions/NEAddMulAdd.cpp"
972 ],
973 "neon": {
974 "fp32":["src/cpu/kernels/addmuladd/generic/neon/fp32.cpp"],
975 "fp16":["src/cpu/kernels/addmuladd/generic/neon/fp16.cpp"],
976 "qasymm8": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8.cpp"],
977 "qasymm8_signed": ["src/cpu/kernels/addmuladd/generic/neon/qasymm8_signed.cpp"]
978 }
979 }
980 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100981 "BatchNormalize": {
Michalis Spyrou20fca522021-06-07 14:23:57 +0100982 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100983 "common": [
984 "src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp",
985 "src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp",
986 "src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp",
987 "src/runtime/NEON/functions/NEFuseBatchNormalization.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100988 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100989 "neon": {
Yair Schwarzbaum41a729e2021-11-15 20:42:47 +0200990 "common": [
Pablo Marquez Tello7e589802023-09-14 09:41:37 +0100991 "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp"
Yair Schwarzbaum41a729e2021-11-15 20:42:47 +0200992 ],
993 "fp16": [
994 "src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp",
995 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp",
Pablo Marquez Tello8d4cdd42023-11-21 10:10:01 +0000996 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp",
997 "src/cpu/kernels/fuse_batch_normalization/nchw/neon/fp16.cpp"
998
Yair Schwarzbaum41a729e2021-11-15 20:42:47 +0200999 ],
1000 "fp32": [
1001 "src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp",
1002 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp",
Pablo Marquez Tello8d4cdd42023-11-21 10:10:01 +00001003 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp",
1004 "src/cpu/kernels/fuse_batch_normalization/nchw/neon/fp32.cpp"
Yair Schwarzbaum41a729e2021-11-15 20:42:47 +02001005 ]
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001006 },
1007 "sve": {
1008 "fp16": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp" ],
1009 "fp32": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001010 }
1011 }
1012 },
1013 "BatchToSpace": {
1014 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001015 "common": [
1016 "src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp",
1017 "src/runtime/NEON/functions/NEBatchToSpaceLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001018 ]
1019 }
1020 },
1021 "BitwiseAnd": {
1022 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001023 "common": [
1024 "src/core/NEON/kernels/NEBitwiseAndKernel.cpp",
1025 "src/runtime/NEON/functions/NEBitwiseAnd.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001026 ]
1027 }
1028 },
1029 "BitwiseNot": {
1030 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001031 "common": [
1032 "src/core/NEON/kernels/NEBitwiseNotKernel.cpp",
1033 "src/runtime/NEON/functions/NEBitwiseNot.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001034 ]
1035 }
1036 },
1037 "BitwiseOr": {
1038 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001039 "common": [
1040 "src/core/NEON/kernels/NEBitwiseOrKernel.cpp",
1041 "src/runtime/NEON/functions/NEBitwiseOr.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001042 ]
1043 }
1044 },
1045 "BitwiseXor": {
1046 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001047 "common": [
1048 "src/core/NEON/kernels/NEBitwiseXorKernel.cpp",
1049 "src/runtime/NEON/functions/NEBitwiseXor.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001050 ]
1051 }
1052 },
1053 "BoundingBoxTransform": {
1054 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001055 "common": [
1056 "src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp",
1057 "src/runtime/NEON/functions/NEBoundingBoxTransform.cpp"
Dana Zlotnik3475ffe2022-01-03 14:37:10 +02001058 ],
1059 "neon":{
1060 "common":["src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp"],
1061 "fp32":["src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp"],
1062 "fp16":["src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp"],
1063 "qsymm16":["src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp"]
Dana Zlotnik3475ffe2022-01-03 14:37:10 +02001064 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001065 }
1066 },
1067 "Cast": {
1068 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001069 "common": [
1070 "src/cpu/operators/CpuCast.cpp",
1071 "src/cpu/kernels/CpuCastKernel.cpp",
Adnan AlSinan40a9d3e2023-09-15 13:46:17 +01001072 "src/runtime/NEON/functions/NECast.cpp"
Yair Schwarzbaum298b2c02022-02-01 08:55:56 +02001073 ],
1074 "neon":{
1075 "fp16":["src/cpu/kernels/cast/generic/neon/fp16.cpp"]
1076 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001077 }
1078 },
1079 "ChannelShuffle": {
1080 "files": {
1081 "common": [
1082 "src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp",
1083 "src/runtime/NEON/functions/NEChannelShuffleLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001084 ]
1085 }
1086 },
1087 "Concatenate": {
1088 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001089 "common": [
1090 "src/cpu/operators/CpuConcatenate.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001091 "src/cpu/kernels/CpuConcatenateWidthKernel.cpp",
1092 "src/cpu/kernels/CpuConcatenateBatchKernel.cpp",
1093 "src/cpu/kernels/CpuConcatenateDepthKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001094 "src/cpu/kernels/CpuConcatenateHeightKernel.cpp",
1095 "src/runtime/NEON/functions/NEConcatenateLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001096 ]
1097 }
1098 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001099 "Conv2d": {
1100 "deps": [
1101 "Activation",
1102 "ElementwiseBinary",
1103 "FFT2D",
1104 "Gemm",
1105 "Mul",
1106 "Pad",
1107 "Permute",
1108 "Reshape",
1109 "Reverse",
1110 "Slice"
1111 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001112 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001113 "common": [
1114 "src/cpu/operators/CpuConv2d.cpp",
1115 "src/cpu/operators/CpuDirectConv2d.cpp",
1116 "src/cpu/operators/CpuGemmDirectConv2d.cpp",
1117 "src/cpu/operators/CpuGemmConv2d.cpp",
1118 "src/cpu/operators/CpuWinogradConv2d.cpp",
1119 "src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp",
1120 "src/cpu/kernels/CpuDirectConv2dKernel.cpp",
1121 "src/cpu/kernels/CpuDirectConv2dOutputStageKernel.cpp",
1122 "src/cpu/kernels/CpuWinogradConv2dKernel.cpp",
1123 "src/cpu/kernels/CpuCol2ImKernel.cpp",
1124 "src/cpu/kernels/CpuIm2ColKernel.cpp",
1125 "src/cpu/kernels/CpuWeightsReshapeKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001126 "src/runtime/NEON/functions/NEConvolutionLayer.cpp",
1127 "src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp",
1128 "src/runtime/NEON/functions/NEFFTConvolutionLayer.cpp",
1129 "src/runtime/NEON/functions/NEGEMMConv2d.cpp",
1130 "src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp",
1131 "src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001132 ],
1133 "neon": {
1134 "common": [
1135 "src/core/NEON/kernels/convolution/common/padding.cpp",
1136 "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
1137 "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
1138 "src/core/NEON/kernels/convolution/common/utils.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001139 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp32.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001140 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp32.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001141 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp32.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001142 "src/core/NEON/kernels/convolution/winograd/winograd_fp32.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001143 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp",
1144 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp",
1145 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp",
1146 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001147 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp",
1148 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp",
1149 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp",
1150 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp",
1151 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp",
1152 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001153 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_3x3.cpp",
1154 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_5x5.cpp",
1155 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_4x4_3x3.cpp",
1156 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x2_1x7.cpp",
1157 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp",
1158 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x6_1x3.cpp",
alerah01c9e519d2022-01-31 19:04:10 +02001159 "src/cpu/kernels/directconv2d/nhwc/neon/impl.cpp",
Pablo Marquez Telloe5362e72023-10-16 14:14:14 +01001160 "src/cpu/kernels/directconv2d/nhwc/neon/qasymm8.cpp",
alerah01c9e519d2022-01-31 19:04:10 +02001161 "src/cpu/kernels/directconv2d/nchw/all.cpp"
Pablo Marquez Telloe5362e72023-10-16 14:14:14 +01001162
alerah01c9e519d2022-01-31 19:04:10 +02001163 ],
1164 "fp32": [
1165 "src/cpu/kernels/directconv2d/nhwc/neon/fp32.cpp"
Pablo Marquez Telloe5362e72023-10-16 14:14:14 +01001166 ],
1167 "fp16": [
1168 "src/cpu/kernels/directconv2d/nchw/fp16.cpp",
Pablo Marquez Tellob5264312023-12-12 10:09:59 +00001169 "src/core/NEON/kernels/convolution/winograd/winograd_fp16.cpp",
1170 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp16.cpp",
1171 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp16.cpp",
1172 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp16.cpp",
1173 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp",
1174 "src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp",
1175 "src/core/NEON/kernels/convolution/winograd/weight_transforms/a64_fp16_4x4_3x3.cpp",
Pablo Marquez Telloe5362e72023-10-16 14:14:14 +01001176 "src/cpu/kernels/directconv2d/nhwc/neon/fp16.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001177 ]
ramelg0199c46472022-07-15 13:08:39 +01001178 },
1179 "sve": {
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001180 "common": [
1181 "src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp",
1182 "src/core/NEON/kernels/convolution/winograd/input_transforms/sme_fp32_mla_6x6.cpp",
1183 "src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp"
1184 ]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001185 }
Michalis Spyroub55f8e82021-07-22 11:23:11 +01001186 }
1187 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001188 "Copy": {
1189 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001190 "common": [
1191 "src/cpu/operators/CpuCopy.cpp",
1192 "src/cpu/kernels/CpuCopyKernel.cpp",
1193 "src/runtime/NEON/functions/NECopy.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001194 ]
1195 }
1196 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001197 "CropResize": {
1198 "deps": [ "Scale" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001199 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001200 "common": [
1201 "src/core/NEON/kernels/NECropKernel.cpp",
1202 "src/runtime/NEON/functions/NECropResize.cpp"
alerah0127619932021-11-24 17:21:43 +02001203 ],
1204 "neon": {
alerah0127619932021-11-24 17:21:43 +02001205 "fp32": [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ],
1206 "fp16": [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ],
1207 "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp" ]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001208 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001209 }
1210 },
1211 "Deconv2d": {
1212 "deps": [ "Conv2d", "Reverse", "Transpose"],
1213 "files": {
1214 "common": [
1215 "src/runtime/NEON/functions/NEDeconvolutionLayer.cpp"
1216 ]
1217 }
1218 },
1219 "DepthConvert": {
1220 "deps": [ "Cast"],
1221 "files": {
1222 "common": [
1223 "src/runtime/NEON/functions/NEDepthConvertLayer.cpp"
1224 ]
1225 }
1226 },
1227 "DepthToSpace": {
1228 "files": {
1229 "common": [
1230 "src/core/NEON/kernels/NEDepthToSpaceLayerKernel.cpp",
Viet-Hoa Do47370942023-11-13 17:20:45 +00001231 "src/runtime/NEON/functions/NEDepthToSpaceLayer.cpp",
1232 "src/cpu/kernels/depth_to_space/nhwc/any/impl.cpp",
1233 "src/cpu/kernels/depth_to_space/nchw/any/impl.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001234 ]
1235 }
1236 },
1237 "DepthwiseConv2d": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001238 "deps": [ "Activation", "Permute" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001239 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001240 "common": [
Georgios Pinitas7891a732021-08-20 21:39:25 +01001241 "src/cpu/operators/CpuDepthwiseConv2d.cpp",
1242 "src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001243 "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.cpp",
1244 "src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001245 "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001246 ],
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001247 "neon": {
1248 "common": [
1249 "src/core/NEON/kernels/convolution/common/padding.cpp",
1250 "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
1251 "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
1252 "src/core/NEON/kernels/convolution/common/utils.cpp",
ramelg01c827e992022-04-08 03:52:28 +01001253 "src/core/NEON/kernels/arm_conv/addressing.cpp",
Pablo Marquez Tello4e2bbbb2023-01-09 17:21:01 +00001254 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_common.cpp",
ramelg018a164882022-04-07 02:42:52 +01001255 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_strategies_common.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001256 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
1257 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001258 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
1259 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001260 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
1261 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
1262 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1263 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1264 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1265 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1266 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1267 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1268 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1269 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001270 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1271 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1272 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1273 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1274 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1275 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1276 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1277 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1278 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1279 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1280 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1281 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1282 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1283 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1284 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1285 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1286 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1287 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1288 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1289 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1290 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1291 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1292 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1293 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1294 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1295 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
ramelg018a164882022-04-07 02:42:52 +01001296 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1297 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1298 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001299 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
ramelg01c827e992022-04-08 03:52:28 +01001300 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp",
ramelg018a164882022-04-07 02:42:52 +01001301 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp",
1302 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp",
Michael Tyler8deee9b2023-06-30 11:26:05 +01001303 "src/core/NEON/kernels/arm_conv/depthwise/premultiply.cpp",
Dana Zlotnikebbae942022-02-03 12:52:15 +02001304 "src/cpu/kernels/depthwiseconv2d/generic/neon/impl.cpp"
ramelg01a1f78512022-06-29 16:28:10 +01001305 ],
Pablo Marquez Tello64f4a302023-11-06 15:15:53 +00001306 "fp16":[
1307 "src/cpu/kernels/depthwiseconv2d/generic/neon/fp16.cpp",
1308 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
1309 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1310 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1311 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1312 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1313 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1314 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1315 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1316 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1317 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1318 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1319 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1320 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp"
1321 ],
ramelg01a1f78512022-06-29 16:28:10 +01001322 "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"],
1323 "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"],
Dana Zlotnikebbae942022-02-03 12:52:15 +02001324 "qasymm8_signed":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8_signed.cpp"]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001325 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001326 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001327 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +01001328 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp",
1329 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp",
1330 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1331 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1332 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1333 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1334 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1335 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1336 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1337 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1338 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1339 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001340 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1341 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001342 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1343 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1344 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1345 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1346 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1347 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1348 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1349 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1350 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1351 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
1352 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
1353 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1354 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1355 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1356 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1357 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1358 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1359 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1360 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1361 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1362 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1363 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1364 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1365 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1366 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1367 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1368 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001369 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
David Mansellb2758f32023-03-30 19:10:52 +01001370 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1371 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1372 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1373 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1374 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1375 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1376 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1377 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1378 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1379 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001380 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1381 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1382 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1383 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1384 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1385 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1386 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1387 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1388 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s1_4rows_mla_za/generic.cpp",
1389 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp",
1390 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp",
1391 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s2_4rows_mla_za/generic.cpp",
1392 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp",
1393 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s2_4rows_dot_za/generic.cpp",
1394 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp",
1395 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp",
1396 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001397 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1398 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1399 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
1400 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001401 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1402 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1403 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
1404 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001405 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1406 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1407 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001408 ]
1409 }
1410 }
1411 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001412 "Dequantize": {
1413 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001414 "common": [
1415 "src/cpu/operators/CpuDequantize.cpp",
1416 "src/cpu/kernels/CpuDequantizeKernel.cpp",
1417 "src/runtime/NEON/functions/NEDequantizationLayer.cpp"
Ramy Elgammald3d2e9b2024-05-15 22:23:46 +01001418 ],
1419 "neon":{
1420 "fp32":["src/cpu/kernels/dequantize/generic/neon/fp32.cpp"],
1421 "fp16":["src/cpu/kernels/dequantize/generic/neon/fp16.cpp"]
1422 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001423 }
1424 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001425 "DetectionPostProcess": {
1426 "deps": [ "Dequantize" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001427 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001428 "common" : [ "src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001429 }
1430 },
Sheri Zhang6d9c9822021-09-24 16:02:57 +01001431 "Conv3d": {
1432 "deps": [
1433 "Activation"
1434 ],
1435 "files": {
1436 "common": [
1437 "src/cpu/operators/CpuDirectConv3d.cpp",
1438 "src/cpu/kernels/CpuDirectConv3dKernel.cpp",
1439 "src/runtime/NEON/functions/NEConv3D.cpp"
1440 ]
1441 }
1442 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001443 "ElementwiseBinary": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001444 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001445 "common": [
1446 "src/cpu/operators/CpuElementwise.cpp",
1447 "src/cpu/kernels/CpuElementwiseKernel.cpp",
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001448 "src/runtime/NEON/functions/NEElementwiseOperations.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001449 ],
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001450 "neon":{
1451 "fp32": ["src/cpu/kernels/elementwise_binary/generic/neon/fp32.cpp"],
1452 "fp16": ["src/cpu/kernels/elementwise_binary/generic/neon/fp16.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001453 "integer": ["src/cpu/kernels/elementwise_binary/generic/neon/integer.cpp"],
1454 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8.cpp"],
1455 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp"]
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001456 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001457 "sve": {
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001458 "common": ["src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp" ],
1459 "integer": ["src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp"],
1460 "fp32": ["src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001461 "fp16": ["src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp"]
1462
ramelg01a1f78512022-06-29 16:28:10 +01001463 },
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001464 "sve2":{
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001465 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8.cpp"],
1466 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001467 }
1468 }
1469 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001470 "ElementwiseUnary":{
Viet-Hoa Dofd472f02023-03-15 14:05:06 +00001471 "deps": [ "LUT" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001472 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001473 "common": [
1474 "src/cpu/operators/CpuElementwiseUnary.cpp",
1475 "src/cpu/kernels/CpuElementwiseUnaryKernel.cpp",
1476 "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001477 ],
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001478 "neon": {
1479 "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"],
1480 "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"],
Viet-Hoa Dofd472f02023-03-15 14:05:06 +00001481 "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"],
Ramy Elgammal8b7f42a2023-03-31 16:16:15 +01001482 "qasymm8": [
1483 "src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp",
1484 "src/cpu/kernels/elementwise_unary/generic/neon/qasymm8.cpp"
1485 ],
1486 "qasymm8_signed": [
1487 "src/cpu/kernels/elementwise_unary/generic/neon/q8.cpp",
1488 "src/cpu/kernels/elementwise_unary/generic/neon/qasymm8_signed.cpp"
1489 ]
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001490 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001491 "sve": {
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001492 "common": ["src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp" ],
1493 "integer": ["src/cpu/kernels/elementwise_unary/generic/sve/integer.cpp"],
1494 "fp32": ["src/cpu/kernels/elementwise_unary/generic/sve/fp32.cpp"],
SiCong Lic0463a22023-05-17 13:46:13 +01001495 "fp16": ["src/cpu/kernels/elementwise_unary/generic/sve/fp16.cpp"]
1496 },
1497 "sve2": {
1498 "qasymm8": ["src/cpu/kernels/elementwise_unary/generic/sve2/q8.cpp"],
1499 "qasymm8_signed": ["src/cpu/kernels/elementwise_unary/generic/sve2/q8.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001500 }
1501 }
1502 },
1503 "FFT1D": {
Freddie Liardet487d3902021-09-21 12:36:43 +01001504 "deps": [ "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001505 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001506 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +01001507 "src/core/NEON/kernels/NEFFTDigitReverseKernel.cpp",
1508 "src/core/NEON/kernels/NEFFTRadixStageKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001509 "src/core/NEON/kernels/NEFFTScaleKernel.cpp",
1510 "src/runtime/NEON/functions/NEFFT1D.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001511 ]
1512 }
1513 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001514 "FFT2D": {
1515 "deps": [ "FFT1D" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001516 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001517 "common": [
1518 "src/runtime/NEON/functions/NEFFT2D.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001519 ]
1520 }
1521 },
1522 "Fill": {
1523 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001524 "common": [
1525 "src/cpu/operators/CpuFill.cpp",
1526 "src/cpu/kernels/CpuFillKernel.cpp",
1527 "src/runtime/NEON/functions/NEFill.cpp"
1528 ]
1529 }
1530 },
1531 "Flatten": {
1532 "deps: ": [ "Reshape" ],
1533 "files": {
1534 "common": [
1535 "src/cpu/operators/CpuFlatten.cpp",
1536 "src/runtime/NEON/functions/NEFlattenLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001537 ]
1538 }
1539 },
1540 "Floor": {
1541 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001542 "common": [
1543 "src/cpu/operators/CpuFloor.cpp",
1544 "src/cpu/kernels/CpuFloorKernel.cpp",
1545 "src/runtime/NEON/functions/NEFloor.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001546 ],
1547 "neon": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001548 "fp32": [ "src/cpu/kernels/floor/neon/fp32.cpp" ],
1549 "fp16": [ "src/cpu/kernels/floor/neon/fp16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001550 }
1551 }
1552 },
Michele Di Giorgiod9cdf142021-07-02 15:17:08 +01001553 "FullyConnected": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001554 "deps": [ "Flatten", "Gemm", "Transpose"],
Michele Di Giorgiod9cdf142021-07-02 15:17:08 +01001555 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001556 "common": [
1557 "src/cpu/kernels/CpuConvertFullyConnectedWeightsKernel.cpp",
1558 "src/cpu/operators/CpuConvertFullyConnectedWeights.cpp",
1559 "src/cpu/operators/CpuFullyConnected.cpp",
1560 "src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp",
1561 "src/runtime/NEON/functions/NEFullyConnectedLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001562 ]
1563 }
1564 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001565 "Gather": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001566 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001567 "common": [
1568 "src/core/NEON/kernels/NEGatherKernel.cpp",
1569 "src/runtime/NEON/functions/NEGather.cpp"
1570 ]
1571 }
1572 },
1573 "Gemm": {
Freddie Liardet487d3902021-09-21 12:36:43 +01001574 "deps": [ "Quantize", "Add"],
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001575 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001576 "common": [
1577 "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001578 "src/cpu/kernels/CpuGemmMatrixAdditionKernel.cpp",
1579 "src/cpu/kernels/CpuGemmMatrixMultiplyKernel.cpp",
1580 "src/cpu/kernels/CpuGemmTranspose1xWKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001581 "src/cpu/kernels/CpuGemmInterleave4x4Kernel.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001582 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp",
1583 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
1584 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
1585 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
1586 "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp",
1587 "src/cpu/kernels/CpuGemmLowpMatrixReductionKernel.cpp",
1588 "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001589 "src/cpu/kernels/CpuGemmLowpOffsetContributionKernel.cpp",
1590 "src/cpu/operators/CpuGemm.cpp",
1591 "src/cpu/operators/CpuGemmLowpOutputStage.cpp",
1592 "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001593 "src/runtime/NEON/functions/NEGEMM.cpp",
1594 "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001595 "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001596 ],
1597 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001598 "common": [
1599 "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001600 "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
Renato Arantes36a75da2024-01-26 17:31:18 +00001601 "src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001602 "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp",
1603 "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp",
Jonathan Deakina668f9f2024-01-24 09:15:38 +00001604 "src/core/NEON/kernels/arm_gemm/gemm_s8fp32.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001605 "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp",
1606 "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp",
1607 "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp",
1608 "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp",
Gunes Bayiref637392024-02-12 21:32:51 +00001609 "src/core/NEON/kernels/arm_gemm/interleave-8way.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001610 "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001611 "src/core/NEON/kernels/arm_gemm/mergeresults.cpp",
1612 "src/core/NEON/kernels/arm_gemm/misc.cpp",
1613 "src/core/NEON/kernels/arm_gemm/quantized.cpp",
1614 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp",
1615 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp",
1616 "src/core/NEON/kernels/arm_gemm/transform.cpp",
1617 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
1618 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
1619 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
1620 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
1621 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
1622 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
1623 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
1624 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
1625 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
1626 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001627 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
1628 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001629 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
1630 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
1631 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
1632 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
1633 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
1634 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
1635 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
1636 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
1637 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
1638 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
1639 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
1640 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
1641 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
1642 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
1643 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
1644 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
1645 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
1646 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
1647 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
1648 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
1649 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
1650 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
1651 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
1652 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
1653 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
1654 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
1655 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp",
1656 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
1657 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp",
1658 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
1659 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp",
1660 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
1661 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
1662 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
1663 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
1664 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
1665 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
1666 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
1667 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
1668 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
1669 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
1670 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
1671 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
1672 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
1673 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
1674 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001675 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp",
Dana Zlotnik256ac622022-02-02 15:06:11 +02001676 "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp",
1677 "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001678 ],
Dana Zlotnik256ac622022-02-02 15:06:11 +02001679 "fp32":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp",
1680 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp"],
1681 "fp16":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp",
Pablo Marquez Tello2217f1e2024-05-14 07:54:19 +01001682 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
1683 "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
Pablo Marquez Telloc0e01282024-05-16 11:29:48 +01001684 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp",
1685 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp",
Pablo Marquez Tello2217f1e2024-05-14 07:54:19 +01001686 "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp",
1687 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
1688 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
1689 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
1690 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
Dana Zlotnik256ac622022-02-02 15:06:11 +02001691 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp"],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001692 "estate32": [
1693 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp",
1694 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp",
1695 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp"
1696 ],
1697 "estate64": [
Freddie Liardet487d3902021-09-21 12:36:43 +01001698 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp"
Pablo Marquez Tello2217f1e2024-05-14 07:54:19 +01001699
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001700 ],
Nathan John Sircombed7113e42023-04-26 15:02:43 +01001701 "fixed_format_kernels": [
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001702 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp",
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001703 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp",
1704 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
Milos Puzovic905786e2024-03-26 14:34:30 +00001705 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001706 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp",
1707 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp",
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001708 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001709 ]
1710 },
1711 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001712 "common": [
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001713 "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp",
1714 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp",
David Mansell0fa92b82023-10-17 13:33:24 +01001715 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16fp32fp16_dot_16VL/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001716 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp",
1717 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp",
1718 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp",
1719 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp",
1720 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp",
1721 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp",
1722 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp",
David Mansellaaa9da12023-03-10 13:48:50 +00001723 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp",
1724 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp",
1725 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001726 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp",
1727 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp",
1728 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp",
1729 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp",
1730 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp",
1731 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp",
Jonathan Deakina668f9f2024-01-24 09:15:38 +00001732 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/generic.cpp",
1733 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/generic.cpp",
1734 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001735 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp",
1736 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp",
1737 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp",
1738 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp",
1739 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp",
1740 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001741 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001742 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp",
1743 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001744 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001745 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001746 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001747 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001748 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001749 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp",
1750 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001751 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001752 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001753 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001754 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp",
1755 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001756 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001757 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001758 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001759 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp",
1760 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001761 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001762 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001763 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp",
1764 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001765 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001766 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001767 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001768 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp",
1769 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001770 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001771 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp",
1772 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001773 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001774 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp",
1775 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001776 "src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001777 "src/core/NEON/kernels/arm_gemm/transform-sve.cpp",
1778 "src/core/NEON/kernels/arm_gemm/misc-sve.cpp"
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001779 ],
Nathan John Sircombed7113e42023-04-26 15:02:43 +01001780 "fixed_format_kernels": [
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001781 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp",
1782 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp",
1783 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp",
1784 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp",
1785 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp",
1786 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp",
1787 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp",
1788 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp",
1789 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp",
1790 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp",
1791 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp"
1792 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001793 }
1794 }
1795 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001796 "GenerateProposals": {
1797 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001798 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001799 "common": [
1800 "src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp",
1801 "src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp"
Dana Zlotnik50ce2772021-12-29 13:55:56 +02001802 ],
1803 "neon":{
1804 "common":["src/cpu/kernels/genproposals/generic/neon/impl.cpp"],
1805 "fp16":["src/cpu/kernels/genproposals/generic/neon/fp16.cpp"],
1806 "fp32":["src/cpu/kernels/genproposals/generic/neon/fp32.cpp"],
1807 "qsymm16":["src/cpu/kernels/genproposals/generic/neon/qsymm16.cpp"]
1808 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001809 }
1810 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001811 "InstanceNormalize": {
1812 "deps": [ "Permute", "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001813 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001814 "common": [
1815 "src/core/NEON/kernels/NEInstanceNormalizationLayerKernel.cpp",
1816 "src/runtime/NEON/functions/NEInstanceNormalizationLayer.cpp"
ramelg01a1f78512022-06-29 16:28:10 +01001817 ],
Dana Zlotnikd7e2ec52022-01-03 10:59:41 +02001818 "neon":{
1819 "common":["src/cpu/kernels/instancenorm/generic/neon/impl.cpp"],
1820 "fp16":["src/cpu/kernels/instancenorm/generic/neon/fp16.cpp"],
1821 "fp32":["src/cpu/kernels/instancenorm/generic/neon/fp32.cpp"]
1822 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001823 }
1824 },
1825 "L2Normalize": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001826 "deps": [ "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001827 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001828 "common": [
1829 "src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp",
1830 "src/runtime/NEON/functions/NEL2NormalizeLayer.cpp"
Yair Schwarzbaum5e993182022-01-10 15:11:07 +02001831 ],
1832 "neon":{
Yair Schwarzbaum5e993182022-01-10 15:11:07 +02001833 "fp32":["src/cpu/kernels/l2normlayer/generic/neon/fp32.cpp"],
1834 "fp16":["src/cpu/kernels/l2normlayer/generic/neon/fp16.cpp"]
1835 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001836 }
1837 },
1838 "Logical": {
1839 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001840 "common": [
1841 "src/core/NEON/kernels/NELogicalKernel.cpp",
1842 "src/runtime/NEON/functions/NELogical.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001843 ]
1844 }
1845 },
Viet-Hoa Dofd472f02023-03-15 14:05:06 +00001846 "LUT": {
1847 "files": {
1848 "common": [],
1849 "neon":{
1850 "qasymm8": ["src/cpu/kernels/lut/generic/neon/u8.cpp"],
1851 "qasymm8_signed": ["src/cpu/kernels/lut/generic/neon/u8.cpp"]
1852 },
Mohammed Suhail Munshi7467ba82023-12-05 14:27:31 +00001853 "sve": {
1854 "fp16": ["src/cpu/kernels/lut/generic/sve/u16.cpp"],
1855 "qasymm16": ["src/cpu/kernels/lut/generic/sve/u16.cpp"],
1856 "qasymm16_signed": ["src/cpu/kernels/lut/generic/sve/u16.cpp"]
1857 },
SiCong Lic0463a22023-05-17 13:46:13 +01001858 "sve2": {
1859 "qasymm8": ["src/cpu/kernels/lut/generic/sve2/u8.cpp"],
1860 "qasymm8_signed": ["src/cpu/kernels/lut/generic/sve2/u8.cpp"]
Viet-Hoa Dofd472f02023-03-15 14:05:06 +00001861 }
1862 }
1863 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001864 "LSTM": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001865 "deps": [
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001866 "Activation",
1867 "Concatenate",
1868 "Copy",
1869 "Dequantize",
1870 "ElementwiseBinary",
1871 "Fill",
1872 "FullyConnected",
1873 "Gemm",
1874 "MeanStdDevNormalize",
1875 "Mul",
1876 "Quantize",
1877 "Slice",
1878 "Transpose"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001879 ],
1880 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001881 "common": [
1882 "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp",
1883 "src/runtime/NEON/functions/NELSTMLayer.cpp",
1884 "src/runtime/NEON/functions/NELSTMLayerQuantized.cpp",
1885 "src/runtime/NEON/functions/NEQLSTMLayer.cpp"
1886 ]
1887 }
1888 },
1889 "MaxUnpool2d": {
1890 "deps": [ "Fill" ],
1891 "files": {
1892 "common": [
Dana Zlotnik149203b2022-01-26 12:38:03 +02001893 "src/cpu/kernels/CpuMaxUnpoolingLayerKernel.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001894 "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp",
Dana Zlotnik149203b2022-01-26 12:38:03 +02001895 "src/cpu/operators/CpuMaxUnpooling.cpp"
Dana Zlotnik143310c2021-12-30 13:24:52 +02001896 ],
1897 "neon":{
Dana Zlotnik143310c2021-12-30 13:24:52 +02001898 "fp32":["src/cpu/kernels/maxunpool/generic/neon/fp32.cpp"],
1899 "fp16":["src/cpu/kernels/maxunpool/generic/neon/fp16.cpp"],
1900 "qasymm8":["src/cpu/kernels/maxunpool/generic/neon/qasymm8.cpp"],
Pablo Marquez Tello6b6ba9e2023-09-22 14:36:03 +01001901 "qasymm8_signed":[ "src/cpu/kernels/maxunpool/generic/neon/qasymm8_signed.cpp"]
Dana Zlotnik143310c2021-12-30 13:24:52 +02001902 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001903 }
1904 },
1905 "Mean": {
1906 "deps" : [ "Reduction" ],
1907 "files": {
1908 "common": [ "src/runtime/NEON/functions/NEReduceMean.cpp" ]
1909 }
1910 },
1911 "MeanStdDevNormalize": {
1912 "files": {
1913 "common": [
1914 "src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp",
1915 "src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp"
Dana Zlotnik027bcef2021-12-27 17:35:00 +02001916 ],
1917 "neon":{
1918 "common":["src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp"],
1919 "fp32":["src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp"],
Murray Kornelsen6e09e142022-07-13 21:40:26 -04001920 "fp16":["src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp"],
1921 "qasymm8":["src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp"]
Dana Zlotnik027bcef2021-12-27 17:35:00 +02001922 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001923 }
1924 },
Mohammed Suhail Munshia1b1e412023-03-23 22:21:31 +00001925 "MatMul" : {
Viet-Hoa Do99b6fd22023-04-27 17:06:10 +01001926 "deps": [ "Transpose" ],
Mohammed Suhail Munshia1b1e412023-03-23 22:21:31 +00001927 "files": {
1928 "common": [
1929 "src/cpu/operators/CpuMatMul.cpp",
1930 "src/runtime/NEON/functions/NEMatMul.cpp"
1931 ]
1932 }
1933 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001934 "Mul": {
1935 "files": {
1936 "common": [
1937 "src/cpu/operators/CpuMul.cpp",
1938 "src/cpu/kernels/CpuMulKernel.cpp",
1939 "src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp"
Pablo Marquez Tello568aab62023-11-20 14:20:01 +00001940 ],
1941 "neon":{
1942 "fp16":["src/cpu/kernels/mul/generic/neon/fp16.cpp"],
1943 "fp32":["src/cpu/kernels/mul/generic/neon/fp32.cpp"]
1944 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001945 }
1946 },
1947 "Normalize": {
1948 "deps": [ "Mul" ],
1949 "files": {
1950 "common": [
1951 "src/core/NEON/kernels/NENormalizationLayerKernel.cpp",
1952 "src/runtime/NEON/functions/NENormalizationLayer.cpp"
Pablo Marquez Telloba933712023-11-13 15:57:54 +00001953 ],
1954 "neon":{
1955 "fp16":["src/cpu/kernels/norm_layer/generic/neon/fp16.cpp"],
1956 "fp32":["src/cpu/kernels/norm_layer/generic/neon/fp32.cpp"]
1957 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001958 }
1959 },
1960 "Pad": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001961 "deps": [ "Concatenate", "Copy", "StridedSlice" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001962 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001963 "common": [
1964 "src/core/NEON/kernels/NEPadLayerKernel.cpp",
1965 "src/runtime/NEON/functions/NEPadLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001966 ]
1967 }
1968 },
1969 "Permute": {
1970 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001971 "common": [
1972 "src/cpu/operators/CpuPermute.cpp",
1973 "src/cpu/kernels/CpuPermuteKernel.cpp",
1974 "src/runtime/NEON/functions/NEPermute.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001975 ]
1976 }
1977 },
1978 "Pool2d": {
1979 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001980 "common": [
1981 "src/cpu/operators/CpuPool2d.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001982 "src/cpu/kernels/CpuPool2dKernel.cpp",
1983 "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001984 "src/runtime/NEON/functions/NEPoolingLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001985 ],
1986 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001987 "common": [
1988 "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001989 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp",
1990 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp",
1991 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp",
1992 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
1993 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001994 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1995 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
1996 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1997 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
1998 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
1999 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2000 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
2001 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
2002 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
2003 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
2004 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2005 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
2006 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01002007 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02002008 ],
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002009 "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
Pablo Marquez Tello01b0f9b2023-11-07 15:42:46 +00002010 "fp16": [
2011 "src/cpu/kernels/pool2d/neon/fp16.cpp",
2012 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp",
2013 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
2014 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
2015 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2016 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp"
2017 ],
ramelg01a1f78512022-06-29 16:28:10 +01002018 "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ],
2019 "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
Dana Zlotnika538ae52022-02-21 13:12:41 +02002020 "qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002021 },
2022 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002023 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +01002024 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
2025 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
2026 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2027 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002028 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
2029 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
2030 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2031 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_generic_depthfirst/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01002032 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_avg_generic_depthfirst/generic.cpp",
2033 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2034 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp",
2035 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
2036 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp",
2037 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp",
2038 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2039 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp",
2040 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01002041 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp",
2042 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
2043 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
2044 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2045 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_generic_depthfirst/generic.cpp",
2046 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
2047 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
2048 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2049 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp",
2050 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_avg_generic_depthfirst/generic.cpp",
2051 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2052 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_generic_depthfirst/generic.cpp",
2053 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
2054 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp",
2055 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_avg_generic_depthfirst/generic.cpp",
2056 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
2057 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_generic_depthfirst/generic.cpp",
2058 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
2059 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002060 ]
2061 }
2062 }
2063 },
Adnan AlSinan171fc3d2022-03-15 18:46:42 +00002064 "Pool3d": {
2065 "files": {
2066 "common": [
2067 "src/cpu/operators/CpuPool3d.cpp",
2068 "src/cpu/kernels/CpuPool3dKernel.cpp",
2069 "src/runtime/NEON/functions/NEPooling3dLayer.cpp"
2070 ],
2071 "neon": {
Adnan AlSinan9104cd52022-04-06 16:19:31 +01002072 "fp16": [ "src/cpu/kernels/pool3d/neon/fp16.cpp" ],
2073 "fp32": [ "src/cpu/kernels/pool3d/neon/fp32.cpp" ],
2074 "qasymm8": [ "src/cpu/kernels/pool3d/neon/qasymm8.cpp" ],
2075 "qasymm8_signed": [ "src/cpu/kernels/pool3d/neon/qasymm8_signed.cpp" ]
Adnan AlSinan171fc3d2022-03-15 18:46:42 +00002076 }
2077 }
2078 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002079 "PRelu": {
2080 "deps": [ "ElementwiseBinary" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01002081 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002082 "common": [
2083 "src/runtime/NEON/functions/NEPReluLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002084 ]
2085 }
2086 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002087 "PriorBox": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01002088 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002089 "common": [
2090 "src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp",
2091 "src/runtime/NEON/functions/NEPriorBoxLayer.cpp"
2092 ]
2093 }
2094 },
2095 "Quantize": {
2096 "files": {
2097 "common": [
2098 "src/cpu/operators/CpuQuantize.cpp",
2099 "src/cpu/kernels/CpuQuantizeKernel.cpp",
Ramy Elgammald3d2e9b2024-05-15 22:23:46 +01002100 "src/runtime/NEON/functions/NEQuantizationLayer.cpp"
Ramy Elgammalb4b61a62024-05-14 15:21:07 +01002101 ],
2102 "neon":{
2103 "fp32":["src/cpu/kernels/quantize/generic/neon/fp32.cpp"],
2104 "fp16":["src/cpu/kernels/quantize/generic/neon/fp16.cpp"],
2105 "integer":["src/cpu/kernels/quantize/generic/neon/integer.cpp"]
2106 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002107 }
2108 },
2109 "Range": {
2110 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002111 "common": [
2112 "src/core/NEON/kernels/NERangeKernel.cpp",
2113 "src/runtime/NEON/functions/NERange.cpp"
Yair Schwarzbaum9e8a7702021-11-08 10:58:06 +02002114 ],
2115 "neon": {
Yair Schwarzbaum9e8a7702021-11-08 10:58:06 +02002116 "fp32": [ "src/cpu/kernels/range/generic/neon/fp32.cpp" ],
2117 "fp16": [ "src/cpu/kernels/range/generic/neon/fp16.cpp" ],
2118 "integer": [ "src/cpu/kernels/range/generic/neon/integer.cpp" ]
2119 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002120 }
2121 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002122 "Reduction":{
2123 "deps": [ "Reshape" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01002124 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002125 "common": [
2126 "src/core/NEON/kernels/NEReductionOperationKernel.cpp",
2127 "src/runtime/NEON/functions/NEReductionOperation.cpp"
Ramy Elgammal21fb2ad2024-05-13 11:12:11 +01002128 ],
2129 "neon":{
2130 "fp32":["src/cpu/kernels/reduction_layer/generic/neon/fp32.cpp"],
2131 "fp16":["src/cpu/kernels/reduction_layer/generic/neon/fp16.cpp"],
2132 "integer":["src/cpu/kernels/reduction_layer/generic/neon/integer.cpp"],
2133 "qasymm8":["src/cpu/kernels/reduction_layer/generic/neon/qasymm8.cpp"],
2134 "qasymm8_signed":["src/cpu/kernels/reduction_layer/generic/neon/qasymm8_signed.cpp"]
2135 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002136 }
2137 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002138 "Reorg": {
2139 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002140 "common": [
2141 "src/core/NEON/kernels/NEReorgLayerKernel.cpp",
2142 "src/runtime/NEON/functions/NEReorgLayer.cpp"
2143 ]
2144 }
2145 },
2146 "Reshape": {
2147 "files": {
2148 "common": [
2149 "src/cpu/operators/CpuReshape.cpp",
2150 "src/cpu/kernels/CpuReshapeKernel.cpp",
2151 "src/runtime/NEON/functions/NEReshapeLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002152 ]
2153 }
2154 },
David Svantesson3b162e52023-03-28 14:13:32 +00002155 "Reorder": {
2156 "files": {
2157 "common": [
David Svantessonf0ff76d2023-04-28 16:34:52 +00002158 "src/runtime/NEON/functions/NEReorderLayer.cpp",
David Svantesson3b162e52023-03-28 14:13:32 +00002159 "src/core/NEON/kernels/NEReorderKernel.cpp"
2160 ]
2161 }
2162 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002163 "Reverse": {
2164 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002165 "common": [
2166 "src/core/NEON/kernels/NEReverseKernel.cpp",
2167 "src/runtime/NEON/functions/NEReverse.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002168 ]
2169 }
2170 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002171 "RNN": {
2172 "deps": [ "Activation", "Add", "FullyConnected", "Gemm"],
2173 "files": {
2174 "common": [ "src/runtime/NEON/functions/NERNNLayer.cpp" ]
2175 }
2176 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002177 "ROIAlign": {
2178 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002179 "common": [
2180 "src/core/NEON/kernels/NEROIAlignLayerKernel.cpp",
2181 "src/runtime/NEON/functions/NEROIAlignLayer.cpp"
Dana Zlotnik9ae5a4e2022-01-03 16:22:47 +02002182 ],
2183 "neon":{
Dana Zlotnik9ae5a4e2022-01-03 16:22:47 +02002184 "fp32":["src/cpu/kernels/roialign/generic/neon/fp32.cpp"],
2185 "fp16":["src/cpu/kernels/roialign/generic/neon/fp16.cpp"],
2186 "qasymm8":["src/cpu/kernels/roialign/generic/neon/qasymm8.cpp"],
2187 "qasymm8_signed":["src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp"]
2188 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002189 }
2190 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002191 "ROIPool2d": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01002192 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002193 "common": [
2194 "src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp",
2195 "src/runtime/NEON/functions/NEROIPoolingLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002196 ]
2197 }
2198 },
2199 "Scale": {
2200 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002201 "common": [
2202 "src/cpu/operators/CpuScale.cpp",
2203 "src/cpu/kernels/CpuScaleKernel.cpp",
2204 "src/runtime/NEON/functions/NEScale.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002205 ],
2206 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002207 "fp16": [ "src/cpu/kernels/scale/sve/fp16.cpp" ],
2208 "fp32": [ "src/cpu/kernels/scale/sve/fp32.cpp" ],
2209 "integer": [ "src/cpu/kernels/scale/sve/integer.cpp" ],
2210 "qasymm8": [ "src/cpu/kernels/scale/sve/qasymm8.cpp" ],
2211 "qasymm8_signed": [ "src/cpu/kernels/scale/sve/qasymm8_signed.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002212 },
2213 "neon": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002214 "fp16": [ "src/cpu/kernels/scale/neon/fp16.cpp" ],
2215 "integer": [ "src/cpu/kernels/scale/neon/integer.cpp" ],
Gunes Bayirc4f27432022-09-11 15:59:19 +01002216 "qasymm8": [ "src/cpu/kernels/scale/neon/qasymm8.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ],
2217 "qasymm8_signed": [ "src/cpu/kernels/scale/neon/qasymm8_signed.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002218 }
2219 }
2220 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002221 "Select": {
2222 "files": {
2223 "common": [
2224 "src/core/NEON/kernels/NESelectKernel.cpp",
2225 "src/runtime/NEON/functions/NESelect.cpp"
Anton Vainer8a9a0fb2022-01-09 14:37:12 +02002226 ],
2227 "neon": {
Anton Vainer8a9a0fb2022-01-09 14:37:12 +02002228 "fp32": [ "src/cpu/kernels/select/generic/neon/fp32.cpp" ],
2229 "fp16": [ "src/cpu/kernels/select/generic/neon/fp16.cpp" ],
2230 "integer": [ "src/cpu/kernels/select/generic/neon/integer.cpp" ]
2231 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002232 }
2233 },
2234 "Slice": {
2235 "deps": [ "StridedSlice" ],
2236 "files": {
2237 "common": [ "src/runtime/NEON/functions/NESlice.cpp" ]
2238 }
2239 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002240 "Softmax": {
2241 "deps": [
2242 "Permute"
2243 ],
2244 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002245 "common": [
2246 "src/cpu/operators/CpuSoftmax.cpp",
2247 "src/cpu/kernels/CpuSoftmaxKernel.cpp",
2248 "src/runtime/NEON/functions/NESoftmaxLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002249 ],
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002250 "neon":{
Dana Zlotnika538ae52022-02-21 13:12:41 +02002251 "common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"],
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002252 "fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"],
ramelg01a1f78512022-06-29 16:28:10 +01002253 "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"],
2254 "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02002255 "qasymm8_signed":["src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp"]
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002256 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002257 "sve": {
Gunes Bayirfadc9b12023-11-07 05:43:07 +00002258 "common": [ "src/cpu/kernels/softmax/generic/sve/impl.cpp" ]
Dana Zlotnik6a2df882022-01-17 09:54:26 +02002259 },
2260 "sve2":{
Viet-Hoa Do77bbe2e2023-12-06 11:01:15 +00002261 "common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"],
Gunes Bayircfca87b2024-04-09 23:13:04 +01002262 "fp32" :["src/cpu/kernels/softmax/generic/sme2/fp32.cpp"],
Omar Al Khatibc1575b22024-04-23 16:26:56 +01002263 "fp16" :["src/cpu/kernels/softmax/generic/sme2/fp16.cpp"],
2264 "qasymm8" :["src/cpu/kernels/softmax/generic/sme2/qasymm8.cpp"],
2265 "qasymm8_signed" :["src/cpu/kernels/softmax/generic/sme2/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002266 }
2267 }
2268 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002269 "SpaceToBatch": {
2270 "files": {
2271 "common": [
2272 "src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp",
2273 "src/runtime/NEON/functions/NESpaceToBatchLayer.cpp"
2274 ]
2275 }
2276 },
2277 "SpaceToDepth": {
2278 "files": {
2279 "common": [
2280 "src/core/NEON/kernels/NESpaceToDepthLayerKernel.cpp",
2281 "src/runtime/NEON/functions/NESpaceToDepthLayer.cpp"
2282 ]
2283 }
2284 },
2285 "Split": {
2286 "deps": [ "StridedSlice" ],
2287 "files": {
2288 "common": [
2289 "src/runtime/NEON/functions/NESplit.cpp"
2290 ]
2291 }
2292 },
2293 "Stack": {
2294 "files": {
2295 "common": [
2296 "src/core/NEON/kernels/NEStackLayerKernel.cpp",
2297 "src/runtime/NEON/functions/NEStackLayer.cpp"
2298 ]
2299 }
2300 },
2301 "StridedSlice": {
2302 "files": {
2303 "common": [
2304 "src/core/NEON/kernels/NEStridedSliceKernel.cpp",
2305 "src/runtime/NEON/functions/NEStridedSlice.cpp"
2306 ]
2307 }
2308 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002309 "Sub": {
2310 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002311 "common": [
2312 "src/cpu/operators/CpuSub.cpp",
2313 "src/cpu/kernels/CpuSubKernel.cpp",
Pablo Marquez Tello67773592023-10-06 13:49:44 +01002314 "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp"
2315 ],
2316 "neon": {
2317 "fp16":["src/cpu/kernels/sub/neon/fp16.cpp"],
2318 "qasymm8": ["src/cpu/kernels/sub/neon/qasymm8.cpp"],
2319 "qasymm8_signed": ["src/cpu/kernels/sub/neon/qasymm8_signed.cpp"],
2320 "qsymm16": ["src/cpu/kernels/sub/neon/qsymm16.cpp"]
2321 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002322 }
2323 },
Pablo Marquez Tello67773592023-10-06 13:49:44 +01002324
Michalis Spyrou20fca522021-06-07 14:23:57 +01002325 "Tile": {
2326 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002327 "common": [
2328 "src/core/NEON/kernels/NETileKernel.cpp",
2329 "src/runtime/NEON/functions/NETile.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002330 ]
2331 }
2332 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002333 "Transpose": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01002334 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002335 "common": [
2336 "src/cpu/kernels/CpuTransposeKernel.cpp",
2337 "src/cpu/operators/CpuTranspose.cpp",
2338 "src/runtime/NEON/functions/NETranspose.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002339 ]
2340 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002341 },
2342 "Unstack": {
2343 "deps": [ "StridedSlice" ],
2344 "files": {
2345 "common": [ "src/runtime/NEON/functions/NEUnstack.cpp" ]
2346 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002347 }
2348 }
Giorgio Arena232c4522022-03-03 10:09:01 +00002349 },
2350 "experimental": {
SiCong Li23882a92023-06-28 09:49:45 +01002351 "dynamic_fusion": {
2352 "common": [
2353 "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp",
2354 "src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp",
2355 "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp",
2356 "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp",
2357 "src/dynamic_fusion/sketch/attributes/Conv2dAttributes.cpp",
2358 "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp",
Adnan AlSinand9c1d442023-10-06 19:07:48 +01002359 "src/dynamic_fusion/sketch/attributes/MatMulAttributes.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002360 "src/dynamic_fusion/sketch/attributes/Pool2dAttributes.cpp",
2361 "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp",
2362 "src/dynamic_fusion/sketch/attributes/SoftmaxAttributes.cpp",
2363 "src/dynamic_fusion/sketch/attributes/ReshapeAttributes.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002364 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp",
2365 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp",
2366 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.cpp",
2367 "src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp",
2368 "src/dynamic_fusion/sketch/gpu/GpuOperatorGroup.cpp",
2369 "src/dynamic_fusion/sketch/gpu/GpuWorkloadContext.cpp",
2370 "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.cpp",
2371 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.cpp",
2372 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentCast.cpp",
2373 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp",
2374 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp",
Adnan AlSinand9c1d442023-10-06 19:07:48 +01002375 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentMatMul.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002376 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentPool2d.cpp",
2377 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002378 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentReshape.cpp",
2379 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp",
2380 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp",
2381 "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp",
2382 "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp",
2383 "src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp",
2384 "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp",
2385 "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp",
Adnan AlSinand9c1d442023-10-06 19:07:48 +01002386 "src/dynamic_fusion/sketch/gpu/operators/GpuMatMul.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002387 "src/dynamic_fusion/sketch/gpu/operators/GpuMul.cpp",
2388 "src/dynamic_fusion/sketch/gpu/operators/GpuReshape.cpp",
2389 "src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp",
2390 "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp",
2391 "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp",
2392 "src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp",
2393 "src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp",
2394 "src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp",
2395 "src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp",
2396 "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp"
2397 ],
SiCong Li23882a92023-06-28 09:49:45 +01002398 "ckw_driver": [
2399 "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwComponentArgument.cpp",
2400 "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwDriver.cpp",
2401 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwActivation.cpp",
2402 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp",
ramy.elgammal@arm.coma04ae3e2023-07-27 18:23:17 +01002403 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDepthwiseConv2d.cpp",
Jakub Sujake1c96e72023-07-31 13:36:58 +01002404 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwDirectConv2d.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002405 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwElementwiseBinary.cpp",
Adnan AlSinan2e6d6592023-08-21 13:54:27 +01002406 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwPool2d.cpp",
Adnan AlSinand9c1d442023-10-06 19:07:48 +01002407 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwMatMul.cpp",
Gunes Bayir91cb7332023-07-25 17:00:33 +01002408 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwResize.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002409 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwStore.cpp",
Gunes Bayir2b9fa592024-01-17 16:07:03 +00002410 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/CkwHelper.cpp",
2411 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.cpp",
2412 "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/ElementwiseBinary.cpp",
SiCong Li23882a92023-06-28 09:49:45 +01002413 "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwScopedKernelWriter.cpp",
2414 "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwVariableTable.cpp",
2415 "src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.cpp"
2416 ]
2417 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002418 }
ramelg018a164882022-04-07 02:42:52 +01002419}