blob: 5828b43ec12a332ddfd56e071c31af4f93cad955 [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",
17 "src/core/IAccessWindow.cpp",
18 "src/core/IKernel.cpp",
19 "src/core/ITensor.cpp",
20 "src/core/ITensorPack.cpp",
21 "src/core/Rounding.cpp",
22 "src/core/Size2D.cpp",
23 "src/core/SubTensorInfo.cpp",
24 "src/core/TensorInfo.cpp",
25 "src/core/Utils.cpp",
26 "src/core/Validate.cpp",
27 "src/core/Version.cpp",
28 "src/core/helpers/SoftmaxHelpers.cpp",
29 "src/core/helpers/WindowHelpers.cpp",
30 "src/core/utils/AssemblyUtils.cpp",
31 "src/core/utils/ScaleUtils.cpp",
32 "src/core/utils/helpers/fft.cpp",
33 "src/core/utils/helpers/tensor_transform.cpp",
34 "src/core/utils/io/FileHandler.cpp",
35 "src/core/utils/misc/MMappedFile.cpp",
36 "src/core/utils/quantization/AsymmHelpers.cpp",
37 "src/core/CPP/CPPTypes.cpp",
38 "src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp",
39 "src/core/CPP/kernels/CPPNonMaximumSuppressionKernel.cpp",
40 "src/core/CPP/kernels/CPPPermuteKernel.cpp",
41 "src/core/CPP/kernels/CPPTopKVKernel.cpp",
42 "src/core/CPP/kernels/CPPUpsampleKernel.cpp",
43 "src/runtime/Allocator.cpp",
44 "src/runtime/BlobLifetimeManager.cpp",
45 "src/runtime/BlobMemoryPool.cpp",
46 "src/runtime/ISimpleLifetimeManager.cpp",
47 "src/runtime/ITensorAllocator.cpp",
48 "src/runtime/IWeightsManager.cpp",
49 "src/runtime/IScheduler.cpp",
50 "src/runtime/Memory.cpp",
51 "src/runtime/MemoryManagerOnDemand.cpp",
52 "src/runtime/OffsetLifetimeManager.cpp",
53 "src/runtime/OffsetMemoryPool.cpp",
54 "src/runtime/OperatorTensor.cpp",
55 "src/runtime/PoolManager.cpp",
56 "src/runtime/RuntimeContext.cpp",
57 "src/runtime/Scheduler.cpp",
58 "src/runtime/SchedulerFactory.cpp",
59 "src/runtime/SchedulerUtils.cpp",
60 "src/runtime/SubTensor.cpp",
61 "src/runtime/Tensor.cpp",
62 "src/runtime/TensorAllocator.cpp",
63 "src/runtime/Utils.cpp",
64 "src/runtime/CPP/ICPPSimpleFunction.cpp",
65 "src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp",
66 "src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp",
67 "src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp",
68 "src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp",
69 "src/runtime/CPP/functions/CPPPermute.cpp",
70 "src/runtime/CPP/functions/CPPTopKV.cpp",
71 "src/runtime/CPP/functions/CPPUpsample.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +010072 ],
Georgios Pinitasb6af4822021-09-14 12:33:34 +010073 "logging": [
74 "src/core/utils/logging/FilePrinter.cpp",
75 "src/core/utils/logging/Helpers.cpp",
76 "src/core/utils/logging/Logger.cpp",
77 "src/core/utils/logging/LoggerRegistry.cpp"
78 ],
79 "scheduler": {
80 "single": [ "src/runtime/CPP/SingleThreadScheduler.cpp" ],
81 "threads": [ "src/runtime/CPP/CPPScheduler.cpp" ],
82 "omp": [ "src/runtime/OMP/OMPScheduler.cpp"]
83 },
Michalis Spyrou20fca522021-06-07 14:23:57 +010084 "c_api": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +010085 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +010086 "src/c/AclContext.cpp",
Georgios Pinitas06ac6e42021-07-05 08:08:52 +010087 "src/c/AclOperator.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +010088 "src/c/AclQueue.cpp",
89 "src/c/AclTensor.cpp",
90 "src/c/AclTensorPack.cpp",
91 "src/c/AclVersion.cpp"
Georgios Pinitasbdcdc392021-04-22 16:42:03 +010092 ],
Georgios Pinitas13ef1762021-07-14 17:14:43 +010093 "cpu": [
94 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +010095 "gpu": [
96 "src/c/cl/AclOpenClExt.cpp"
Georgios Pinitas41648142021-08-03 08:24:00 +010097 ],
98 "operators":
99 [
100 "src/c/operators/AclActivation.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100101 ]
102 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100103 "high_priority": [
104 "Activation",
105 "DepthwiseConv2d",
106 "Conv2d",
107 "Permute",
108 "Pool2d",
109 "Reshape"
110 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100111 "gpu": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100112 "common": [
113 "src/core/CL/CLCompileContext.cpp",
Giorgio Arenaba2dd822021-07-28 16:10:03 +0100114 "src/core/CL/DefaultLWSHeuristics.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100115 "src/core/CL/CLHelpers.cpp",
116 "src/core/CL/CLKernelLibrary.cpp",
117 "src/core/CL/CLUtils.cpp",
118 "src/core/CL/ICLKernel.cpp",
119 "src/core/CL/ICLSimple2DKernel.cpp",
120 "src/core/CL/ICLSimple3DKernel.cpp",
121 "src/core/CL/ICLSimpleKernel.cpp",
122 "src/core/CL/ICLTensor.cpp",
123 "src/core/CL/OpenCL.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100124 "src/gpu/cl/ClContext.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100125 "src/gpu/cl/ClKernelLibrary.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100126 "src/gpu/cl/ClQueue.cpp",
127 "src/gpu/cl/ClTensor.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100128 "src/core/CL/kernels/CLFillBorderKernel.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100129 "src/runtime/CL/CLBufferAllocator.cpp",
130 "src/runtime/CL/CLGEMMHeuristicsHandle.cpp",
131 "src/runtime/CL/CLHelpers.cpp",
132 "src/runtime/CL/CLMemory.cpp",
133 "src/runtime/CL/CLMemoryRegion.cpp",
134 "src/runtime/CL/CLOperator.cpp",
135 "src/runtime/CL/CLRuntimeContext.cpp",
136 "src/runtime/CL/CLScheduler.cpp",
137 "src/runtime/CL/CLSubTensor.cpp",
138 "src/runtime/CL/CLTensor.cpp",
139 "src/runtime/CL/CLTensorAllocator.cpp",
140 "src/runtime/CL/CLTuner.cpp",
141 "src/runtime/CL/ICLSimpleFunction.cpp",
142 "src/runtime/CL/Utils.cpp",
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100143 "src/runtime/CL/mlgo/HeuristicTree.cpp",
144 "src/runtime/CL/mlgo/MLGOHeuristics.cpp",
145 "src/runtime/CL/mlgo/MLGOParser.cpp",
146 "src/runtime/CL/mlgo/Utils.cpp",
147 "src/runtime/CL/tuners/CLTuningParametersList.cpp"
148 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100149 "operators": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100150 "Activation":{
151 "files": {
152 "common": [
153 "src/gpu/cl/kernels/ClActivationKernel.cpp",
154 "src/gpu/cl/operators/ClActivation.cpp",
155 "src/runtime/CL/functions/CLActivationLayer.cpp"
156 ]
157 }
158 },
159 "ArgMinMax": {
160 "deps": [ "Reshape" ],
161 "files": {
162 "common": [
163 "src/core/CL/kernels/CLArgMinMaxLayerKernel.cpp",
164 "src/runtime/CL/functions/CLArgMinMaxLayer.cpp"
165 ]
166 }
167 },
168 "Add": {
169 "files": {
170 "common": [
171 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
172 "src/gpu/cl/operators/ClAdd.cpp"
173 ]
174 }
175 },
176 "BatchNormalization": {
177 "files": {
178 "common": [
179 "src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp",
180 "src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp",
181 "src/runtime/CL/functions/CLBatchNormalizationLayer.cpp",
182 "src/runtime/CL/functions/CLFuseBatchNormalization.cpp"
183 ]
184 }
185 },
186 "BatchToSpace": {
187 "files": {
188 "common": [
189 "src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp",
190 "src/runtime/CL/functions/CLBatchToSpaceLayer.cpp"
191 ]
192 }
193 },
194 "Bitwise": {
195 "files": {
196 "common": [ "src/core/CL/kernels/CLBitwiseKernel.cpp" ]
197 }
198 },
199 "BitwiseAnd": {
200 "deps": [ "Bitwise" ],
201 "files": {
202 "common": [ "src/runtime/CL/functions/CLBitwiseAnd.cpp" ]
203 }
204 },
205 "BitwiseNot": {
206 "deps": [ "Bitwise" ],
207 "files": {
208 "common": [ "src/runtime/CL/functions/CLBitwiseNot.cpp" ]
209 }
210 },
211 "BitwiseOr": {
212 "deps": [ "Bitwise" ],
213 "files": {
214 "common": [ "src/runtime/CL/functions/CLBitwiseOr.cpp" ]
215 }
216 },
217 "BitwiseXor": {
218 "deps": [ "Bitwise" ],
219 "files": {
220 "common": [ "src/runtime/CL/functions/CLBitwiseXor.cpp" ]
221 }
222 },
223 "BoundingBoxTransform": {
224 "files": {
225 "common": [
226 "src/core/CL/kernels/CLBoundingBoxTransformKernel.cpp",
227 "src/runtime/CL/functions/CLBoundingBoxTransform.cpp"
228 ]
229 }
230 },
231 "Cast": {
232 "files": {
233 "common": [
234 "src/gpu/cl/kernels/ClCastKernel.cpp",
235 "src/gpu/cl/operators/ClCast.cpp",
236 "src/runtime/CL/functions/CLCast.cpp"
237 ]
238 }
239 },
240 "ChannelShuffle": {
241 "files": {
242 "common": [
243 "src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp",
244 "src/runtime/CL/functions/CLChannelShuffleLayer.cpp"
245 ]
246 }
247 },
248 "Comparison": {
249 "files": {
250 "common": [
251 "src/core/CL/kernels/CLComparisonKernel.cpp",
252 "src/runtime/CL/functions/CLComparison.cpp"
253 ]
254 }
255 },
256 "Concatenate": {
257 "files": {
258 "common": [
259 "src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp",
260 "src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp",
261 "src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp",
262 "src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp",
263 "src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp",
264 "src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp",
265 "src/gpu/cl/operators/ClConcatenate.cpp",
266 "src/runtime/CL/functions/CLConcatenateLayer.cpp"
267 ]
268 }
269 },
270 "Conv2d": {
271 "deps": [
272 "Activation",
273 "ElementwiseBinary",
274 "FFT2D",
275 "Gemm",
276 "Mul",
277 "Pad",
278 "Permute",
279 "Reduction",
280 "Reshape",
281 "Reverse",
282 "Slice"
283 ],
284 "files": {
285 "common": [
286 "src/gpu/cl/kernels/ClDirectConv2dKernel.cpp",
287 "src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp",
288 "src/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp",
289 "src/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp",
290 "src/gpu/cl/kernels/ClIm2ColKernel.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000291 "src/gpu/cl/kernels/ClIndirectConv2dKernel.cpp",
Gian Marco Iodice5d016812022-11-17 11:03:39 +0000292 "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100293 "src/gpu/cl/kernels/ClCol2ImKernel.cpp",
294 "src/gpu/cl/operators/ClConv2d.cpp",
295 "src/gpu/cl/operators/ClDirectConv2d.cpp",
296 "src/gpu/cl/operators/ClGemmConv2d.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000297 "src/gpu/cl/operators/ClIndirectConv2d.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100298 "src/gpu/cl/operators/ClWinogradConv2d.cpp",
299 "src/gpu/cl/kernels/ClWeightsReshapeKernel.cpp",
300 "src/runtime/CL/functions/CLConvolutionLayer.cpp",
301 "src/runtime/CL/functions/CLDirectConvolutionLayer.cpp",
302 "src/runtime/CL/functions/CLFFTConvolutionLayer.cpp",
303 "src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp",
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000304 "src/runtime/CL/functions/CLIndirectConvolutionLayer.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100305 "src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp"
306 ]
307 }
308 },
Giorgio Arena273c28c2021-10-14 15:59:15 +0100309 "Conv3d": {
310 "files": {
311 "common": [
312 "src/gpu/cl/kernels/ClDirectConv3dKernel.cpp",
313 "src/gpu/cl/operators/ClDirectConv3d.cpp",
314 "src/runtime/CL/functions/CLConv3D.cpp"
315 ]
316 }
317 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100318 "Copy": {
319 "files": {
320 "common": [
321 "src/gpu/cl/kernels/ClCopyKernel.cpp",
322 "src/gpu/cl/operators/ClCopy.cpp",
323 "src/runtime/CL/functions/CLCopy.cpp"
324 ]
325 }
326 },
327 "CropResize": {
328 "deps": [ "Copy", "Fill", "Scale" ],
329 "files": {
330 "common": [
331 "src/gpu/cl/kernels/ClCropKernel.cpp",
332 "src/gpu/cl/operators/ClCrop.cpp",
333 "src/runtime/CL/functions/CLCrop.cpp",
334 "src/runtime/CL/functions/CLCropResize.cpp"
335 ]
336 }
337 },
338 "Deconv2d": {
339 "deps": [ "Conv2d", "Reverse", "Transpose"],
340 "files": {
341 "common": [
342 "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp",
343 "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp",
Gunes Bayirec0113d2022-11-09 09:26:27 +0000344 "src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp",
345 "src/gpu/cl/operators/ClTransposedConvolution.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100346 "src/runtime/CL/functions/CLDeconvolutionLayer.cpp",
347 "src/runtime/CL/functions/CLDeconvolutionLayerUpsample.cpp",
348 "src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp",
349 "src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp"
350 ]
351 }
352 },
353 "DepthConvert": {
354 "deps": [ "Cast"],
355 "files": {
356 "common": [ "src/runtime/CL/functions/CLDepthConvertLayer.cpp" ]
357 }
358 },
359 "DepthToSpace": {
360 "files": {
361 "common": [
362 "src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp",
363 "src/runtime/CL/functions/CLDepthToSpaceLayer.cpp"
364 ]
365 }
366 },
367 "DepthwiseConv2d": {
368 "deps": [ "Permute" ],
369 "files": {
370 "common": [
371 "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp",
372 "src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp"
373 ]
374 }
375 },
376 "Dequantize": {
377 "files": {
378 "common": [
379 "src/gpu/cl/kernels/ClDequantizeKernel.cpp",
380 "src/gpu/cl/operators/ClDequantize.cpp",
381 "src/runtime/CL/functions/CLDequantizationLayer.cpp"
382 ]
383 }
384 },
385 "ElementwiseBinary": {
386 "deps": ["Add", "Sub"],
387 "files": {
388 "common": [
389 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
390 "src/gpu/cl/operators/ClElementwiseOperations.cpp",
391 "src/runtime/CL/functions/CLElementwiseOperations.cpp"
392 ]
393 }
394 },
395 "ElementwiseUnary":{
396 "files": {
397 "common": [
398 "src/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp",
399 "src/gpu/cl/operators/ClElementwiseUnary.cpp",
400 "src/runtime/CL/functions/CLElementwiseUnaryLayer.cpp"
401 ]
402 }
403 },
404 "FFT1D": {
405 "files": {
406 "common": [
407 "src/core/CL/kernels/CLFFTDigitReverseKernel.cpp",
408 "src/core/CL/kernels/CLFFTRadixStageKernel.cpp",
409 "src/core/CL/kernels/CLFFTScaleKernel.cpp",
410 "src/runtime/CL/functions/CLFFT1D.cpp"
411 ]
412 }
413 },
414 "FFT2D": {
415 "deps": [ "FFT1D" ],
416 "files": {
417 "common": [ "src/runtime/CL/functions/CLFFT2D.cpp" ]
418 }
419 },
420 "Fill": {
421 "files": {
422 "common": [
423 "src/gpu/cl/kernels/ClFillKernel.cpp",
424 "src/gpu/cl/operators/ClFill.cpp",
425 "src/runtime/CL/functions/CLFill.cpp"
426 ]
427 }
428 },
429 "Flatten": {
430 "files": {
431 "common": [
432 "src/gpu/cl/operators/ClFlatten.cpp",
433 "src/runtime/CL/functions/CLFlattenLayer.cpp"
434 ]
435 }
436 },
437 "Floor": {
438 "files": {
439 "common": [
440 "src/gpu/cl/kernels/ClFloorKernel.cpp",
441 "src/gpu/cl/operators/ClFloor.cpp",
442 "src/runtime/CL/functions/CLFloor.cpp"
443 ]
444 }
445 },
446 "FullyConnected": {
447 "deps": [ "Flatten", "Gemm", "Transpose"],
448 "files": {
449 "common": [
450 "src/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp",
451 "src/gpu/cl/operators/ClConvertFullyConnectedWeights.cpp",
452 "src/gpu/cl/operators/ClFullyConnected.cpp",
453 "src/runtime/CL/functions/CLConvertFullyConnectedWeights.cpp",
454 "src/runtime/CL/functions/CLFullyConnectedLayer.cpp"
455 ]
456 }
457 },
458 "Gather": {
459 "files": {
460 "common": [
461 "src/core/CL/kernels/CLGatherKernel.cpp",
462 "src/runtime/CL/functions/CLGather.cpp"]
463 }
464 },
465 "Gemm": {
466 "deps": [ "Cast" ],
467 "files": {
468 "common": [
469 "src/gpu/cl/kernels/gemm/ClGemmHelpers.cpp",
470 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeBifrost.cpp",
471 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeMidgard.cpp",
472 "src/gpu/cl/kernels/gemm/native/ClGemmDefaultConfigNativeValhall.cpp",
473 "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedBifrost.cpp",
474 "src/gpu/cl/kernels/gemm/reshaped/ClGemmDefaultConfigReshapedValhall.cpp",
475 "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyBifrost.cpp",
476 "src/gpu/cl/kernels/gemm/reshaped_only_rhs/ClGemmDefaultConfigReshapedRhsOnlyValhall.cpp",
477 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp",
478 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp",
479 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.cpp",
Freddie Liardete572dff2022-05-16 14:09:10 +0100480 "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100481 "src/gpu/cl/kernels/ClGemmLowpOffsetContributionKernel.cpp",
482 "src/gpu/cl/kernels/ClGemmLowpOffsetContributionOutputStageKernel.cpp",
483 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.cpp",
484 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp",
485 "src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100486 "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp",
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000487 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100488 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.cpp",
489 "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp",
490 "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.cpp",
491 "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.cpp",
492 "src/gpu/cl/kernels/ClGemmLowpReductionKernel.cpp",
493 "src/gpu/cl/operators/ClGemm.cpp",
494 "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.cpp",
495 "src/gpu/cl/operators/ClGemmLowpOutputStage.cpp",
496 "src/runtime/CL/gemm/CLGEMMDefaultTypeBifrost.cpp",
497 "src/runtime/CL/gemm/CLGEMMDefaultTypeMidgard.cpp",
498 "src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp",
499 "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp",
500 "src/runtime/CL/functions/CLGEMM.cpp",
501 "src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp",
Ramy Elgammaldf6a3b02022-11-30 16:23:10 +0000502 "src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp",
503 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.cpp",
Gian Marco Iodicea5cb79f2022-12-28 13:53:51 +0000504 "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.cpp",
505 "src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp"
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100506 ]
507 }
508 },
509 "GenerateProposals": {
510 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ],
511 "files": {
512 "common": [
513 "src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp",
514 "src/runtime/CL/functions/CLGenerateProposalsLayer.cpp"
515 ]
516 }
517 },
518 "InstanceNormalize": {
519 "files": {
520 "common": [
521 "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp",
522 "src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp"
523 ]
524 }
525 },
526 "L2Normalize": {
527 "deps": [ "Reduction" ],
528 "files": {
529 "common": [
530 "src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp",
531 "src/runtime/CL/functions/CLL2NormalizeLayer.cpp"
532 ]
533 }
534 },
535 "Logical": {
536 "files": {
537 "common": [
538 "src/gpu/cl/operators/ClLogicalNot.cpp",
539 "src/runtime/CL/functions/CLLogicalAnd.cpp",
540 "src/runtime/CL/functions/CLLogicalNot.cpp",
541 "src/runtime/CL/functions/CLLogicalOr.cpp"
542 ]
543 }
544 },
545 "LSTM": {
546 "deps": [
547 "Activation",
548 "Concatenate",
549 "Copy",
550 "Dequantize",
551 "ElementwiseBinary",
552 "Fill",
553 "FullyConnected",
554 "Gemm",
555 "MeanStdDevNormalize",
556 "Mul",
557 "Quantize",
558 "Slice",
559 "Transpose"
560 ],
561 "files": {
562 "common": [
563 "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.cpp",
564 "src/runtime/CL/functions/CLQLSTMLayer.cpp",
565 "src/runtime/CL/functions/CLLSTMLayer.cpp",
566 "src/runtime/CL/functions/CLLSTMLayerQuantized.cpp"
567 ]
568 }
569 },
570 "MaxUnpool2d": {
571 "deps": [ "Fill" ],
572 "files": {
573 "common": [
574 "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.cpp",
575 "src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp"
576 ]
577 }
578 },
579 "MeanStdDevNormalize": {
580 "deps": [ "Reduction" ],
581 "files": {
582 "common": [
583 "src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp",
584 "src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp",
585 "src/runtime/CL/functions/CLReduceMean.cpp"
586 ]
587 }
588 },
589 "Mul": {
590 "files": {
591 "common": [
592 "src/gpu/cl/kernels/ClMulKernel.cpp",
593 "src/gpu/cl/operators/ClMul.cpp",
594 "src/runtime/CL/functions/CLPixelWiseMultiplication.cpp"
595 ]
596 }
597 },
598 "Normalize": {
599 "files": {
600 "common": [
601 "src/core/CL/kernels/CLNormalizationLayerKernel.cpp",
602 "src/runtime/CL/functions/CLNormalizationLayer.cpp"
603 ]
604 }
605 },
606 "Pad": {
607 "deps": [ "Copy" ],
608 "files": {
609 "common": [
610 "src/core/CL/kernels/CLPadLayerKernel.cpp",
611 "src/runtime/CL/functions/CLPadLayer.cpp"
612 ]
613 }
614 },
615 "Permute": {
616 "files": {
617 "common": [
618 "src/gpu/cl/kernels/ClPermuteKernel.cpp",
619 "src/gpu/cl/operators/ClPermute.cpp",
620 "src/runtime/CL/functions/CLPermute.cpp"
621 ]
622 }
623 },
624 "Pool2d": {
625 "files": {
626 "common": [
627 "src/gpu/cl/kernels/ClPool2dKernel.cpp",
628 "src/gpu/cl/operators/ClPool2d.cpp",
629 "src/runtime/CL/functions/CLPoolingLayer.cpp"
630 ]
631 }
632 },
ramelg0137515692022-02-26 22:06:20 +0000633 "Pool3d": {
634 "files": {
635 "common": [
636 "src/gpu/cl/kernels/ClPool3dKernel.cpp",
637 "src/gpu/cl/operators/ClPool3d.cpp",
638 "src/runtime/CL/functions/CLPooling3dLayer.cpp"
639 ]
640 }
641 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100642 "PRelu": {
643 "deps": [ "ElementwiseBinary" ],
644 "files": {
645 "common": [
646 "src/gpu/cl/operators/ClPRelu.cpp",
647 "src/runtime/CL/functions/CLPReluLayer.cpp"
648 ]
649 }
650 },
651 "PriorBox": {
652 "files": {
653 "common": [
654 "src/core/CL/kernels/CLPriorBoxLayerKernel.cpp",
655 "src/runtime/CL/functions/CLPriorBoxLayer.cpp"
656 ]
657 }
658 },
659 "Quantize": {
660 "files": {
661 "common": [
662 "src/gpu/cl/kernels/ClQuantizeKernel.cpp",
663 "src/gpu/cl/operators/ClQuantize.cpp",
664 "src/runtime/CL/functions/CLQuantizationLayer.cpp"
665 ]
666 }
667 },
668 "Range": {
669 "files": {
670 "common": [
671 "src/core/CL/kernels/CLRangeKernel.cpp",
672 "src/runtime/CL/functions/CLRange.cpp"
673 ]
674 }
675 },
676 "Reduction": {
677 "deps": [ "Reshape" ],
678 "files": {
ramelg01a1f78512022-06-29 16:28:10 +0100679 "common": [
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100680 "src/core/CL/kernels/CLReductionOperationKernel.cpp",
681 "src/runtime/CL/functions/CLReductionOperation.cpp"
682 ]
683 }
684 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100685 "Reorg": {
686 "files": {
687 "common": [
688 "src/core/CL/kernels/CLReorgLayerKernel.cpp",
689 "src/runtime/CL/functions/CLReorgLayer.cpp"
690 ]
691 }
692 },
693 "Reshape": {
694 "files": {
695 "common": [
696 "src/gpu/cl/kernels/ClReshapeKernel.cpp",
697 "src/gpu/cl/operators/ClReshape.cpp",
698 "src/runtime/CL/functions/CLReshapeLayer.cpp"
699 ]
700 }
701 },
702 "Reverse": {
703 "files": {
704 "common": [
705 "src/core/CL/kernels/CLReverseKernel.cpp",
706 "src/runtime/CL/functions/CLReverse.cpp"
707 ]
708 }
709 },
710 "RNN": {
711 "deps": [ "Activation", "Cast", "ElementwiseBinary", "FullyConnected", "Gemm"],
712 "files": {
713 "common": [ "src/runtime/CL/functions/CLRNNLayer.cpp" ]
714 }
715 },
716 "ROIAlign": {
717 "files": {
718 "common": [
719 "src/core/CL/kernels/CLROIAlignLayerKernel.cpp",
720 "src/runtime/CL/functions/CLROIAlignLayer.cpp"
721 ]
722 }
723 },
724 "ROIPool2d": {
725 "files": {
726 "common": [
727 "src/core/CL/kernels/CLROIPoolingLayerKernel.cpp",
728 "src/runtime/CL/functions/CLROIPoolingLayer.cpp"
729 ]
730 }
731 },
732 "Scale": {
733 "files": {
734 "common": [
735 "src/gpu/cl/kernels/ClScaleKernel.cpp",
736 "src/gpu/cl/operators/ClScale.cpp",
737 "src/runtime/CL/functions/CLScale.cpp"
738 ]
739 }
740 },
741 "Select": {
742 "files": {
743 "common": [
744 "src/core/CL/kernels/CLSelectKernel.cpp",
745 "src/runtime/CL/functions/CLSelect.cpp"
746 ]
747 }
748 },
749 "Slice": {
750 "deps": [ "StridedSlice" ],
751 "files": {
752 "common": [ "src/runtime/CL/functions/CLSlice.cpp" ]
753 }
754 },
755 "Softmax": {
756 "deps": [ "Permute" ],
757 "files": {
758 "common": [
759 "src/gpu/cl/kernels/ClSoftmaxKernel.cpp",
760 "src/gpu/cl/operators/ClSoftmax.cpp",
761 "src/runtime/CL/functions/CLSoftmaxLayer.cpp"
762 ]
763 }
764 },
765 "SpaceToBatch": {
766 "files": {
767 "common": [
768 "src/core/CL/kernels/CLSpaceToBatchLayerKernel.cpp",
769 "src/runtime/CL/functions/CLSpaceToBatchLayer.cpp"
770 ]
771 }
772 },
773 "SpaceToDepth": {
774 "files": {
775 "common": [
776 "src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp",
777 "src/runtime/CL/functions/CLSpaceToDepthLayer.cpp"
778 ]
779 }
780 },
781 "Split": {
782 "deps": [ "StridedSlice" ],
783 "files": {
784 "common": [ "src/runtime/CL/functions/CLSplit.cpp" ]
785 }
786 },
787 "Stack": {
788 "files": {
789 "common": [
790 "src/core/CL/kernels/CLStackLayerKernel.cpp",
791 "src/runtime/CL/functions/CLStackLayer.cpp"
792 ]
793 }
794 },
795 "StridedSlice": {
796 "files": {
797 "common": [
798 "src/core/CL/kernels/CLStridedSliceKernel.cpp",
799 "src/runtime/CL/functions/CLStridedSlice.cpp"
800 ]
801 }
802 },
803 "Sub": {
804 "files": {
805 "common": [
806 "src/gpu/cl/kernels/ClElementwiseKernel.cpp",
807 "src/gpu/cl/operators/ClSub.cpp"
808 ]
809 }
810 },
811 "Tile": {
812 "files": {
813 "common": [
814 "src/core/CL/kernels/CLTileKernel.cpp",
815 "src/runtime/CL/functions/CLTile.cpp"
816 ]
817 }
818 },
819 "Transpose": {
820 "files": {
821 "common": [
822 "src/gpu/cl/kernels/ClTransposeKernel.cpp",
823 "src/gpu/cl/operators/ClTranspose.cpp",
824 "src/runtime/CL/functions/CLTranspose.cpp"
825 ]
826 }
827 },
828 "Unstack": {
829 "deps": [ "StridedSlice" ],
830 "files": {
831 "common": [ "src/runtime/CL/functions/CLUnstack.cpp" ]
832 }
833 },
834 "YUVNormalize": {
835 "files": {
836 "common": [
837 "src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp",
838 "src/runtime/CL/functions/CLNormalizePlanarYUVLayer.cpp"
839 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100840 }
Georgios Pinitasbdcdc392021-04-22 16:42:03 +0100841 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100842 }
843},
Michalis Spyrou20fca522021-06-07 14:23:57 +0100844 "cpu": {
Georgios Pinitas13ef1762021-07-14 17:14:43 +0100845 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +0100846 "src/cpu/CpuContext.cpp",
847 "src/cpu/CpuQueue.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100848 "src/cpu/CpuTensor.cpp",
849 "src/core/NEON/kernels/NEFillBorderKernel.cpp",
850 "src/runtime/NEON/INEOperator.cpp",
851 "src/runtime/NEON/INESimpleFunction.cpp",
852 "src/runtime/NEON/INESimpleFunctionNoBorder.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100853 ],
854 "operators": {
855 "Activation": {
856 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100857 "common": [
858 "src/cpu/operators/CpuActivation.cpp",
859 "src/cpu/kernels/CpuActivationKernel.cpp",
Dana Zlotnika538ae52022-02-21 13:12:41 +0200860 "src/runtime/NEON/functions/NEActivationLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100861 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100862 "neon": {
Dana Zlotnik32291712021-11-25 09:58:27 +0200863 "fp16": [ "src/cpu/kernels/activation/generic/neon/fp16.cpp" ],
ramelg01c827e992022-04-08 03:52:28 +0100864 "fp32": [ "src/cpu/kernels/activation/generic/neon/fp32.cpp" ],
Viet-Hoa Do29db3d22022-08-10 11:56:49 +0100865 "qasymm8": [
866 "src/cpu/kernels/activation/generic/neon/qasymm8.cpp",
867 "src/cpu/kernels/activation/generic/neon/lut.cpp"
868 ],
869 "qasymm8_signed": [
870 "src/cpu/kernels/activation/generic/neon/qasymm8_signed.cpp",
871 "src/cpu/kernels/activation/generic/neon/lut.cpp"
872 ],
Dana Zlotnika538ae52022-02-21 13:12:41 +0200873 "qsymm16": [ "src/cpu/kernels/activation/generic/neon/qsymm16.cpp" ]
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100874 },
875 "sve": {
Dana Zlotnik32291712021-11-25 09:58:27 +0200876 "fp16": [ "src/cpu/kernels/activation/generic/sve/fp16.cpp" ],
Pablo Marquez Tellod1586092022-06-10 14:37:10 +0100877 "fp32": [ "src/cpu/kernels/activation/generic/sve/fp32.cpp" ],
878 "qasymm8": ["src/cpu/kernels/activation/generic/sve/lut.cpp"]
ramelg01c827e992022-04-08 03:52:28 +0100879 },
alerah01c9e519d2022-01-31 19:04:10 +0200880 "sve2":{
Dana Zlotnik32291712021-11-25 09:58:27 +0200881 "qasymm8": [ "src/cpu/kernels/activation/generic/sve2/qasymm8.cpp" ],
882 "qasymm8_signed": [ "src/cpu/kernels/activation/generic/sve2/qasymm8_signed.cpp" ],
883 "qsymm16": [ "src/cpu/kernels/activation/generic/sve2/qsymm16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100884 }
885 }
886 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100887 "ArgMinMax": {
888 "deps": [ "Reduction" ],
889 "files": {
890 "common": [ "src/runtime/NEON/functions/NEArgMinMaxLayer.cpp" ]
891 }
892 },
Michalis Spyrou20fca522021-06-07 14:23:57 +0100893 "Add": {
894 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100895 "common": [
896 "src/cpu/operators/CpuAdd.cpp",
897 "src/cpu/kernels/CpuAddKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200898 "src/runtime/NEON/functions/NEArithmeticAddition.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100899 ],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200900 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200901 "common": ["src/cpu/kernels/add/generic/neon/impl.cpp"],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200902 "fp32":["src/cpu/kernels/add/generic/neon/fp32.cpp"],
903 "fp16":["src/cpu/kernels/add/generic/neon/fp16.cpp"],
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200904 "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"],
905 "qasymm8": ["src/cpu/kernels/add/generic/neon/qasymm8.cpp"],
906 "qasymm8_signed": ["src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp"],
907 "qsymm16": ["src/cpu/kernels/add/generic/neon/qsymm16.cpp"]
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200908 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100909 "sve": {
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200910 "common": [ "src/cpu/kernels/add/generic/sve/impl.cpp" ],
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200911 "integer":["src/cpu/kernels/add/generic/sve/integer.cpp"],
Motti Gondabi9d9ad332022-01-23 12:42:24 +0200912 "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"],
913 "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"]
914 },
915 "sve2": {
Dana Zlotnikbd2942d2021-11-15 08:46:04 +0200916 "qasymm8": [ "src/cpu/kernels/add/generic/sve2/qasymm8.cpp" ],
917 "qasymm8_signed": [ "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp" ],
918 "qsymm16": [ "src/cpu/kernels/add/generic/sve2/qsymm16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100919 }
920 }
921 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100922 "BatchNormalize": {
Michalis Spyrou20fca522021-06-07 14:23:57 +0100923 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100924 "common": [
925 "src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp",
926 "src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp",
927 "src/runtime/NEON/functions/NEBatchNormalizationLayer.cpp",
928 "src/runtime/NEON/functions/NEFuseBatchNormalization.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100929 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +0100930 "neon": {
Yair Schwarzbaum41a729e2021-11-15 20:42:47 +0200931 "common": [
932 "src/cpu/kernels/fuse_batch_normalization/generic/impl.cpp",
933 "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp",
934 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp"
935 ],
936 "fp16": [
937 "src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp",
938 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp",
939 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp"
940 ],
941 "fp32": [
942 "src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp",
943 "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp",
944 "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp"
945 ]
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100946 },
947 "sve": {
948 "fp16": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp" ],
949 "fp32": [ "src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +0100950 }
951 }
952 },
953 "BatchToSpace": {
954 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100955 "common": [
956 "src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp",
957 "src/runtime/NEON/functions/NEBatchToSpaceLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100958 ]
959 }
960 },
961 "BitwiseAnd": {
962 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100963 "common": [
964 "src/core/NEON/kernels/NEBitwiseAndKernel.cpp",
965 "src/runtime/NEON/functions/NEBitwiseAnd.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100966 ]
967 }
968 },
969 "BitwiseNot": {
970 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100971 "common": [
972 "src/core/NEON/kernels/NEBitwiseNotKernel.cpp",
973 "src/runtime/NEON/functions/NEBitwiseNot.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100974 ]
975 }
976 },
977 "BitwiseOr": {
978 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100979 "common": [
980 "src/core/NEON/kernels/NEBitwiseOrKernel.cpp",
981 "src/runtime/NEON/functions/NEBitwiseOr.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100982 ]
983 }
984 },
985 "BitwiseXor": {
986 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100987 "common": [
988 "src/core/NEON/kernels/NEBitwiseXorKernel.cpp",
989 "src/runtime/NEON/functions/NEBitwiseXor.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +0100990 ]
991 }
992 },
993 "BoundingBoxTransform": {
994 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +0100995 "common": [
996 "src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp",
997 "src/runtime/NEON/functions/NEBoundingBoxTransform.cpp"
Dana Zlotnik3475ffe2022-01-03 14:37:10 +0200998 ],
999 "neon":{
1000 "common":["src/cpu/kernels/boundingboxtransform/generic/neon/impl.cpp"],
1001 "fp32":["src/cpu/kernels/boundingboxtransform/generic/neon/fp32.cpp"],
1002 "fp16":["src/cpu/kernels/boundingboxtransform/generic/neon/fp16.cpp"],
1003 "qsymm16":["src/cpu/kernels/boundingboxtransform/generic/neon/qsymm16.cpp"]
Dana Zlotnik3475ffe2022-01-03 14:37:10 +02001004 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001005 }
1006 },
1007 "Cast": {
1008 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001009 "common": [
1010 "src/cpu/operators/CpuCast.cpp",
1011 "src/cpu/kernels/CpuCastKernel.cpp",
Yair Schwarzbaum298b2c02022-02-01 08:55:56 +02001012 "src/runtime/NEON/functions/NECast.cpp",
1013 "src/cpu/kernels/cast/generic/neon/bfloat16.cpp"
1014 ],
1015 "neon":{
1016 "fp16":["src/cpu/kernels/cast/generic/neon/fp16.cpp"]
1017 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001018 }
1019 },
1020 "ChannelShuffle": {
1021 "files": {
1022 "common": [
1023 "src/core/NEON/kernels/NEChannelShuffleLayerKernel.cpp",
1024 "src/runtime/NEON/functions/NEChannelShuffleLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001025 ]
1026 }
1027 },
1028 "Concatenate": {
1029 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001030 "common": [
1031 "src/cpu/operators/CpuConcatenate.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001032 "src/cpu/kernels/CpuConcatenateWidthKernel.cpp",
1033 "src/cpu/kernels/CpuConcatenateBatchKernel.cpp",
1034 "src/cpu/kernels/CpuConcatenateDepthKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001035 "src/cpu/kernels/CpuConcatenateHeightKernel.cpp",
1036 "src/runtime/NEON/functions/NEConcatenateLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001037 ]
1038 }
1039 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001040 "Conv2d": {
1041 "deps": [
1042 "Activation",
1043 "ElementwiseBinary",
1044 "FFT2D",
1045 "Gemm",
1046 "Mul",
1047 "Pad",
1048 "Permute",
1049 "Reshape",
1050 "Reverse",
1051 "Slice"
1052 ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001053 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001054 "common": [
1055 "src/cpu/operators/CpuConv2d.cpp",
1056 "src/cpu/operators/CpuDirectConv2d.cpp",
1057 "src/cpu/operators/CpuGemmDirectConv2d.cpp",
1058 "src/cpu/operators/CpuGemmConv2d.cpp",
1059 "src/cpu/operators/CpuWinogradConv2d.cpp",
1060 "src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp",
1061 "src/cpu/kernels/CpuDirectConv2dKernel.cpp",
1062 "src/cpu/kernels/CpuDirectConv2dOutputStageKernel.cpp",
1063 "src/cpu/kernels/CpuWinogradConv2dKernel.cpp",
1064 "src/cpu/kernels/CpuCol2ImKernel.cpp",
1065 "src/cpu/kernels/CpuIm2ColKernel.cpp",
1066 "src/cpu/kernels/CpuWeightsReshapeKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001067 "src/runtime/NEON/functions/NEConvolutionLayer.cpp",
1068 "src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp",
1069 "src/runtime/NEON/functions/NEFFTConvolutionLayer.cpp",
1070 "src/runtime/NEON/functions/NEGEMMConv2d.cpp",
1071 "src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp",
1072 "src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001073 ],
1074 "neon": {
1075 "common": [
1076 "src/core/NEON/kernels/convolution/common/padding.cpp",
1077 "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
1078 "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
1079 "src/core/NEON/kernels/convolution/common/utils.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001080 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp16.cpp",
1081 "src/core/NEON/kernels/convolution/winograd/input_transforms_fp32.cpp",
1082 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp16.cpp",
1083 "src/core/NEON/kernels/convolution/winograd/output_transforms_fp32.cpp",
1084 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp16.cpp",
1085 "src/core/NEON/kernels/convolution/winograd/weight_transforms_fp32.cpp",
1086 "src/core/NEON/kernels/convolution/winograd/winograd_fp16.cpp",
1087 "src/core/NEON/kernels/convolution/winograd/winograd_fp32.cpp",
1088 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp",
1089 "src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp",
1090 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp",
1091 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp",
1092 "src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001093 "src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp",
1094 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp",
1095 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp",
1096 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp",
1097 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp",
1098 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp",
1099 "src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp",
1100 "src/core/NEON/kernels/convolution/winograd/weight_transforms/a64_fp16_4x4_3x3.cpp",
1101 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_3x3.cpp",
1102 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_2x2_5x5.cpp",
1103 "src/core/NEON/kernels/convolution/winograd/weight_transforms/arm_fp32_4x4_3x3.cpp",
1104 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x2_1x7.cpp",
1105 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp",
1106 "src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x6_1x3.cpp",
alerah01c9e519d2022-01-31 19:04:10 +02001107 "src/cpu/kernels/directconv2d/nhwc/neon/impl.cpp",
1108 "src/cpu/kernels/directconv2d/nchw/all.cpp"
1109 ],
1110 "fp32": [
1111 "src/cpu/kernels/directconv2d/nhwc/neon/fp32.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001112 ]
ramelg0199c46472022-07-15 13:08:39 +01001113 },
1114 "sve": {
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001115 "common": [
1116 "src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp",
1117 "src/core/NEON/kernels/convolution/winograd/input_transforms/sme_fp32_mla_6x6.cpp",
1118 "src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp"
1119 ]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001120 }
Michalis Spyroub55f8e82021-07-22 11:23:11 +01001121 }
1122 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001123 "Copy": {
1124 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001125 "common": [
1126 "src/cpu/operators/CpuCopy.cpp",
1127 "src/cpu/kernels/CpuCopyKernel.cpp",
1128 "src/runtime/NEON/functions/NECopy.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001129 ]
1130 }
1131 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001132 "CropResize": {
1133 "deps": [ "Scale" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001134 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001135 "common": [
1136 "src/core/NEON/kernels/NECropKernel.cpp",
1137 "src/runtime/NEON/functions/NECropResize.cpp"
alerah0127619932021-11-24 17:21:43 +02001138 ],
1139 "neon": {
1140 "common": [ "src/cpu/kernels/crop/generic/neon/impl.cpp" ],
1141 "fp32": [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ],
1142 "fp16": [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ],
1143 "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp" ]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001144 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001145 }
1146 },
1147 "Deconv2d": {
1148 "deps": [ "Conv2d", "Reverse", "Transpose"],
1149 "files": {
1150 "common": [
1151 "src/runtime/NEON/functions/NEDeconvolutionLayer.cpp"
1152 ]
1153 }
1154 },
1155 "DepthConvert": {
1156 "deps": [ "Cast"],
1157 "files": {
1158 "common": [
1159 "src/runtime/NEON/functions/NEDepthConvertLayer.cpp"
1160 ]
1161 }
1162 },
1163 "DepthToSpace": {
1164 "files": {
1165 "common": [
1166 "src/core/NEON/kernels/NEDepthToSpaceLayerKernel.cpp",
1167 "src/runtime/NEON/functions/NEDepthToSpaceLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001168 ]
1169 }
1170 },
1171 "DepthwiseConv2d": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001172 "deps": [ "Activation", "Permute" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001173 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001174 "common": [
Georgios Pinitas7891a732021-08-20 21:39:25 +01001175 "src/cpu/operators/CpuDepthwiseConv2d.cpp",
1176 "src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001177 "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.cpp",
1178 "src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001179 "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001180 ],
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001181 "neon": {
1182 "common": [
1183 "src/core/NEON/kernels/convolution/common/padding.cpp",
1184 "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
1185 "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
1186 "src/core/NEON/kernels/convolution/common/utils.cpp",
ramelg01c827e992022-04-08 03:52:28 +01001187 "src/core/NEON/kernels/arm_conv/addressing.cpp",
ramelg018a164882022-04-07 02:42:52 +01001188 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_strategies_common.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001189 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
1190 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
1191 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp",
1192 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
1193 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
1194 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1195 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
1196 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
1197 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1198 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1199 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1200 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1201 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1202 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1203 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1204 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
ramelg018a164882022-04-07 02:42:52 +01001205 "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001206 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1207 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1208 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1209 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1210 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1211 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1212 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1213 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1214 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1215 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1216 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1217 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1218 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1219 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1220 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1221 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1222 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1223 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1224 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1225 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1226 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1227 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1228 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1229 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1230 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1231 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1232 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1233 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1234 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1235 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1236 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1237 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1238 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1239 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1240 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1241 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1242 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
ramelg018a164882022-04-07 02:42:52 +01001243 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1244 "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1245 "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 +02001246 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
ramelg01c827e992022-04-08 03:52:28 +01001247 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp",
ramelg018a164882022-04-07 02:42:52 +01001248 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp",
1249 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp",
Dana Zlotnikebbae942022-02-03 12:52:15 +02001250 "src/cpu/kernels/depthwiseconv2d/generic/neon/impl.cpp"
ramelg01a1f78512022-06-29 16:28:10 +01001251 ],
Dana Zlotnikebbae942022-02-03 12:52:15 +02001252 "fp16":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp16.cpp"],
ramelg01a1f78512022-06-29 16:28:10 +01001253 "fp32":["src/cpu/kernels/depthwiseconv2d/generic/neon/fp32.cpp"],
1254 "qasymm8":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8.cpp"],
Dana Zlotnikebbae942022-02-03 12:52:15 +02001255 "qasymm8_signed":["src/cpu/kernels/depthwiseconv2d/generic/neon/qasymm8_signed.cpp"]
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001256 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001257 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001258 "common": [
Freddie Liardetd216f572021-08-03 15:57:32 +01001259 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_8b_mla.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001260 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp",
1261 "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp",
1262 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1263 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1264 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1265 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1266 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1267 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1268 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1269 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1270 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1271 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1272 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1273 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1274 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1275 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst_strided/generic.cpp",
1276 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1277 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1278 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1279 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1280 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1281 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1282 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1283 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1284 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
1285 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
1286 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
1287 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
1288 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1289 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1290 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1291 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1292 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1293 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1294 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1295 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
1296 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1297 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
1298 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1299 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
1300 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
1301 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
1302 "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 +01001303 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
1304 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
1305 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
1306 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
1307 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
1308 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
1309 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
1310 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
1311 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
1312 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s1_4rows_mla_za/generic.cpp",
1313 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_3x3_s2_4rows_mla_za/generic.cpp",
1314 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s1_4rows_mla_za/generic.cpp",
1315 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_planar_5x5_s2_4rows_mla_za/generic.cpp",
1316 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za/generic.cpp",
1317 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_3x3_s2_4rows_dot_za/generic.cpp",
1318 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s1_4rows_dot_za/generic.cpp",
1319 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp",
1320 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1321 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
1322 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1323 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1324 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_s8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
1325 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1326 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
1327 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1328 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1329 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8q_planar_5x5_s2_4rows_dot_za/generic.cpp",
1330 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za/generic.cpp",
1331 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_2rows_dot_za/generic.cpp",
1332 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_3x3_s2_4rows_dot_za/generic.cpp",
1333 "src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s1_4rows_dot_za/generic.cpp",
1334 "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 +01001335 ]
1336 }
1337 }
1338 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001339 "Dequantize": {
1340 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001341 "common": [
1342 "src/cpu/operators/CpuDequantize.cpp",
1343 "src/cpu/kernels/CpuDequantizeKernel.cpp",
1344 "src/runtime/NEON/functions/NEDequantizationLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001345 ]
1346 }
1347 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001348 "DetectionPostProcess": {
1349 "deps": [ "Dequantize" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001350 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001351 "common" : [ "src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001352 }
1353 },
Sheri Zhang6d9c9822021-09-24 16:02:57 +01001354 "Conv3d": {
1355 "deps": [
1356 "Activation"
1357 ],
1358 "files": {
1359 "common": [
1360 "src/cpu/operators/CpuDirectConv3d.cpp",
1361 "src/cpu/kernels/CpuDirectConv3dKernel.cpp",
1362 "src/runtime/NEON/functions/NEConv3D.cpp"
1363 ]
1364 }
1365 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001366 "ElementwiseBinary": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001367 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001368 "common": [
1369 "src/cpu/operators/CpuElementwise.cpp",
1370 "src/cpu/kernels/CpuElementwiseKernel.cpp",
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001371 "src/runtime/NEON/functions/NEElementwiseOperations.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001372 ],
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001373 "neon":{
1374 "fp32": ["src/cpu/kernels/elementwise_binary/generic/neon/fp32.cpp"],
1375 "fp16": ["src/cpu/kernels/elementwise_binary/generic/neon/fp16.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001376 "integer": ["src/cpu/kernels/elementwise_binary/generic/neon/integer.cpp"],
1377 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8.cpp"],
1378 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/neon/qasymm8_signed.cpp"]
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001379 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001380 "sve": {
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001381 "common": ["src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp" ],
1382 "integer": ["src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp"],
1383 "fp32": ["src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001384 "fp16": ["src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp"]
1385
ramelg01a1f78512022-06-29 16:28:10 +01001386 },
Dana Zlotnik6a2df882022-01-17 09:54:26 +02001387 "sve2":{
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001388 "qasymm8": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8.cpp"],
1389 "qasymm8_signed": ["src/cpu/kernels/elementwise_binary/generic/sve2/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001390 }
1391 }
1392 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001393 "ElementwiseUnary":{
Michalis Spyrou20fca522021-06-07 14:23:57 +01001394 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001395 "common": [
1396 "src/cpu/operators/CpuElementwiseUnary.cpp",
1397 "src/cpu/kernels/CpuElementwiseUnaryKernel.cpp",
1398 "src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001399 ],
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001400 "neon": {
Dana Zlotnika538ae52022-02-21 13:12:41 +02001401 "common":["src/cpu/kernels/elementwise_unary/generic/neon/impl.cpp"],
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001402 "integer": ["src/cpu/kernels/elementwise_unary/generic/neon/integer.cpp"],
1403 "fp32": ["src/cpu/kernels/elementwise_unary/generic/neon/fp32.cpp"],
1404 "fp16": ["src/cpu/kernels/elementwise_unary/generic/neon/fp16.cpp"]
1405 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001406 "sve": {
Dana Zlotnikd5c496d2021-11-28 14:46:12 +02001407 "common": ["src/cpu/kernels/elementwise_unary/generic/sve/impl.cpp" ],
1408 "integer": ["src/cpu/kernels/elementwise_unary/generic/sve/integer.cpp"],
1409 "fp32": ["src/cpu/kernels/elementwise_unary/generic/sve/fp32.cpp"],
1410 "fp16": ["src/cpu/kernels/elementwise_unary/generic/sve/fp16.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001411 }
1412 }
1413 },
1414 "FFT1D": {
Freddie Liardet487d3902021-09-21 12:36:43 +01001415 "deps": [ "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001416 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001417 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +01001418 "src/core/NEON/kernels/NEFFTDigitReverseKernel.cpp",
1419 "src/core/NEON/kernels/NEFFTRadixStageKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001420 "src/core/NEON/kernels/NEFFTScaleKernel.cpp",
1421 "src/runtime/NEON/functions/NEFFT1D.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001422 ]
1423 }
1424 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001425 "FFT2D": {
1426 "deps": [ "FFT1D" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001427 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001428 "common": [
1429 "src/runtime/NEON/functions/NEFFT2D.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001430 ]
1431 }
1432 },
1433 "Fill": {
1434 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001435 "common": [
1436 "src/cpu/operators/CpuFill.cpp",
1437 "src/cpu/kernels/CpuFillKernel.cpp",
1438 "src/runtime/NEON/functions/NEFill.cpp"
1439 ]
1440 }
1441 },
1442 "Flatten": {
1443 "deps: ": [ "Reshape" ],
1444 "files": {
1445 "common": [
1446 "src/cpu/operators/CpuFlatten.cpp",
1447 "src/runtime/NEON/functions/NEFlattenLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001448 ]
1449 }
1450 },
1451 "Floor": {
1452 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001453 "common": [
1454 "src/cpu/operators/CpuFloor.cpp",
1455 "src/cpu/kernels/CpuFloorKernel.cpp",
1456 "src/runtime/NEON/functions/NEFloor.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001457 ],
1458 "neon": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001459 "fp32": [ "src/cpu/kernels/floor/neon/fp32.cpp" ],
1460 "fp16": [ "src/cpu/kernels/floor/neon/fp16.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001461 }
1462 }
1463 },
Michele Di Giorgiod9cdf142021-07-02 15:17:08 +01001464 "FullyConnected": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001465 "deps": [ "Flatten", "Gemm", "Transpose"],
Michele Di Giorgiod9cdf142021-07-02 15:17:08 +01001466 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001467 "common": [
1468 "src/cpu/kernels/CpuConvertFullyConnectedWeightsKernel.cpp",
1469 "src/cpu/operators/CpuConvertFullyConnectedWeights.cpp",
1470 "src/cpu/operators/CpuFullyConnected.cpp",
1471 "src/runtime/NEON/functions/NEConvertFullyConnectedWeights.cpp",
1472 "src/runtime/NEON/functions/NEFullyConnectedLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001473 ]
1474 }
1475 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001476 "Gather": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001477 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001478 "common": [
1479 "src/core/NEON/kernels/NEGatherKernel.cpp",
1480 "src/runtime/NEON/functions/NEGather.cpp"
1481 ]
1482 }
1483 },
1484 "Gemm": {
Freddie Liardet487d3902021-09-21 12:36:43 +01001485 "deps": [ "Quantize", "Add"],
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001486 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001487 "common": [
1488 "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001489 "src/cpu/kernels/CpuGemmMatrixAdditionKernel.cpp",
1490 "src/cpu/kernels/CpuGemmMatrixMultiplyKernel.cpp",
1491 "src/cpu/kernels/CpuGemmTranspose1xWKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001492 "src/cpu/kernels/CpuGemmInterleave4x4Kernel.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001493 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ScaleKernel.cpp",
1494 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp",
1495 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp",
1496 "src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp",
1497 "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp",
1498 "src/cpu/kernels/CpuGemmLowpMatrixReductionKernel.cpp",
1499 "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001500 "src/cpu/kernels/CpuGemmLowpOffsetContributionKernel.cpp",
1501 "src/cpu/operators/CpuGemm.cpp",
1502 "src/cpu/operators/CpuGemmLowpOutputStage.cpp",
1503 "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001504 "src/runtime/NEON/functions/NEGEMM.cpp",
1505 "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001506 "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001507 ],
1508 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001509 "common": [
1510 "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
1511 "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
1512 "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
1513 "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp",
1514 "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp",
1515 "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp",
1516 "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp",
1517 "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp",
1518 "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp",
1519 "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp",
1520 "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp",
1521 "src/core/NEON/kernels/arm_gemm/mergeresults.cpp",
1522 "src/core/NEON/kernels/arm_gemm/misc.cpp",
1523 "src/core/NEON/kernels/arm_gemm/quantized.cpp",
1524 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp",
1525 "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp",
1526 "src/core/NEON/kernels/arm_gemm/transform.cpp",
1527 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
1528 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
1529 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
1530 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
1531 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
1532 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
1533 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
1534 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
1535 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
1536 "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
1537 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
1538 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
1539 "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
1540 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
1541 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
1542 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
1543 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
1544 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
1545 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
1546 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
1547 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
1548 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
1549 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
1550 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
1551 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
1552 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
1553 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
1554 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
1555 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
1556 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
1557 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
1558 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
1559 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
1560 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
1561 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
1562 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
1563 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
1564 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
1565 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
1566 "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
1567 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
1568 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
1569 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
1570 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp",
1571 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
1572 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp",
1573 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
1574 "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp",
1575 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
1576 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
1577 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
1578 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
1579 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
1580 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
1581 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
1582 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
1583 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
1584 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
1585 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
1586 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
1587 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
1588 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
1589 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001590 "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp",
Dana Zlotnik256ac622022-02-02 15:06:11 +02001591 "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp",
1592 "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001593 ],
Dana Zlotnik256ac622022-02-02 15:06:11 +02001594 "fp32":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp",
1595 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp"],
1596 "fp16":["src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp",
1597 "src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp"],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001598 "estate32": [
1599 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp",
1600 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp",
1601 "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp"
1602 ],
1603 "estate64": [
Freddie Liardet487d3902021-09-21 12:36:43 +01001604 "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp"
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001605 ],
1606 "experimental_fixed_format_kernels": [
1607 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp",
1608 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp",
1609 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp",
1610 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
1611 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp",
1612 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp",
1613 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp",
1614 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001615 ]
1616 },
1617 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001618 "common": [
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001619 "src/core/NEON/kernels/arm_gemm/interleave_indirect-sve.cpp",
1620 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp",
1621 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp",
1622 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp",
1623 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp",
1624 "src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp",
1625 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp",
1626 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp",
1627 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp",
1628 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp",
1629 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp",
1630 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp",
1631 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp",
1632 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp",
1633 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp",
1634 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp",
1635 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp",
1636 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp",
1637 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp",
1638 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp",
1639 "src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001640 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001641 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp",
1642 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001643 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001644 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001645 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001646 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001647 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001648 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp",
1649 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001650 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001651 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001652 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001653 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp",
1654 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001655 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001656 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001657 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001658 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp",
1659 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001660 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001661 "src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001662 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp",
1663 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001664 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001665 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001666 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001667 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp",
1668 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001669 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001670 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp",
1671 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001672 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001673 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp",
1674 "src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp",
1675 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_fp32_mla_8x1VL/generic.cpp",
1676 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_s8s32_dot_8x1VL/generic.cpp",
Georgios Pinitas4ee8b152021-07-16 16:16:43 +01001677 "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_u8u32_dot_8x1VL/generic.cpp",
1678 "src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001679 "src/core/NEON/kernels/arm_gemm/transform-sve.cpp",
1680 "src/core/NEON/kernels/arm_gemm/misc-sve.cpp"
Francesco.Petrogalli@arm.com5fcf22d2022-04-05 10:31:08 +00001681 ],
1682 "experimental_fixed_format_kernels": [
1683 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp",
1684 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp",
1685 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp",
1686 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp",
1687 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp",
1688 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp",
1689 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp",
1690 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp",
1691 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp",
1692 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp",
1693 "src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp"
1694 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001695 }
1696 }
1697 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001698 "GenerateProposals": {
1699 "deps": [ "BoundingBoxTransform", "Dequantize", "Pad", "Permute", "Quantize", "Reshape" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001700 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001701 "common": [
1702 "src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp",
1703 "src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp"
Dana Zlotnik50ce2772021-12-29 13:55:56 +02001704 ],
1705 "neon":{
1706 "common":["src/cpu/kernels/genproposals/generic/neon/impl.cpp"],
1707 "fp16":["src/cpu/kernels/genproposals/generic/neon/fp16.cpp"],
1708 "fp32":["src/cpu/kernels/genproposals/generic/neon/fp32.cpp"],
1709 "qsymm16":["src/cpu/kernels/genproposals/generic/neon/qsymm16.cpp"]
1710 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001711 }
1712 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001713 "InstanceNormalize": {
1714 "deps": [ "Permute", "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001715 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001716 "common": [
1717 "src/core/NEON/kernels/NEInstanceNormalizationLayerKernel.cpp",
1718 "src/runtime/NEON/functions/NEInstanceNormalizationLayer.cpp"
ramelg01a1f78512022-06-29 16:28:10 +01001719 ],
Dana Zlotnikd7e2ec52022-01-03 10:59:41 +02001720 "neon":{
1721 "common":["src/cpu/kernels/instancenorm/generic/neon/impl.cpp"],
1722 "fp16":["src/cpu/kernels/instancenorm/generic/neon/fp16.cpp"],
1723 "fp32":["src/cpu/kernels/instancenorm/generic/neon/fp32.cpp"]
1724 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001725 }
1726 },
1727 "L2Normalize": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001728 "deps": [ "Reduction" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001729 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001730 "common": [
1731 "src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp",
1732 "src/runtime/NEON/functions/NEL2NormalizeLayer.cpp"
Yair Schwarzbaum5e993182022-01-10 15:11:07 +02001733 ],
1734 "neon":{
1735 "common":["src/cpu/kernels/l2normlayer/generic/neon/impl.cpp"],
1736 "fp32":["src/cpu/kernels/l2normlayer/generic/neon/fp32.cpp"],
1737 "fp16":["src/cpu/kernels/l2normlayer/generic/neon/fp16.cpp"]
1738 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001739 }
1740 },
1741 "Logical": {
1742 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001743 "common": [
1744 "src/core/NEON/kernels/NELogicalKernel.cpp",
1745 "src/runtime/NEON/functions/NELogical.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001746 ]
1747 }
1748 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001749 "LSTM": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001750 "deps": [
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001751 "Activation",
1752 "Concatenate",
1753 "Copy",
1754 "Dequantize",
1755 "ElementwiseBinary",
1756 "Fill",
1757 "FullyConnected",
1758 "Gemm",
1759 "MeanStdDevNormalize",
1760 "Mul",
1761 "Quantize",
1762 "Slice",
1763 "Transpose"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001764 ],
1765 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001766 "common": [
1767 "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.cpp",
1768 "src/runtime/NEON/functions/NELSTMLayer.cpp",
1769 "src/runtime/NEON/functions/NELSTMLayerQuantized.cpp",
1770 "src/runtime/NEON/functions/NEQLSTMLayer.cpp"
1771 ]
1772 }
1773 },
1774 "MaxUnpool2d": {
1775 "deps": [ "Fill" ],
1776 "files": {
1777 "common": [
Dana Zlotnik149203b2022-01-26 12:38:03 +02001778 "src/cpu/kernels/CpuMaxUnpoolingLayerKernel.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001779 "src/runtime/NEON/functions/NEMaxUnpoolingLayer.cpp",
Dana Zlotnik149203b2022-01-26 12:38:03 +02001780 "src/cpu/operators/CpuMaxUnpooling.cpp"
Dana Zlotnik143310c2021-12-30 13:24:52 +02001781 ],
1782 "neon":{
Dana Zlotnik143310c2021-12-30 13:24:52 +02001783 "fp32":["src/cpu/kernels/maxunpool/generic/neon/fp32.cpp"],
1784 "fp16":["src/cpu/kernels/maxunpool/generic/neon/fp16.cpp"],
1785 "qasymm8":["src/cpu/kernels/maxunpool/generic/neon/qasymm8.cpp"],
Dana Zlotnik149203b2022-01-26 12:38:03 +02001786 "qasymm8_signed":[ "src/cpu/kernels/maxunpool/generic/neon/qasymm8_signed.cpp"],
1787 "common":["src/cpu/kernels/maxunpool/generic/neon/impl.cpp"]
Dana Zlotnik143310c2021-12-30 13:24:52 +02001788 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001789 }
1790 },
1791 "Mean": {
1792 "deps" : [ "Reduction" ],
1793 "files": {
1794 "common": [ "src/runtime/NEON/functions/NEReduceMean.cpp" ]
1795 }
1796 },
1797 "MeanStdDevNormalize": {
1798 "files": {
1799 "common": [
1800 "src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp",
1801 "src/runtime/NEON/functions/NEMeanStdDevNormalizationLayer.cpp"
Dana Zlotnik027bcef2021-12-27 17:35:00 +02001802 ],
1803 "neon":{
1804 "common":["src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp"],
1805 "fp32":["src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp"],
Murray Kornelsen6e09e142022-07-13 21:40:26 -04001806 "fp16":["src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp"],
1807 "qasymm8":["src/cpu/kernels/meanstddevnorm/generic/neon/qasymm8.cpp"]
Dana Zlotnik027bcef2021-12-27 17:35:00 +02001808 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001809 }
1810 },
1811 "Mul": {
1812 "files": {
1813 "common": [
1814 "src/cpu/operators/CpuMul.cpp",
1815 "src/cpu/kernels/CpuMulKernel.cpp",
1816 "src/runtime/NEON/functions/NEPixelWiseMultiplication.cpp"
1817 ]
1818 }
1819 },
1820 "Normalize": {
1821 "deps": [ "Mul" ],
1822 "files": {
1823 "common": [
1824 "src/core/NEON/kernels/NENormalizationLayerKernel.cpp",
1825 "src/runtime/NEON/functions/NENormalizationLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001826 ]
1827 }
1828 },
1829 "Pad": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001830 "deps": [ "Concatenate", "Copy", "StridedSlice" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001831 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001832 "common": [
1833 "src/core/NEON/kernels/NEPadLayerKernel.cpp",
1834 "src/runtime/NEON/functions/NEPadLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001835 ]
1836 }
1837 },
1838 "Permute": {
1839 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001840 "common": [
1841 "src/cpu/operators/CpuPermute.cpp",
1842 "src/cpu/kernels/CpuPermuteKernel.cpp",
1843 "src/runtime/NEON/functions/NEPermute.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001844 ]
1845 }
1846 },
1847 "Pool2d": {
1848 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001849 "common": [
1850 "src/cpu/operators/CpuPool2d.cpp",
Georgios Pinitas7891a732021-08-20 21:39:25 +01001851 "src/cpu/kernels/CpuPool2dKernel.cpp",
1852 "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp",
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001853 "src/runtime/NEON/functions/NEPoolingLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001854 ],
1855 "neon": {
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001856 "common": [
1857 "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
1858 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp",
1859 "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp",
1860 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp",
1861 "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp",
1862 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
1863 "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
1864 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1865 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
1866 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1867 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
1868 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1869 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
1870 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1871 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
1872 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
1873 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1874 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
1875 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
1876 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
1877 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
1878 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1879 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
1880 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
ramelg01a1f78512022-06-29 16:28:10 +01001881 "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
Motti Gondabi9d9ad332022-01-23 12:42:24 +02001882 ],
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001883 "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
1884 "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ],
ramelg01a1f78512022-06-29 16:28:10 +01001885 "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ],
1886 "qasymm8":[ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
Dana Zlotnika538ae52022-02-21 13:12:41 +02001887 "qasymm8_signed":["src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01001888 },
1889 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001890 "common": [
Michalis Spyrou20fca522021-06-07 14:23:57 +01001891 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1892 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
1893 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1894 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_max_generic_depthfirst/generic.cpp",
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001895 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1896 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
1897 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1898 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp32_nhwc_max_generic_depthfirst/generic.cpp",
Michalis Spyrou20fca522021-06-07 14:23:57 +01001899 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_avg_generic_depthfirst/generic.cpp",
1900 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1901 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp",
1902 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
1903 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp",
1904 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp",
1905 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1906 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp",
1907 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
Viet-Hoa Do03b29712022-06-01 11:47:14 +01001908 "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp",
1909 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1910 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
1911 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1912 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_max_generic_depthfirst/generic.cpp",
1913 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
1914 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
1915 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1916 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp",
1917 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_avg_generic_depthfirst/generic.cpp",
1918 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1919 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8_nhwc_max_generic_depthfirst/generic.cpp",
1920 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
1921 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp",
1922 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_avg_generic_depthfirst/generic.cpp",
1923 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
1924 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8_nhwc_max_generic_depthfirst/generic.cpp",
1925 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
1926 "src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001927 ]
1928 }
1929 }
1930 },
Adnan AlSinan171fc3d2022-03-15 18:46:42 +00001931 "Pool3d": {
1932 "files": {
1933 "common": [
1934 "src/cpu/operators/CpuPool3d.cpp",
1935 "src/cpu/kernels/CpuPool3dKernel.cpp",
1936 "src/runtime/NEON/functions/NEPooling3dLayer.cpp"
1937 ],
1938 "neon": {
Adnan AlSinan9104cd52022-04-06 16:19:31 +01001939 "common": [ "src/cpu/kernels/pool3d/neon/impl.cpp" ],
1940 "fp16": [ "src/cpu/kernels/pool3d/neon/fp16.cpp" ],
1941 "fp32": [ "src/cpu/kernels/pool3d/neon/fp32.cpp" ],
1942 "qasymm8": [ "src/cpu/kernels/pool3d/neon/qasymm8.cpp" ],
1943 "qasymm8_signed": [ "src/cpu/kernels/pool3d/neon/qasymm8_signed.cpp" ]
Adnan AlSinan171fc3d2022-03-15 18:46:42 +00001944 }
1945 }
1946 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001947 "PRelu": {
1948 "deps": [ "ElementwiseBinary" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001949 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001950 "common": [
1951 "src/runtime/NEON/functions/NEPReluLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001952 ]
1953 }
1954 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001955 "PriorBox": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01001956 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001957 "common": [
1958 "src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp",
1959 "src/runtime/NEON/functions/NEPriorBoxLayer.cpp"
1960 ]
1961 }
1962 },
1963 "Quantize": {
1964 "files": {
1965 "common": [
1966 "src/cpu/operators/CpuQuantize.cpp",
1967 "src/cpu/kernels/CpuQuantizeKernel.cpp",
1968 "src/runtime/NEON/functions/NEQuantizationLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001969 ]
1970 }
1971 },
1972 "Range": {
1973 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001974 "common": [
1975 "src/core/NEON/kernels/NERangeKernel.cpp",
1976 "src/runtime/NEON/functions/NERange.cpp"
Yair Schwarzbaum9e8a7702021-11-08 10:58:06 +02001977 ],
1978 "neon": {
1979 "common": [ "src/cpu/kernels/range/generic/neon/impl.cpp" ],
1980 "fp32": [ "src/cpu/kernels/range/generic/neon/fp32.cpp" ],
1981 "fp16": [ "src/cpu/kernels/range/generic/neon/fp16.cpp" ],
1982 "integer": [ "src/cpu/kernels/range/generic/neon/integer.cpp" ]
1983 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01001984 }
1985 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001986 "Reduction":{
1987 "deps": [ "Reshape" ],
Michalis Spyrou20fca522021-06-07 14:23:57 +01001988 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001989 "common": [
1990 "src/core/NEON/kernels/NEReductionOperationKernel.cpp",
1991 "src/runtime/NEON/functions/NEReductionOperation.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01001992 ]
1993 }
1994 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01001995 "Reorg": {
1996 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01001997 "common": [
1998 "src/core/NEON/kernels/NEReorgLayerKernel.cpp",
1999 "src/runtime/NEON/functions/NEReorgLayer.cpp"
2000 ]
2001 }
2002 },
2003 "Reshape": {
2004 "files": {
2005 "common": [
2006 "src/cpu/operators/CpuReshape.cpp",
2007 "src/cpu/kernels/CpuReshapeKernel.cpp",
2008 "src/runtime/NEON/functions/NEReshapeLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002009 ]
2010 }
2011 },
2012 "Reverse": {
2013 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002014 "common": [
2015 "src/core/NEON/kernels/NEReverseKernel.cpp",
2016 "src/runtime/NEON/functions/NEReverse.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002017 ]
2018 }
2019 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002020 "RNN": {
2021 "deps": [ "Activation", "Add", "FullyConnected", "Gemm"],
2022 "files": {
2023 "common": [ "src/runtime/NEON/functions/NERNNLayer.cpp" ]
2024 }
2025 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002026 "ROIAlign": {
2027 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002028 "common": [
2029 "src/core/NEON/kernels/NEROIAlignLayerKernel.cpp",
2030 "src/runtime/NEON/functions/NEROIAlignLayer.cpp"
Dana Zlotnik9ae5a4e2022-01-03 16:22:47 +02002031 ],
2032 "neon":{
2033 "common":["src/cpu/kernels/roialign/generic/neon/impl.cpp"],
2034 "fp32":["src/cpu/kernels/roialign/generic/neon/fp32.cpp"],
2035 "fp16":["src/cpu/kernels/roialign/generic/neon/fp16.cpp"],
2036 "qasymm8":["src/cpu/kernels/roialign/generic/neon/qasymm8.cpp"],
2037 "qasymm8_signed":["src/cpu/kernels/roialign/generic/neon/qasymm8_signed.cpp"]
2038 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002039 }
2040 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002041 "ROIPool2d": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01002042 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002043 "common": [
2044 "src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp",
2045 "src/runtime/NEON/functions/NEROIPoolingLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002046 ]
2047 }
2048 },
2049 "Scale": {
2050 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002051 "common": [
2052 "src/cpu/operators/CpuScale.cpp",
2053 "src/cpu/kernels/CpuScaleKernel.cpp",
2054 "src/runtime/NEON/functions/NEScale.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002055 ],
2056 "sve": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002057 "fp16": [ "src/cpu/kernels/scale/sve/fp16.cpp" ],
2058 "fp32": [ "src/cpu/kernels/scale/sve/fp32.cpp" ],
2059 "integer": [ "src/cpu/kernels/scale/sve/integer.cpp" ],
2060 "qasymm8": [ "src/cpu/kernels/scale/sve/qasymm8.cpp" ],
2061 "qasymm8_signed": [ "src/cpu/kernels/scale/sve/qasymm8_signed.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002062 },
2063 "neon": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002064 "fp16": [ "src/cpu/kernels/scale/neon/fp16.cpp" ],
2065 "integer": [ "src/cpu/kernels/scale/neon/integer.cpp" ],
Gunes Bayirc4f27432022-09-11 15:59:19 +01002066 "qasymm8": [ "src/cpu/kernels/scale/neon/qasymm8.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ],
2067 "qasymm8_signed": [ "src/cpu/kernels/scale/neon/qasymm8_signed.cpp", "src/cpu/kernels/scale/neon/integer.cpp" ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002068 }
2069 }
2070 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002071 "Select": {
2072 "files": {
2073 "common": [
2074 "src/core/NEON/kernels/NESelectKernel.cpp",
2075 "src/runtime/NEON/functions/NESelect.cpp"
Anton Vainer8a9a0fb2022-01-09 14:37:12 +02002076 ],
2077 "neon": {
2078 "common": [ "src/cpu/kernels/select/generic/neon/impl.cpp" ],
2079 "fp32": [ "src/cpu/kernels/select/generic/neon/fp32.cpp" ],
2080 "fp16": [ "src/cpu/kernels/select/generic/neon/fp16.cpp" ],
2081 "integer": [ "src/cpu/kernels/select/generic/neon/integer.cpp" ]
2082 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002083 }
2084 },
2085 "Slice": {
2086 "deps": [ "StridedSlice" ],
2087 "files": {
2088 "common": [ "src/runtime/NEON/functions/NESlice.cpp" ]
2089 }
2090 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002091 "Softmax": {
2092 "deps": [
2093 "Permute"
2094 ],
2095 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002096 "common": [
2097 "src/cpu/operators/CpuSoftmax.cpp",
2098 "src/cpu/kernels/CpuSoftmaxKernel.cpp",
2099 "src/runtime/NEON/functions/NESoftmaxLayer.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002100 ],
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002101 "neon":{
Dana Zlotnika538ae52022-02-21 13:12:41 +02002102 "common":["src/cpu/kernels/softmax/generic/neon/impl.cpp"],
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002103 "fp32": ["src/cpu/kernels/softmax/generic/neon/fp32.cpp"],
ramelg01a1f78512022-06-29 16:28:10 +01002104 "fp16": ["src/cpu/kernels/softmax/generic/neon/fp16.cpp"],
2105 "qasymm8":[ "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02002106 "qasymm8_signed":["src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp"]
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002107 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002108 "sve": {
Dana Zlotnikc48a3e52021-12-21 13:34:42 +02002109 "common": [ "src/cpu/kernels/softmax/generic/sve/impl.cpp" ],
2110 "fp32": ["src/cpu/kernels/softmax/generic/sve/fp32.cpp"],
2111 "fp16": ["src/cpu/kernels/softmax/generic/sve/fp16.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02002112 "qasymm8": ["src/cpu/kernels/softmax/generic/sve/qasymm8.cpp" ],
2113 "qasymm8_signed": ["src/cpu/kernels/softmax/generic/sve/qasymm8_signed.cpp"]
2114 },
2115 "sve2":{
Dana Zlotnika538ae52022-02-21 13:12:41 +02002116 "common" :["src/cpu/kernels/softmax/generic/sve2/impl.cpp"],
ramelg01a1f78512022-06-29 16:28:10 +01002117 "qasymm8":[ "src/cpu/kernels/softmax/generic/sve2/qasymm8.cpp"],
Dana Zlotnik6a2df882022-01-17 09:54:26 +02002118 "qasymm8_signed":["src/cpu/kernels/softmax/generic/sve2/qasymm8_signed.cpp"]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002119 }
2120 }
2121 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002122 "SpaceToBatch": {
2123 "files": {
2124 "common": [
2125 "src/core/NEON/kernels/NESpaceToBatchLayerKernel.cpp",
2126 "src/runtime/NEON/functions/NESpaceToBatchLayer.cpp"
2127 ]
2128 }
2129 },
2130 "SpaceToDepth": {
2131 "files": {
2132 "common": [
2133 "src/core/NEON/kernels/NESpaceToDepthLayerKernel.cpp",
2134 "src/runtime/NEON/functions/NESpaceToDepthLayer.cpp"
2135 ]
2136 }
2137 },
2138 "Split": {
2139 "deps": [ "StridedSlice" ],
2140 "files": {
2141 "common": [
2142 "src/runtime/NEON/functions/NESplit.cpp"
2143 ]
2144 }
2145 },
2146 "Stack": {
2147 "files": {
2148 "common": [
2149 "src/core/NEON/kernels/NEStackLayerKernel.cpp",
2150 "src/runtime/NEON/functions/NEStackLayer.cpp"
2151 ]
2152 }
2153 },
2154 "StridedSlice": {
2155 "files": {
2156 "common": [
2157 "src/core/NEON/kernels/NEStridedSliceKernel.cpp",
2158 "src/runtime/NEON/functions/NEStridedSlice.cpp"
2159 ]
2160 }
2161 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002162 "Sub": {
2163 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002164 "common": [
2165 "src/cpu/operators/CpuSub.cpp",
2166 "src/cpu/kernels/CpuSubKernel.cpp",
Freddie Liardet487d3902021-09-21 12:36:43 +01002167 "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp",
2168 "src/cpu/kernels/sub/neon/qasymm8.cpp",
2169 "src/cpu/kernels/sub/neon/qasymm8_signed.cpp",
2170 "src/cpu/kernels/sub/neon/qsymm16.cpp"
2171 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002172 }
2173 },
Michalis Spyrou20fca522021-06-07 14:23:57 +01002174 "Tile": {
2175 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002176 "common": [
2177 "src/core/NEON/kernels/NETileKernel.cpp",
2178 "src/runtime/NEON/functions/NETile.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002179 ]
2180 }
2181 },
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002182 "Transpose": {
Michalis Spyrou20fca522021-06-07 14:23:57 +01002183 "files": {
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002184 "common": [
2185 "src/cpu/kernels/CpuTransposeKernel.cpp",
2186 "src/cpu/operators/CpuTranspose.cpp",
2187 "src/runtime/NEON/functions/NETranspose.cpp"
Michalis Spyrou20fca522021-06-07 14:23:57 +01002188 ]
2189 }
Georgios Pinitasb6af4822021-09-14 12:33:34 +01002190 },
2191 "Unstack": {
2192 "deps": [ "StridedSlice" ],
2193 "files": {
2194 "common": [ "src/runtime/NEON/functions/NEUnstack.cpp" ]
2195 }
Michalis Spyrou20fca522021-06-07 14:23:57 +01002196 }
2197 }
Giorgio Arena232c4522022-03-03 10:09:01 +00002198 },
2199 "experimental": {
2200 "dynamic_fusion": [
SiCong Lif44bbc52022-08-29 18:25:51 +01002201 "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp",
2202 "src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp",
Gunes Bayir1dc6ff12022-12-06 20:48:31 +00002203 "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp",
Jakub Sujak32741722022-11-25 16:43:18 +00002204 "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp",
Gunes Bayir7dc02342022-11-21 21:46:50 +00002205 "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp",
Jakub Sujak8ae57142022-12-02 16:09:06 +00002206 "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp",
Gunes Bayiraecb5d92022-12-18 21:31:29 +00002207 "src/dynamic_fusion/sketch/attributes/SoftmaxAttributes.cpp",
SiCong Lif44bbc52022-08-29 18:25:51 +01002208 "src/dynamic_fusion/sketch/OperatorAttributes.cpp",
2209 "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.cpp",
2210 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp",
2211 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp",
2212 "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.cpp",
2213 "src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp",
2214 "src/dynamic_fusion/sketch/gpu/GpuOperatorGroup.cpp",
2215 "src/dynamic_fusion/sketch/gpu/GpuWorkloadContext.cpp",
2216 "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.cpp",
Jakub Sujak32741722022-11-25 16:43:18 +00002217 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.cpp",
Gunes Bayir1dc6ff12022-12-06 20:48:31 +00002218 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentCast.cpp",
Gunes Bayir7dc02342022-11-21 21:46:50 +00002219 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp",
SiCong Lif44bbc52022-08-29 18:25:51 +01002220 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp",
Ramy Elgammal404462a2022-11-08 02:14:46 +00002221 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp",
Gunes Bayiraecb5d92022-12-18 21:31:29 +00002222 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentLogits1DMaxShiftExpSum.cpp",
Jakub Sujak8ae57142022-12-02 16:09:06 +00002223 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp",
SiCong Lif44bbc52022-08-29 18:25:51 +01002224 "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp",
Ramy Elgammal404462a2022-11-08 02:14:46 +00002225 "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp",
Gunes Bayir1dc6ff12022-12-06 20:48:31 +00002226 "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp",
Jakub Sujak32741722022-11-25 16:43:18 +00002227 "src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp",
Gunes Bayir1dc6ff12022-12-06 20:48:31 +00002228 "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp",
Viet-Hoa Dob84e2532022-12-13 13:09:10 +00002229 "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp",
Jakub Sujak8ae57142022-12-02 16:09:06 +00002230 "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp",
Ramy Elgammal404462a2022-11-08 02:14:46 +00002231 "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp",
Gunes Bayir7dc02342022-11-21 21:46:50 +00002232 "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp",
Jakub Sujak32741722022-11-25 16:43:18 +00002233 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp",
Gunes Bayir1dc6ff12022-12-06 20:48:31 +00002234 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateCast.cpp",
Gunes Bayir7dc02342022-11-21 21:46:50 +00002235 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp",
SiCong Lif44bbc52022-08-29 18:25:51 +01002236 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp",
Ramy Elgammal404462a2022-11-08 02:14:46 +00002237 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp",
Gunes Bayiraecb5d92022-12-18 21:31:29 +00002238 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DMaxShiftExpSum.cpp",
Jakub Sujak8ae57142022-12-02 16:09:06 +00002239 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp",
SiCong Lif44bbc52022-08-29 18:25:51 +01002240 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp",
2241 "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp",
2242 "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp"
Giorgio Arena232c4522022-03-03 10:09:01 +00002243 ]
Michalis Spyrou20fca522021-06-07 14:23:57 +01002244 }
ramelg018a164882022-04-07 02:42:52 +01002245}