pytorch

e51d937e - move cuda abs to Aten (#25857)

move cuda abs to Aten (#25857) Summary: VitalyFedyunin, this PR fix the https://github.com/pytorch/pytorch/issues/24531 Benchmark script : ``` import timeit device = "cuda" for n, t in [(10, 100000),(1000, 10000)]: print('a.abs() (a.numel() == {}) for {} times'.format(n, t)) for dtype in ('torch.int8', 'torch.uint8', 'torch.int16', 'torch.int32', 'torch.int64', 'torch.float', 'torch.double', 'torch.half'): print(f'device: {device}, dtype: {dtype}, {t} times', end='\t\t') print(timeit.timeit(f'a.abs()\nif "{device}" == "cuda": torch.cuda.synchronize()', setup=f'import torch; a = torch.ones({n}, device="{device}", dtype={dtype})', number=t)) ``` Device: **Tesla P4** Cuda verison: **9.0.176** Before this change: ``` a.abs() (a.numel() == 10) for 100000 times device: cuda, dtype: torch.int8, 100000 times 1.8391285985708237 device: cuda, dtype: torch.uint8, 100000 times 1.8831938095390797 device: cuda, dtype: torch.int16, 100000 times 1.8131775446236134 device: cuda, dtype: torch.int32, 100000 times 1.832334715873003 device: cuda, dtype: torch.int64, 100000 times 1.8218239657580853 device: cuda, dtype: torch.float, 100000 times 1.7942761108279228 device: cuda, dtype: torch.double, 100000 times 1.8193779103457928 device: cuda, dtype: torch.half, 100000 times 1.796515878289938 a.abs() (a.numel() == 1000) for 10000 times device: cuda, dtype: torch.int8, 10000 times 0.18348361551761627 device: cuda, dtype: torch.uint8, 10000 times 0.1892806850373745 device: cuda, dtype: torch.int16, 10000 times 0.18253886327147484 device: cuda, dtype: torch.int32, 10000 times 0.18509215489029884 device: cuda, dtype: torch.int64, 10000 times 0.18291602283716202 device: cuda, dtype: torch.float, 10000 times 0.1796952784061432 device: cuda, dtype: torch.double, 10000 times 0.18088893592357635 device: cuda, dtype: torch.half, 10000 times 0.18222836777567863 ``` After change: ```a.abs() (a.numel() == 10) for 100000 times device: cuda, dtype: torch.int8, 100000 times 1.7365420907735825 device: cuda, dtype: torch.uint8, 100000 times 1.7433889284729958 device: cuda, dtype: torch.int16, 100000 times 1.7034666128456593 device: cuda, dtype: torch.int32, 100000 times 1.6825932636857033 device: cuda, dtype: torch.int64, 100000 times 1.6896217577159405 device: cuda, dtype: torch.float, 100000 times 1.7211194895207882 device: cuda, dtype: torch.double, 100000 times 1.6823345720767975 device: cuda, dtype: torch.half, 100000 times 1.7027524448931217 a.abs() (a.numel() == 1000) for 10000 times device: cuda, dtype: torch.int8, 10000 times 0.17180879414081573 device: cuda, dtype: torch.uint8, 10000 times 0.17316896095871925 device: cuda, dtype: torch.int16, 10000 times 0.16990498825907707 device: cuda, dtype: torch.int32, 10000 times 0.1681906059384346 device: cuda, dtype: torch.int64, 10000 times 0.16994905844330788 device: cuda, dtype: torch.float, 10000 times 0.1719626784324646 device: cuda, dtype: torch.double, 10000 times 0.16886932775378227 device: cuda, dtype: torch.half, 10000 times 0.16957201063632965 ``` Pull Request resolved: https://github.com/pytorch/pytorch/pull/25857 Differential Revision: D18299368 Pulled By: VitalyFedyunin fbshipit-source-id: 173eb0f6ca5a12a27f3d53466ff373a5f81f1da8