pytorch

5a02930d - Vectorize (CPU) generic types for binary bitwise operators (#34338)

Vectorize (CPU) generic types for binary bitwise operators (#34338) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/34338 For those types not optimized for AVX2, this commit would give bitwise operations on them a boost. Benchmark (RHEL 7.7, Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz, Turbo off, Release build): ```python import timeit for op in ('bitwise_and', 'bitwise_or', 'bitwise_xor'): for dtype in ('torch.int8', 'torch.uint8'): for n, t in [(10_000, 200000), (100_000, 20000)]: print(f'a.{op}_(b), numel() == {n} for {t} times, dtype={dtype}') print(timeit.timeit(f'a.{op}_(b)', setup=f'import torch; a = torch.arange(1, {n}, dtype={dtype}); b = torch.arange({n}, 1, -1, dtype={dtype})', number=t)) ``` Before: ``` a.bitwise_and_(b), numel() == 10000 for 200000 times, dtype=torch.int8 1.353799690001324 a.bitwise_and_(b), numel() == 100000 for 20000 times, dtype=torch.int8 1.056434961999912 a.bitwise_and_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 1.2957618809996347 a.bitwise_and_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 1.0591609650000464 a.bitwise_or_(b), numel() == 10000 for 200000 times, dtype=torch.int8 1.3113185389993305 a.bitwise_or_(b), numel() == 100000 for 20000 times, dtype=torch.int8 1.0693870880022587 a.bitwise_or_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 1.3075691039994126 a.bitwise_or_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 1.0589785859992844 a.bitwise_xor_(b), numel() == 10000 for 200000 times, dtype=torch.int8 1.3036618039986934 a.bitwise_xor_(b), numel() == 100000 for 20000 times, dtype=torch.int8 1.0595013140009542 a.bitwise_xor_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 1.2947387999993225 a.bitwise_xor_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 1.059969027999614 ``` After: ``` a.bitwise_and_(b), numel() == 10000 for 200000 times, dtype=torch.int8 0.9562859639991075 a.bitwise_and_(b), numel() == 100000 for 20000 times, dtype=torch.int8 0.6811799210008758 a.bitwise_and_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 0.9522694869992847 a.bitwise_and_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 0.6815469840003061 a.bitwise_or_(b), numel() == 10000 for 200000 times, dtype=torch.int8 0.8609786279994296 a.bitwise_or_(b), numel() == 100000 for 20000 times, dtype=torch.int8 0.5794818879985542 a.bitwise_or_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 0.8534434389985108 a.bitwise_or_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 0.5764101290005783 a.bitwise_xor_(b), numel() == 10000 for 200000 times, dtype=torch.int8 0.9634105910008657 a.bitwise_xor_(b), numel() == 100000 for 20000 times, dtype=torch.int8 0.6819724230008433 a.bitwise_xor_(b), numel() == 10000 for 200000 times, dtype=torch.uint8 1.0901075929978106 a.bitwise_xor_(b), numel() == 100000 for 20000 times, dtype=torch.uint8 0.816546294001455 ``` Test Plan: Imported from OSS Differential Revision: D20687081 Pulled By: ezyang fbshipit-source-id: 59b06460430ce181fb761e45a5bdd6379611b391