pytorch

79b9c034 - Optimize torch zeros (#45636)

Optimize torch zeros (#45636) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/45636 After creating empty tensor 'memset' used to zero out items of tensor Test Plan: pytorch benchmark tool results: timer = benchmark_utils.Timer(stmt="torch.zeros((1024, 4096))") Before: 1007 us After: 841.26 us 1 measurement, 10000 runs , 1 thread timer = benchmark_utils.Timer(stmt="torch.zeros((128))") Before: 4 - 7.6 us After: 2.4 - 2.8 us 1 measurement, 10000 runs , 1 thread torch.int8 | 1 | 4096 | 8192 | 16384 | 32768 | 1 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 500 | 600 | 700 | 2000 | (Reference) x.zero_() | 800 | 1000 | 1000 | 2000 | 2000 | 2 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 500 | 600 | 700 | 2000 | (Reference) x.zero_() | 800 | 1000 | 1000 | 2000 | 3000 | 4 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 500 | 600 | 700 | 2000 | (Reference) x.zero_() | 800 | 1000 | 1000 | 2000 | 3000 | torch.int32 | 1 | 4096 | 8192 | 16384 | 32768 | 1 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 400 | 700 | 2000 | 2900 | 5500 | (Reference) x.zero_() | 800 | 2000 | 3000 | 4400 | 7300 | 2 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 700 | 2000 | 3000 | 5600 | (Reference) x.zero_() | 900 | 2000 | 2000 | 3600 | 7200 | 4 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 400 | 700 | 2000 | 3000 | 5700 | (Reference) x.zero_() | 800 | 2000 | 3100 | 4300 | 9000 | torch.float16 | 1 | 4096 | 8192 | 16384 | 32768 | 1 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 500 | 700 | 2000 | 3000 | (Reference) x.zero_() | 800 | 1000 | 2000 | 2000 | 3300 | 2 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 600 | 700 | 2000 | 3000 | (Reference) x.zero_() | 800 | 1000 | 2000 | 2000 | 4300 | 4 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 600 | 700 | 2000 | 3300 | (Reference) x.zero_() | 900 | 1000 | 2000 | 2000 | 4400 | torch.float32 | 1 | 4096 | 8192 | 16384 | 32768 | 1 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 700 | 2000 | 3200 | 6100 | (Reference) x.zero_() | 800 | 2000 | 2000 | 3500 | 6100 | 2 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 700 | 2000 | 3100 | 5600 | (Reference) x.zero_() | 800 | 2000 | 2000 | 3300 | 7000 | 4 threads: -------------------------------------------------------------- (PR #45636) x.zero_() | 500 | 700 | 2000 | 3000 | 5600 | (Reference) x.zero_() | 900 | 2000 | 2000 | 3600 | 7500 | Reviewed By: ngimel Differential Revision: D23925113 fbshipit-source-id: 04e97ff6d67c52a8e7a21449113e1a0a7443098f