sha256: small speedup for sse4 path. #13400

pull theuni wants to merge 2 commits into bitcoin:master from theuni:sha2-avx1 changing 1 files +5 −4

theuni commented at 7:06 PM on June 5, 2018: member

This is an optimization described in the intel sha256 whitepaper (page 10). It speeds up the SHA256D64_1024 bench (sse4 path, no avx2) for me by ~6%.
crypto: split out Rotations a2724af487
crypto: sha256 optim: reduce register copies ea3ed0cbd2
sipa commented at 7:07 PM on June 5, 2018: member

ACK, benchmarked to be around 5% faster for SSE4 (when disabling the AVX2 code on my i7-7820HQ).
MarcoFalke added the label Refactoring on Jun 5, 2018
theuni commented at 8:17 PM on June 5, 2018: member

Sadly, @laanwj saw a 25% penalty for this change on pre-avx2 AMD. Closing :(
theuni closed this on Jun 5, 2018
theuni cross-referenced this on Jun 5, 2018 from issue SHA256 implementations based on Intel SHA Extensions by sipa
theuni referenced this in commit 4ed6f4fc90 on Jun 12, 2018
theuni referenced this in commit 6da1fe9d69 on Jun 12, 2018
theuni referenced this in commit 4ee6fbb8b7 on Jun 12, 2018
bitcoin locked this on Sep 8, 2021

Contributors

Labels