-
Notifications
You must be signed in to change notification settings - Fork 2.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Disable auto vectorization of xxhash64, when AVX512 is present. #3819
Comments
Indeed, |
List of updates : https://github.com/Cyan4973/xxHash/releases/tag/v0.8.2 This is also a preparation task before taking care of #3819
Someone asked me about AMD offline so might as well post here. On genoa (Zen4) AVX512 is actually ~30% faster ( 1135871 ns/op vs 885935 ns/op). |
OK, so it makes the situation a bit less clear, However, it doesn't remove the question of "what's a good default", |
Regression on Intel is bigger than gain on AMD, so disabled by default makes sense. Using the same |
List of updates : https://github.com/Cyan4973/xxHash/releases/tag/v0.8.2 This is also a preparation task before taking care of facebook#3819
List of updates : https://github.com/Cyan4973/xxHash/releases/tag/v0.8.2 This is also a preparation task before taking care of facebook#3819
List of updates : https://github.com/Cyan4973/xxHash/releases/tag/v0.8.2 This is also a preparation task before taking care of facebook#3819
AVX512 adds support for
VPMULLQ
instructions, which makes is possible to vectorizeXXH64_update
. Here is a minimal reproducer where clang already vectrozies it with avx512 enabled: https://godbolt.org/z/cnM3986dx . Unfortunately some architectures (e.g skylake) have a very significant latency (uops.info report 15+ cycles), so this makesxxhash
~2x slower. Trivial benchmark showsBM_hash/16M 1.29ms ± 3%
vsBM_hash/16M 2.74ms ± 4%
. We already disable vectorization for xxhash32 if sse4 is present, so we should probably do the same for xxhash64 and avx512.The text was updated successfully, but these errors were encountered: