mirror of
https://github.com/boykopovar/AnyPS5.git
synced 2026-11-06 13:30:00 +00:00
JJenkxandClaude Opus 5.5
eb0b59ccc2
Byte stores beat MASKMOVDQU at every density on a Ryzen 9 7950X, but on a Core i9-14900KF the masked store costs about the same at every density and byte stores win only up to 3 or 4 changed bytes, up to 2.6 times slower for dense chunks (#8930 review). A partly changed chunk now takes byte stores when at most ChangedByteStoreLimit() bytes changed and MASKMOVDQU otherwise; the limit is 16 on an AuthenticAMD CPU and 4 elsewhere. The changed bytes are counted with psadbw, since builds have no popcnt. Co-Authored-By: Claude Opus 5.5 <[email protected]>