mirror of
https://github.com/libretro/RetroArch.git
synced 2026-11-22 10:50:15 -10:00
Rewind diffing (find_change for the next dirty word, find_same for the end of a dirty run) gains an AVX-512BW tier next to SSE2 and AVX2, and find_same gets the AVX2 form it lacked. Tiers are picked once from cpu_features_get(); RETRO_SIMD_AVX512 promises BW. Three things make the wide scans pay off rather than lose to SSE2: - The two blocks are now 64-byte aligned and a whole number of lines apart, so a scan position has the same offset within a line in both. The wide loops step back to a line boundary inside bytes already proven equal, and from there no load straddles a cache line on either block. Before this, AVX-512 lost to SSE2 on dirty states. - Each scan starts narrow (16 bytes, then 32) and only then goes to full width, since most scans in a dirty state end within the first few words. - raw_compress is compiled once per tier with the scans inlined. A target function cannot inline into baseline code, and a call per scan cost more than the scan on dirty states. All tiers write byte-identical patches. Median time against SSE2, paired runs on an Emerald Rapids Xeon (lower is faster): state pattern avx2 avx512 16K clean 0.49 0.28 64K 1 dirty in 4096 0.60 0.45 256K 16-word runs 0.61 0.49 256K 1 dirty in 256 0.70 0.61 1M clean 0.85 0.84 any 1 dirty in 64 ~1.0 ~1.0-1.05 4M+ any ~1.0 ~1.0 (memory bound) samples/state_rewind allocates its blocks the same way and checks that every tier the CPU has writes the same patch over 300 seeded pairs, with the sentinel and scan-pad cases re-run under each tier.