lichess-org/fishnet

support SF armv8-dotprod arch

Closed this issue · 2 comments

Hi,

on Ampere Altra CPUs (Arm v8.2+), SF 16 armv8-dotprod builds consistently run ~9% faster as armv8:

armv8

Total time (ms) : 2939
Nodes searched  : 2593605
Nodes/second    : 882478

armv8-dotprod

Total time (ms) : 2694
Nodes searched  : 2593605
Nodes/second    : 962733

I have no idea how to detect the availability, CARGO_CFG_TARGET_FEATURE doesn't know about it, it shows only neon. Maybe it could be enabled manually via flag. Just wanted to share this finding, roughly 10% is not bad.

Makes sense.

CARGO_CFG_TARGET_FEATURE only has it if it was specifically selected, or with -C target-cpu=native. Runtime detection with https://doc.rust-lang.org/stable/std/arch/macro.is_aarch64_feature_detected.html also works.

Will be included in the next release.

Very cool!