mirror of
https://github.com/amd/blis.git
synced 2026-05-24 02:14:33 +00:00
4c2f436cce
Description: 1. GCC avoiding loading b into registers in m fringe kenrels of int8 kernels. Instead gcc generating fma with memory as an operand for B input. 2. This is causing performance regression for larger n where each fma needs to load the input from memory again and again. 3. This is observed with gcc but not with clang. 4. Inserted dummy shuffle instructions for b data to further explicitly tell compiler that b needs to be in registers. 5. Moved packb_s4_to_bf16 under JIT macro to resovle compilation issue with gcc version < 11.2 AMD-Internal: SWLCSG-2948 Change-Id: I5bd1bad7ad129e0dde91ed78d49a4ede3bff456a