flame / blis

BLAS-like Library Instantiation Software Framework

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

SKX configuration missing gemmsup kernels

m-henning-toptal opened this issue · comments

While working with a client, we noticed that our software running with blis was slower on a recent processor (Intel rocket lake) than it was on an older processor (Haswell). Some testing revealed that this is because the SKX target in blis is missing gemmsup kernels, while the haswell target has them. We're using blis on windows with the x86_64 configuration.

Adding the haswell gemmsup kernels to bli_cntx_init_skx.c was enough to solve the performance issue for us. (We used the same thresholds as defined in bli_cntx_init_haswell.c)