floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sbgemm: spr: tuning for blocking params

This commit is contained in:
Wangyang Guo 2021-09-13 01:44:53 -07:00
parent a70bfb52d5
commit 0abbcd19c1
1 changed files with 14 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
param.h
View File

@ -1771,6 +1771,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif #endif
#define USE_SGEMM_KERNEL_DIRECT 1 #define USE_SGEMM_KERNEL_DIRECT 1
#undef SBGEMM_DEFAULT_UNROLL_N
#undef SBGEMM_DEFAULT_UNROLL_M
#undef SBGEMM_DEFAULT_P
#undef SBGEMM_DEFAULT_R
#undef SBGEMM_DEFAULT_Q
// FIXME: actually UNROLL_M = UNROLL_N = 16
// If M and N is equal, OpenBLAS will reuse OCOPY as ICOPY.
// But for AMX, they are not the same, set UNROLL_M = 32 to workaround
#define SBGEMM_DEFAULT_UNROLL_N 16
#define SBGEMM_DEFAULT_UNROLL_M 32
#define SBGEMM_DEFAULT_P 192
#define SBGEMM_DEFAULT_Q 1024
#define SBGEMM_DEFAULT_R sbgemm_r
#ifdef ARCH_X86 #ifdef ARCH_X86
#define SGEMM_DEFAULT_UNROLL_M 4 #define SGEMM_DEFAULT_UNROLL_M 4