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

Restore ZEN SGEMM speed after #2361. 

I partially reverted the changes in #2361 and I received the following
speed up on:
./xsl3blastst -R gemm -N 2048 2048 1 -a 5 1 1 1 1 1

AMD Ryzen 7 2700X (Zen+): 61400 to 63300 MFlops
AMD EPYC 7742 (Zen v2): 91400 to 94500 MFlops

These numbers are single-threaded performance.
This commit is contained in:
Martin Liska 2020-02-18 11:52:06 +01:00
parent 6a517dcb6a
commit 9fb54ee1da
No known key found for this signature in database
GPG Key ID: 4DC182DC0FA73785
2 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
kernel/x86_64/KERNEL.ZEN
View File

@ -30,8 +30,8 @@ DAXPYKERNEL = daxpy.c
CAXPYKERNEL = caxpy.c CAXPYKERNEL = caxpy.c
ZAXPYKERNEL = zaxpy.c ZAXPYKERNEL = zaxpy.c
STRMMKERNEL = sgemm_kernel_8x4_haswell.c STRMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMKERNEL = sgemm_kernel_8x4_haswell.c SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
SGEMMINCOPY = ../generic/gemm_ncopy_8.c SGEMMINCOPY = ../generic/gemm_ncopy_8.c
SGEMMITCOPY = ../generic/gemm_tcopy_8.c SGEMMITCOPY = ../generic/gemm_tcopy_8.c
SGEMMONCOPY = ../generic/gemm_ncopy_4.c SGEMMONCOPY = ../generic/gemm_ncopy_4.c

2
param.h
View File

@ -666,7 +666,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else #else
#define SGEMM_DEFAULT_P 320 #define SGEMM_DEFAULT_P 768
#define DGEMM_DEFAULT_P 512 #define DGEMM_DEFAULT_P 512
#define CGEMM_DEFAULT_P 256 #define CGEMM_DEFAULT_P 256
#define ZGEMM_DEFAULT_P 192 #define ZGEMM_DEFAULT_P 192