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

Tune generic SVE constants closer to other SVE cores 

This allows us to use the `ARMV8` and `ARMV8SVE` targets in a minimal build without losing too much performance.
This commit is contained in:
Chris Sidebottom 2024-08-01 17:01:44 +01:00
parent 9afd0c8afd
commit 22d8803593
1 changed files with 6 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
param.h
View File

@ -3671,8 +3671,10 @@ Until then, just keep it different than DGEMM_DEFAULT_UNROLL_N to keep copy rout
#if defined(XDOUBLE) || defined(DOUBLE)
#define SWITCH_RATIO 8
#define GEMM_PREFERED_SIZE 4
#else
#define SWITCH_RATIO 16
#define GEMM_PREFERED_SIZE 8
#endif
#define SGEMM_DEFAULT_UNROLL_M 4 // Actually 1VL (8) but kept seperate to keep copies seperate
@ -3689,13 +3691,13 @@ Until then, just keep it different than DGEMM_DEFAULT_UNROLL_N to keep copy rout
#define ZGEMM_DEFAULT_UNROLL_N 4
#define ZGEMM_DEFAULT_UNROLL_MN 16
#define SGEMM_DEFAULT_P 128
#define DGEMM_DEFAULT_P 160
#define SGEMM_DEFAULT_P 240
#define DGEMM_DEFAULT_P 240
#define CGEMM_DEFAULT_P 128
#define ZGEMM_DEFAULT_P 128
#define SGEMM_DEFAULT_Q 352
#define DGEMM_DEFAULT_Q 128
#define SGEMM_DEFAULT_Q 640
#define DGEMM_DEFAULT_Q 320
#define CGEMM_DEFAULT_Q 224
#define ZGEMM_DEFAULT_Q 112