Modify aligned address of sa and sb to improve the performance of multi-threads.
This commit is contained in:
25
param.h
25
param.h
@@ -1493,33 +1493,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
#define ZGEMM_DEFAULT_UNROLL_N 2
|
||||
|
||||
#define SGEMM_DEFAULT_P 64
|
||||
#define DGEMM_DEFAULT_P 32
|
||||
#define DGEMM_DEFAULT_P 44
|
||||
#define CGEMM_DEFAULT_P 64
|
||||
#define ZGEMM_DEFAULT_P 32
|
||||
|
||||
#define SGEMM_DEFAULT_Q 192
|
||||
#define DGEMM_DEFAULT_Q 112
|
||||
#define CGEMM_DEFAULT_Q 192
|
||||
#define DGEMM_DEFAULT_Q 92
|
||||
#define CGEMM_DEFAULT_Q 128
|
||||
#define ZGEMM_DEFAULT_Q 80
|
||||
|
||||
#define SGEMM_DEFAULT_R 1024
|
||||
//#define DGEMM_DEFAULT_R 300
|
||||
//#define DGEMM_DEFAULT_R 200
|
||||
//#define DGEMM_DEFAULT_R 400
|
||||
//#define DGEMM_DEFAULT_R 192
|
||||
#define DGEMM_DEFAULT_R dgemm_r
|
||||
//1000
|
||||
//#define DGEMM_DEFAULT_R 160
|
||||
//#define DGEMM_DEFAULT_R 270
|
||||
#define DGEMM_DEFAULT_R dgemm_r
|
||||
#define CGEMM_DEFAULT_R 1024
|
||||
//#define ZGEMM_DEFAULT_R 1000
|
||||
#define ZGEMM_DEFAULT_R 1000
|
||||
#define ZGEMM_DEFAULT_R 1024
|
||||
|
||||
#define GEMM_OFFSET_A1 (DGEMM_DEFAULT_P*DGEMM_DEFAULT_Q*SIZE)
|
||||
//#define GEMM_OFFSET_B1 0x10
|
||||
#define GEMM_OFFSET_B1 (DGEMM_DEFAULT_Q*DGEMM_DEFAULT_R*SIZE)
|
||||
#define GEMM_OFFSET 0x100000
|
||||
#define GEMM_OFFSET1 0x40000
|
||||
#define GEMM_OFFSET_A1 0x10000
|
||||
#define GEMM_OFFSET_B1 0x100000
|
||||
|
||||
#define SYMV_P 16
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user