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

loongarch64: Add buffer offset for target LOONGSON3R5

This commit is contained in:
gxw 2024-05-08 14:40:48 +08:00
parent 6017ad7146
commit 637c650f4f
2 changed files with 9 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
driver/others/blas_server.c
View File

@ -1074,13 +1074,10 @@ fprintf(STDERR, "Server[%2ld] Calculation started. Mode = 0x%03x M = %3ld N=%3l
main_status[cpu] = MAIN_RUNNING1;
#endif
//For Loongson servers, like the 3C5000 (featuring 16 cores), applying an
//offset to the buffer is essential for minimizing cache conflicts and optimizing performance.
#if defined(LOONGSON3R5) && !defined(NO_AFFINITY)
char model_name[128];
get_cpu_model(model_name);
if ((strstr(model_name, "3C5000") != NULL) || (strstr(model_name, "3D5000") != NULL))
if (sa == NULL) sa = (void *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
//For target LOONGSON3R5, applying an offset to the buffer is essential
//for minimizing cache conflicts and optimizing performance.
#if defined(ARCH_LOONGARCH64) && !defined(NO_AFFINITY)
if (sa == NULL) sa = (void *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
#endif
if (sa == NULL) sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
@ -1157,4 +1154,4 @@ if (queue -> mode & BLAS_PTHREAD) {
}
#endif
#endif

13
interface/gemm.c
View File

@ -521,15 +521,10 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
buffer = (XFLOAT *)blas_memory_alloc(0);
//For Loongson servers, like the 3C5000 (featuring 16 cores), applying an
//offset to the buffer is essential for minimizing cache conflicts and optimizing performance.
#if defined(LOONGSON3R5) && !defined(NO_AFFINITY)
char model_name[128];
get_cpu_model(model_name);
if ((strstr(model_name, "3C5000") != NULL) || (strstr(model_name, "3D5000") != NULL))
sa = (XFLOAT *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
else
sa = (XFLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
//For target LOONGSON3R5, applying an offset to the buffer is essential
//for minimizing cache conflicts and optimizing performance.
#if defined(ARCH_LOONGARCH64) && !defined(NO_AFFINITY)
sa = (XFLOAT *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
#else
sa = (XFLOAT *)((BLASLONG)buffer +GEMM_OFFSET_A);
#endif