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

Merge pull request #4686 from XiWeiGu/loongarch64_dgemm_kernel_16x6

Browse Source 
Loongarch64: Improving the Performance and Stability of dgemm
This commit is contained in:
Martin Kroeker
2024-05-10 11:29:12 +02:00
committed by GitHub
parent 5d678f1831 f9a26240a7
commit 8da6f7e5f2
25 changed files with 12711 additions and 2489 deletions
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