Zhang Xianyi
13 years ago
2 changed files with 9 additions and 2 deletions
Unified View
Diff Options
-
+5 -1kernel/x86_64/axpy_sse.S
-
+4 -1kernel/x86_64/axpy_sse2.S
+ 5
- 1
kernel/x86_64/axpy_sse.S
View File
| @@ -69,7 +69,7 @@ | |||||
| #endif | #endif | ||||
| movaps %xmm0, ALPHA | movaps %xmm0, ALPHA | ||||
| #else | #else | ||||
| movaps %xmm3, ALPHA | |||||
| movq 40(%rsp), X | movq 40(%rsp), X | ||||
| movq 48(%rsp), INCX | movq 48(%rsp), INCX | ||||
| @@ -79,6 +79,10 @@ | |||||
| SAVEREGISTERS | SAVEREGISTERS | ||||
| #ifdef WINDOWS_ABI | |||||
| movaps %xmm3, ALPHA | |||||
| #endif | |||||
| shufps $0, ALPHA, ALPHA | shufps $0, ALPHA, ALPHA | ||||
| leaq (, INCX, SIZE), INCX | leaq (, INCX, SIZE), INCX | ||||
+ 4
- 1
kernel/x86_64/axpy_sse2.S
View File
| @@ -69,7 +69,6 @@ | |||||
| #endif | #endif | ||||
| movaps %xmm0, ALPHA | movaps %xmm0, ALPHA | ||||
| #else | #else | ||||
| movaps %xmm3, ALPHA | |||||
| movq 40(%rsp), X | movq 40(%rsp), X | ||||
| movq 48(%rsp), INCX | movq 48(%rsp), INCX | ||||
| @@ -79,6 +78,10 @@ | |||||
| SAVEREGISTERS | SAVEREGISTERS | ||||
| #ifdef WINDOWS_ABI | |||||
| movaps %xmm3, ALPHA | |||||
| #endif | |||||
| unpcklpd ALPHA, ALPHA | unpcklpd ALPHA, ALPHA | ||||
| leaq (, INCX, SIZE), INCX | leaq (, INCX, SIZE), INCX | ||||