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

Refs #113. Support AMD Bobcate using Barcelona kernel codes. Replace 3DNow! with MMX.

Browse Source
This commit is contained in:
Zhang Xianyi
2012-05-31 18:17:45 +08:00
parent 37edae1c90
commit d6cab3f37e
29 changed files with 303 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
kernel/x86/trsm_kernel_RT_2x4_sse2.S
View File

@@ -69,7 +69,7 @@
#define STACK_ALIGN 4096
#define STACK_OFFSET 1024
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
#define PREFETCH prefetch
#define PREFETCHSIZE (8 * 10 + 4)
#endif
@@ -1036,7 +1036,7 @@
.L42:
mulpd %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
mulpd 2 * SIZE(BB), %xmm0
@@ -1066,7 +1066,7 @@
addpd %xmm0, %xmm7
movapd 16 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
#endif
mulpd %xmm1, %xmm2
@@ -2224,7 +2224,7 @@
.L22:
mulsd %xmm0, %xmm2
addsd %xmm2, %xmm4
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movlpd 2 * SIZE(BB), %xmm2
@@ -2273,7 +2273,7 @@
movlpd 40 * SIZE(BB), %xmm3
addsd %xmm0, %xmm7
movlpd 8 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
#endif
mulsd %xmm1, %xmm2