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

Added BULLDOZER target. So far it uses barcelona kernels.

Browse Source
This commit is contained in:
Zhang Xianyi
2012-12-07 00:53:31 +08:00
parent b7c0fa6bd2
commit bfaaa975e6
47 changed files with 156 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
kernel/x86/gemm_kernel_4x4_barcelona.S
View File

@@ -596,7 +596,7 @@
.L22:
mulps %xmm0, %xmm2
addps %xmm2, %xmm4
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movsd 4 * SIZE(BB), %xmm2
@@ -842,7 +842,7 @@
.L32:
mulss %xmm0, %xmm2
addss %xmm2, %xmm4
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movss 4 * SIZE(BB), %xmm2
@@ -1168,7 +1168,7 @@
.L52:
mulps %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
mulps 4 * SIZE(BB), %xmm0
@@ -1198,7 +1198,7 @@
addps %xmm0, %xmm5
movaps 32 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
#endif
mulps %xmm1, %xmm2
@@ -1347,7 +1347,7 @@
ALIGN_4
.L62:
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
@@ -1531,7 +1531,7 @@
.L72:
mulss %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
mulss 4 * SIZE(BB), %xmm0
@@ -1778,7 +1778,7 @@
.L92:
mulps %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movaps 4 * SIZE(AA), %xmm0
@@ -1793,7 +1793,7 @@
mulps 12 * SIZE(BB), %xmm0
addps %xmm0, %xmm7
movaps 32 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
#endif
mulps %xmm1, %xmm3
@@ -1924,7 +1924,7 @@
.L102:
mulps %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movsd 2 * SIZE(AA), %xmm0
@@ -2069,7 +2069,7 @@
.L112:
mulss %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movss 1 * SIZE(AA), %xmm0