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

Support AMD Piledriver by bulldozer kernels.

Browse Source
This commit is contained in:
Zhang Xianyi
2013-07-06 12:06:43 -03:00
parent 0c4074e10b
commit 886cbaf4e4
30 changed files with 377 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
kernel/x86/trsm_kernel_LT_2x4_sse2.S
View File

@@ -69,7 +69,7 @@
#define STACK_ALIGN 4096
#define STACK_OFFSET 1024
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
#define PREFETCH prefetch
#define PREFETCHSIZE (8 * 10 + 4)
#endif
@@ -910,7 +910,7 @@
.L22:
mulsd %xmm0, %xmm2
addsd %xmm2, %xmm4
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
#endif
movlpd 2 * SIZE(BB), %xmm2
@@ -959,7 +959,7 @@
movlpd 40 * SIZE(BB), %xmm3
addsd %xmm0, %xmm7
movlpd 8 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
#endif
mulsd %xmm1, %xmm2
@@ -1439,7 +1439,7 @@
.L42:
mulpd %xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
#endif
mulpd 2 * SIZE(BB), %xmm0
@@ -1469,7 +1469,7 @@
addpd %xmm0, %xmm7
movapd 16 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
#endif
mulpd %xmm1, %xmm2