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
kernel/x86_64/gemm_kernel_8x4_barcelona.S
+19 -19
View File
@@ -930,7 +930,7 @@
.L22:
mulps %xmm8, %xmm9
addps %xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movaps 4 * SIZE(BO), %xmm9
@@ -983,7 +983,7 @@
addps %xmm8, %xmm3
movaps 0 * SIZE(AO), %xmm8
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)
#endif
mulps %xmm10, %xmm9
@@ -1178,7 +1178,7 @@
.L32:
mulps %xmm8, %xmm9
addps %xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movsd 4 * SIZE(BO), %xmm9
@@ -1423,7 +1423,7 @@
.L42:
mulss %xmm8, %xmm9
addss %xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movss 4 * SIZE(BO), %xmm9
@@ -1765,7 +1765,7 @@
.L62:
mulps %xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
mulps 4 * SIZE(BO), %xmm8
@@ -1793,7 +1793,7 @@
addps %xmm8, %xmm5
movaps 32 * SIZE(AO), %xmm8
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)
#endif
mulps %xmm10, %xmm11
@@ -1822,7 +1822,7 @@
addps %xmm10, %xmm5
movaps 48 * SIZE(AO), %xmm10
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 32) * SIZE(AO)
#endif
mulps %xmm12, %xmm13
@@ -1851,7 +1851,7 @@
addps %xmm12, %xmm5
movaps 64 * SIZE(AO), %xmm12
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 48) * SIZE(AO)
#endif
mulps %xmm14, %xmm15
@@ -2024,7 +2024,7 @@
.L72:
mulps %xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
@@ -2208,7 +2208,7 @@
.L82:
mulps %xmm8, %xmm9
addps %xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movsd 4 * SIZE(BO), %xmm9
@@ -2395,7 +2395,7 @@
.L92:
mulps %xmm8, %xmm9
addps %xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movss 4 * SIZE(BO), %xmm9
@@ -2670,7 +2670,7 @@
.L112:
mulps %xmm9, %xmm8
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
@@ -2687,7 +2687,7 @@
addps %xmm9, %xmm4
movaps 8 * SIZE(BO), %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)
#endif
mulps %xmm9, %xmm10
@@ -2704,7 +2704,7 @@
addps %xmm9, %xmm4
movaps 32 * SIZE(BO), %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 32) * SIZE(AO)
#endif
mulps %xmm11, %xmm12
@@ -2721,7 +2721,7 @@
addps %xmm11, %xmm4
movaps 24 * SIZE(BO), %xmm11
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 48) * SIZE(AO)
#endif
mulps %xmm11, %xmm14
@@ -2857,7 +2857,7 @@
.L122:
mulps %xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movaps -28 * SIZE(AO), %xmm8
@@ -2873,7 +2873,7 @@
addps %xmm8, %xmm3
movaps 0 * SIZE(AO), %xmm8
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)
#endif
mulps %xmm10, %xmm11
@@ -3003,7 +3003,7 @@
.L132:
mulps %xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movsd -30 * SIZE(AO), %xmm8
@@ -3150,7 +3150,7 @@
.L142:
mulss %xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
#endif
movss -31 * SIZE(AO), %xmm8
kernel/x86_64/gemm_ncopy_4_opteron.S
+2 -2
View File
@@ -39,7 +39,7 @@
#define ASSEMBLER
#include "common.h"
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define RPREFETCHSIZE (12 + 4)
#define WPREFETCHSIZE (48 + 4)
#define MOVNTQ MOVQ
@@ -79,7 +79,7 @@
#define AO3 %r13
#define AO4 %rax
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define RPREFETCH prefetch
#else
#define RPREFETCH prefetch
kernel/x86_64/gemm_tcopy_4_opteron.S
+2 -2
View File
@@ -39,7 +39,7 @@
#define ASSEMBLER
#include "common.h"
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define RPREFETCHSIZE (12 + 4)
#define WPREFETCHSIZE (12 + 4)
#define MOVNTQ MOVQ
@@ -96,7 +96,7 @@
#endif
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define RPREFETCH prefetch
#else
#define RPREFETCH prefetch
kernel/x86_64/izamax_sse2.S
+1 -1
View File
@@ -469,7 +469,7 @@
ALIGN_4
.L71:
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
prefetch PREFETCHSIZE * SIZE(X)
#endif
kernel/x86_64/scal_sse.S
+1 -1
View File
@@ -266,7 +266,7 @@
sarq $5, I
jle .L113
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
movaps %xmm0, %xmm1
mulps -32 * SIZE(X), %xmm1
kernel/x86_64/scal_sse2.S
+1 -1
View File
@@ -251,7 +251,7 @@
sarq $4, I
jle .L113
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
movaps %xmm0, %xmm1
mulpd -16 * SIZE(X), %xmm1
kernel/x86_64/symv_L_sse.S
+1 -1
View File
@@ -76,7 +76,7 @@
#define movsd movlps
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
kernel/x86_64/symv_L_sse2.S
+1 -1
View File
@@ -76,7 +76,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
kernel/x86_64/symv_U_sse.S
+1 -1
View File
@@ -76,7 +76,7 @@
#define movsd movlps
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
kernel/x86_64/symv_U_sse2.S
+1 -1
View File
@@ -76,7 +76,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
kernel/x86_64/trsm_kernel_LN_8x4_sse.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define movsd movlps
kernel/x86_64/trsm_kernel_LT_8x4_sse.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define movsd movlps
kernel/x86_64/trsm_kernel_RT_8x4_sse.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(OPTERON) || defined(BARCELONA)
#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define movsd movlps
kernel/x86_64/zgemm_ncopy_2.S
+1 -1
View File
@@ -85,7 +85,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define RPREFETCHSIZE 32
#define WPREFETCHSIZE 48
#endif
kernel/x86_64/zsymv_L_sse.S
+1 -1
View File
@@ -160,7 +160,7 @@
#define a3 %xmm14
#define xt1 %xmm15
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define MOVDDUP(a, b, c) movddup a(b), c
#define MOVDDUP2(a, b, c) movddup a##b, c
#else
kernel/x86_64/zsymv_L_sse2.S
+2 -2
View File
@@ -76,7 +76,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
@@ -167,7 +167,7 @@
#define a3 %xmm14
#define xt1 %xmm15
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI)
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define MOVDDUP(a, b, c) movddup a(b), c
#define MOVDDUP2(a, b, c) movddup a##b, c
#else
kernel/x86_64/zsymv_U_sse.S
+2 -2
View File
@@ -76,7 +76,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
@@ -166,7 +166,7 @@
#define xt1 %xmm14
#define xt2 %xmm15
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI)
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define MOVDDUP(a, b, c) movddup a(b), c
#define MOVDDUP2(a, b, c) movddup a##b, c
#else
kernel/x86_64/zsymv_U_sse2.S
+2 -2
View File
@@ -76,7 +76,7 @@
#define movsd movlpd
#endif
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHSIZE (16 * 16)
@@ -166,7 +166,7 @@
#define a3 %xmm14
#define xt1 %xmm15
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI)
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BULLDOZER)
#define MOVDDUP(a, b, c) movddup a(b), c
#define MOVDDUP2(a, b, c) movddup a##b, c
#else
kernel/x86_64/ztrsm_kernel_LN_2x2_sse2.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define BORIG 72(%rsp)
#define BUFFER 128(%rsp)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta
kernel/x86_64/ztrsm_kernel_LN_4x2_sse.S
+1 -1
View File
@@ -95,7 +95,7 @@
#define PREFETCHSIZE (8 * 6 + 4)
#endif
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta
kernel/x86_64/ztrsm_kernel_LT_2x2_sse2.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define BORIG 72(%rsp)
#define BUFFER 128(%rsp)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta
kernel/x86_64/ztrsm_kernel_LT_4x2_sse.S
+1 -1
View File
@@ -95,7 +95,7 @@
#define PREFETCHSIZE (8 * 6 + 4)
#endif
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta
kernel/x86_64/ztrsm_kernel_RT_2x2_sse2.S
+1 -1
View File
@@ -86,7 +86,7 @@
#define BORIG 72(%rsp)
#define BUFFER 128(%rsp)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta
kernel/x86_64/ztrsm_kernel_RT_4x2_sse.S
+1 -1
View File
@@ -95,7 +95,7 @@
#define PREFETCHSIZE (8 * 6 + 4)
#endif
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
#if defined(OPTERON) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
#define PREFETCH prefetch
#define PREFETCHW prefetchw
#define PREFETCHNTA prefetchnta