Refs #113. Fixed the typo BOBCATE -> BOBCAT
This commit is contained in:
parent
d6cab3f37e
commit
d3b67d0bd8
|
@ -28,7 +28,7 @@ OPTERON_SSE3
|
||||||
BARCELONA
|
BARCELONA
|
||||||
SHANGHAI
|
SHANGHAI
|
||||||
ISTANBUL
|
ISTANBUL
|
||||||
BOBCATE
|
BOBCAT
|
||||||
|
|
||||||
c)VIA CPU:
|
c)VIA CPU:
|
||||||
SSE_GENERIC
|
SSE_GENERIC
|
||||||
|
|
4
cpuid.h
4
cpuid.h
|
@ -104,7 +104,7 @@
|
||||||
#define CORE_ATOM 18
|
#define CORE_ATOM 18
|
||||||
#define CORE_NANO 19
|
#define CORE_NANO 19
|
||||||
#define CORE_SANDYBRIDGE 20
|
#define CORE_SANDYBRIDGE 20
|
||||||
#define CORE_BOBCATE 21
|
#define CORE_BOBCAT 21
|
||||||
|
|
||||||
#define HAVE_SSE (1 << 0)
|
#define HAVE_SSE (1 << 0)
|
||||||
#define HAVE_SSE2 (1 << 1)
|
#define HAVE_SSE2 (1 << 1)
|
||||||
|
@ -192,5 +192,5 @@ typedef struct {
|
||||||
#define CPUTYPE_VIAC3 42
|
#define CPUTYPE_VIAC3 42
|
||||||
#define CPUTYPE_NANO 43
|
#define CPUTYPE_NANO 43
|
||||||
#define CPUTYPE_SANDYBRIDGE 44
|
#define CPUTYPE_SANDYBRIDGE 44
|
||||||
#define CPUTYPE_BOBCATE 45
|
#define CPUTYPE_BOBCAT 45
|
||||||
#endif
|
#endif
|
||||||
|
|
12
cpuid_x86.c
12
cpuid_x86.c
|
@ -1029,7 +1029,7 @@ int get_cpuname(void){
|
||||||
case 10:
|
case 10:
|
||||||
return CPUTYPE_BARCELONA;
|
return CPUTYPE_BARCELONA;
|
||||||
case 5:
|
case 5:
|
||||||
return CPUTYPE_BOBCATE;
|
return CPUTYPE_BOBCAT;
|
||||||
}
|
}
|
||||||
break;
|
break;
|
||||||
}
|
}
|
||||||
|
@ -1150,7 +1150,7 @@ static char *cpuname[] = {
|
||||||
"VIAC3",
|
"VIAC3",
|
||||||
"NANO",
|
"NANO",
|
||||||
"SANDYBRIDGE",
|
"SANDYBRIDGE",
|
||||||
"BOBCATE",
|
"BOBCAT",
|
||||||
};
|
};
|
||||||
|
|
||||||
static char *lowercpuname[] = {
|
static char *lowercpuname[] = {
|
||||||
|
@ -1198,7 +1198,7 @@ static char *lowercpuname[] = {
|
||||||
"nsgeode",
|
"nsgeode",
|
||||||
"nano",
|
"nano",
|
||||||
"sandybridge",
|
"sandybridge",
|
||||||
"bobcate",
|
"bobcat",
|
||||||
};
|
};
|
||||||
|
|
||||||
static char *corename[] = {
|
static char *corename[] = {
|
||||||
|
@ -1223,7 +1223,7 @@ static char *corename[] = {
|
||||||
"ATOM",
|
"ATOM",
|
||||||
"NANO",
|
"NANO",
|
||||||
"SANDYBRIDGE",
|
"SANDYBRIDGE",
|
||||||
"BOBCATE",
|
"BOBCAT",
|
||||||
};
|
};
|
||||||
|
|
||||||
static char *corename_lower[] = {
|
static char *corename_lower[] = {
|
||||||
|
@ -1248,7 +1248,7 @@ static char *corename_lower[] = {
|
||||||
"atom",
|
"atom",
|
||||||
"nano",
|
"nano",
|
||||||
"sandybridge",
|
"sandybridge",
|
||||||
"bobcate",
|
"bobcat",
|
||||||
};
|
};
|
||||||
|
|
||||||
|
|
||||||
|
@ -1358,7 +1358,7 @@ int get_coretype(void){
|
||||||
if (family <= 0xe) return CORE_ATHLON;
|
if (family <= 0xe) return CORE_ATHLON;
|
||||||
if (family == 0xf){
|
if (family == 0xf){
|
||||||
if ((exfamily == 0) || (exfamily == 2)) return CORE_OPTERON;
|
if ((exfamily == 0) || (exfamily == 2)) return CORE_OPTERON;
|
||||||
else if (exfamily == 5) return CORE_BOBCATE;
|
else if (exfamily == 5) return CORE_BOBCAT;
|
||||||
else return CORE_BARCELONA;
|
else return CORE_BARCELONA;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
|
@ -163,7 +163,7 @@ int get_L2_size(void){
|
||||||
|
|
||||||
int eax, ebx, ecx, edx;
|
int eax, ebx, ecx, edx;
|
||||||
|
|
||||||
#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE) || \
|
#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || \
|
||||||
defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
|
defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
|
||||||
defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC)
|
defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC)
|
||||||
|
|
||||||
|
@ -446,7 +446,7 @@ void blas_set_parameter(void){
|
||||||
#endif
|
#endif
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(CORE_BARCELONA) || defined(CORE_BOBCATE)
|
#if defined(CORE_BARCELONA) || defined(CORE_BOBCAT)
|
||||||
size >>= 8;
|
size >>= 8;
|
||||||
|
|
||||||
sgemm_p = 232 * size;
|
sgemm_p = 232 * size;
|
||||||
|
|
12
getarch.c
12
getarch.c
|
@ -102,7 +102,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
/* #define FORCE_BARCELONA */
|
/* #define FORCE_BARCELONA */
|
||||||
/* #define FORCE_SHANGHAI */
|
/* #define FORCE_SHANGHAI */
|
||||||
/* #define FORCE_ISTANBUL */
|
/* #define FORCE_ISTANBUL */
|
||||||
/* #define FORCE_BOBCATE */
|
/* #define FORCE_BOBCAT */
|
||||||
/* #define FORCE_SSE_GENERIC */
|
/* #define FORCE_SSE_GENERIC */
|
||||||
/* #define FORCE_VIAC3 */
|
/* #define FORCE_VIAC3 */
|
||||||
/* #define FORCE_NANO */
|
/* #define FORCE_NANO */
|
||||||
|
@ -364,19 +364,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
#define CORENAME "BARCELONA"
|
#define CORENAME "BARCELONA"
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(FORCE_BOBCATE)
|
#if defined(FORCE_BOBCAT)
|
||||||
#define FORCE
|
#define FORCE
|
||||||
#define FORCE_INTEL
|
#define FORCE_INTEL
|
||||||
#define ARCHITECTURE "X86"
|
#define ARCHITECTURE "X86"
|
||||||
#define SUBARCHITECTURE "BOBCATE"
|
#define SUBARCHITECTURE "BOBCAT"
|
||||||
#define ARCHCONFIG "-DBOBCATE " \
|
#define ARCHCONFIG "-DBOBCAT " \
|
||||||
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
|
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
|
||||||
"-DL2_SIZE=524288 -DL2_LINESIZE=64 " \
|
"-DL2_SIZE=524288 -DL2_LINESIZE=64 " \
|
||||||
"-DDTB_DEFAULT_ENTRIES=40 -DDTB_SIZE=4096 " \
|
"-DDTB_DEFAULT_ENTRIES=40 -DDTB_SIZE=4096 " \
|
||||||
"-DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 " \
|
"-DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 " \
|
||||||
"-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_CFLUSH -DHAVE_CMOV"
|
"-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_CFLUSH -DHAVE_CMOV"
|
||||||
#define LIBNAME "bobcate"
|
#define LIBNAME "bobcat"
|
||||||
#define CORENAME "BOBCATE"
|
#define CORENAME "BOBCAT"
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#ifdef FORCE_SSE_GENERIC
|
#ifdef FORCE_SSE_GENERIC
|
||||||
|
|
|
@ -794,7 +794,7 @@ static void init_parameter(void) {
|
||||||
#endif
|
#endif
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#ifdef BOBCATE
|
#ifdef BOBCAT
|
||||||
|
|
||||||
#ifdef DEBUG
|
#ifdef DEBUG
|
||||||
fprintf(stderr, "Bobcate\n");
|
fprintf(stderr, "Bobcate\n");
|
||||||
|
|
|
@ -69,7 +69,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHSIZE (8 * 10 + 4)
|
#define PREFETCHSIZE (8 * 10 + 4)
|
||||||
#endif
|
#endif
|
||||||
|
@ -439,7 +439,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulsd %xmm0, %xmm2
|
mulsd %xmm0, %xmm2
|
||||||
addsd %xmm2, %xmm4
|
addsd %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movlpd 2 * SIZE(BB), %xmm2
|
movlpd 2 * SIZE(BB), %xmm2
|
||||||
|
@ -488,7 +488,7 @@
|
||||||
movlpd 40 * SIZE(BB), %xmm3
|
movlpd 40 * SIZE(BB), %xmm3
|
||||||
addsd %xmm0, %xmm7
|
addsd %xmm0, %xmm7
|
||||||
movlpd 8 * SIZE(AA), %xmm0
|
movlpd 8 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulsd %xmm1, %xmm2
|
mulsd %xmm1, %xmm2
|
||||||
|
@ -1697,7 +1697,7 @@
|
||||||
|
|
||||||
.L42:
|
.L42:
|
||||||
mulpd %xmm0, %xmm2
|
mulpd %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd 2 * SIZE(BB), %xmm0
|
mulpd 2 * SIZE(BB), %xmm0
|
||||||
|
@ -1727,7 +1727,7 @@
|
||||||
addpd %xmm0, %xmm7
|
addpd %xmm0, %xmm7
|
||||||
movapd 16 * SIZE(AA), %xmm0
|
movapd 16 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd %xmm1, %xmm2
|
mulpd %xmm1, %xmm2
|
||||||
|
|
|
@ -64,7 +64,7 @@
|
||||||
#define BORIG 60(%esp)
|
#define BORIG 60(%esp)
|
||||||
#define BUFFER 128(%esp)
|
#define BUFFER 128(%esp)
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
|
@ -437,7 +437,7 @@
|
||||||
.L32:
|
.L32:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
addss %xmm2, %xmm4
|
addss %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 4 * SIZE(BB), %xmm2
|
movss 4 * SIZE(BB), %xmm2
|
||||||
|
@ -833,7 +833,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(BB), %xmm2
|
movaps 4 * SIZE(BB), %xmm2
|
||||||
|
@ -1848,7 +1848,7 @@
|
||||||
|
|
||||||
.L72:
|
.L72:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulss 4 * SIZE(BB), %xmm0
|
mulss 4 * SIZE(BB), %xmm0
|
||||||
|
@ -2109,7 +2109,7 @@
|
||||||
ALIGN_4
|
ALIGN_4
|
||||||
|
|
||||||
.L62:
|
.L62:
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
@ -2429,7 +2429,7 @@
|
||||||
|
|
||||||
.L52:
|
.L52:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps 4 * SIZE(BB), %xmm0
|
mulps 4 * SIZE(BB), %xmm0
|
||||||
|
@ -2459,7 +2459,7 @@
|
||||||
addps %xmm0, %xmm5
|
addps %xmm0, %xmm5
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
|
@ -2952,7 +2952,7 @@
|
||||||
|
|
||||||
.L112:
|
.L112:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 1 * SIZE(AA), %xmm0
|
movss 1 * SIZE(AA), %xmm0
|
||||||
|
@ -3148,7 +3148,7 @@
|
||||||
|
|
||||||
.L102:
|
.L102:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movsd 2 * SIZE(AA), %xmm0
|
movsd 2 * SIZE(AA), %xmm0
|
||||||
|
@ -3389,7 +3389,7 @@
|
||||||
|
|
||||||
.L92:
|
.L92:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(AA), %xmm0
|
movaps 4 * SIZE(AA), %xmm0
|
||||||
|
@ -3404,7 +3404,7 @@
|
||||||
mulps 12 * SIZE(BB), %xmm0
|
mulps 12 * SIZE(BB), %xmm0
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm3
|
mulps %xmm1, %xmm3
|
||||||
|
|
|
@ -69,7 +69,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHSIZE (8 * 10 + 4)
|
#define PREFETCHSIZE (8 * 10 + 4)
|
||||||
#endif
|
#endif
|
||||||
|
@ -910,7 +910,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulsd %xmm0, %xmm2
|
mulsd %xmm0, %xmm2
|
||||||
addsd %xmm2, %xmm4
|
addsd %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movlpd 2 * SIZE(BB), %xmm2
|
movlpd 2 * SIZE(BB), %xmm2
|
||||||
|
@ -959,7 +959,7 @@
|
||||||
movlpd 40 * SIZE(BB), %xmm3
|
movlpd 40 * SIZE(BB), %xmm3
|
||||||
addsd %xmm0, %xmm7
|
addsd %xmm0, %xmm7
|
||||||
movlpd 8 * SIZE(AA), %xmm0
|
movlpd 8 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulsd %xmm1, %xmm2
|
mulsd %xmm1, %xmm2
|
||||||
|
@ -1439,7 +1439,7 @@
|
||||||
|
|
||||||
.L42:
|
.L42:
|
||||||
mulpd %xmm0, %xmm2
|
mulpd %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd 2 * SIZE(BB), %xmm0
|
mulpd 2 * SIZE(BB), %xmm0
|
||||||
|
@ -1469,7 +1469,7 @@
|
||||||
addpd %xmm0, %xmm7
|
addpd %xmm0, %xmm7
|
||||||
movapd 16 * SIZE(AA), %xmm0
|
movapd 16 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd %xmm1, %xmm2
|
mulpd %xmm1, %xmm2
|
||||||
|
|
|
@ -64,7 +64,7 @@
|
||||||
#define BORIG 60(%esp)
|
#define BORIG 60(%esp)
|
||||||
#define BUFFER 128(%esp)
|
#define BUFFER 128(%esp)
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
|
@ -872,7 +872,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(BB), %xmm2
|
movaps 4 * SIZE(BB), %xmm2
|
||||||
|
@ -1316,7 +1316,7 @@
|
||||||
.L32:
|
.L32:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
addss %xmm2, %xmm4
|
addss %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 4 * SIZE(BB), %xmm2
|
movss 4 * SIZE(BB), %xmm2
|
||||||
|
@ -1855,7 +1855,7 @@
|
||||||
|
|
||||||
.L52:
|
.L52:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps 4 * SIZE(BB), %xmm0
|
mulps 4 * SIZE(BB), %xmm0
|
||||||
|
@ -1885,7 +1885,7 @@
|
||||||
addps %xmm0, %xmm5
|
addps %xmm0, %xmm5
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
|
@ -2249,7 +2249,7 @@
|
||||||
ALIGN_4
|
ALIGN_4
|
||||||
|
|
||||||
.L62:
|
.L62:
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
@ -2562,7 +2562,7 @@
|
||||||
|
|
||||||
.L72:
|
.L72:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulss 4 * SIZE(BB), %xmm0
|
mulss 4 * SIZE(BB), %xmm0
|
||||||
|
@ -2957,7 +2957,7 @@
|
||||||
|
|
||||||
.L92:
|
.L92:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(AA), %xmm0
|
movaps 4 * SIZE(AA), %xmm0
|
||||||
|
@ -2972,7 +2972,7 @@
|
||||||
mulps 12 * SIZE(BB), %xmm0
|
mulps 12 * SIZE(BB), %xmm0
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm3
|
mulps %xmm1, %xmm3
|
||||||
|
@ -3280,7 +3280,7 @@
|
||||||
|
|
||||||
.L102:
|
.L102:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movsd 2 * SIZE(AA), %xmm0
|
movsd 2 * SIZE(AA), %xmm0
|
||||||
|
@ -3515,7 +3515,7 @@
|
||||||
|
|
||||||
.L112:
|
.L112:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 1 * SIZE(AA), %xmm0
|
movss 1 * SIZE(AA), %xmm0
|
||||||
|
|
|
@ -69,7 +69,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHSIZE (8 * 10 + 4)
|
#define PREFETCHSIZE (8 * 10 + 4)
|
||||||
#endif
|
#endif
|
||||||
|
@ -1036,7 +1036,7 @@
|
||||||
|
|
||||||
.L42:
|
.L42:
|
||||||
mulpd %xmm0, %xmm2
|
mulpd %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd 2 * SIZE(BB), %xmm0
|
mulpd 2 * SIZE(BB), %xmm0
|
||||||
|
@ -1066,7 +1066,7 @@
|
||||||
addpd %xmm0, %xmm7
|
addpd %xmm0, %xmm7
|
||||||
movapd 16 * SIZE(AA), %xmm0
|
movapd 16 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulpd %xmm1, %xmm2
|
mulpd %xmm1, %xmm2
|
||||||
|
@ -2224,7 +2224,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulsd %xmm0, %xmm2
|
mulsd %xmm0, %xmm2
|
||||||
addsd %xmm2, %xmm4
|
addsd %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movlpd 2 * SIZE(BB), %xmm2
|
movlpd 2 * SIZE(BB), %xmm2
|
||||||
|
@ -2273,7 +2273,7 @@
|
||||||
movlpd 40 * SIZE(BB), %xmm3
|
movlpd 40 * SIZE(BB), %xmm3
|
||||||
addsd %xmm0, %xmm7
|
addsd %xmm0, %xmm7
|
||||||
movlpd 8 * SIZE(AA), %xmm0
|
movlpd 8 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
PREFETCH (PREFETCHSIZE + 8) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulsd %xmm1, %xmm2
|
mulsd %xmm1, %xmm2
|
||||||
|
|
|
@ -64,7 +64,7 @@
|
||||||
#define BORIG 60(%esp)
|
#define BORIG 60(%esp)
|
||||||
#define BUFFER 128(%esp)
|
#define BUFFER 128(%esp)
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
|
@ -439,7 +439,7 @@
|
||||||
|
|
||||||
.L92:
|
.L92:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(AA), %xmm0
|
movaps 4 * SIZE(AA), %xmm0
|
||||||
|
@ -454,7 +454,7 @@
|
||||||
mulps 12 * SIZE(BB), %xmm0
|
mulps 12 * SIZE(BB), %xmm0
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm3
|
mulps %xmm1, %xmm3
|
||||||
|
@ -758,7 +758,7 @@
|
||||||
|
|
||||||
.L102:
|
.L102:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movsd 2 * SIZE(AA), %xmm0
|
movsd 2 * SIZE(AA), %xmm0
|
||||||
|
@ -993,7 +993,7 @@
|
||||||
|
|
||||||
.L112:
|
.L112:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 1 * SIZE(AA), %xmm0
|
movss 1 * SIZE(AA), %xmm0
|
||||||
|
@ -1324,7 +1324,7 @@
|
||||||
|
|
||||||
.L52:
|
.L52:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps 4 * SIZE(BB), %xmm0
|
mulps 4 * SIZE(BB), %xmm0
|
||||||
|
@ -1354,7 +1354,7 @@
|
||||||
addps %xmm0, %xmm5
|
addps %xmm0, %xmm5
|
||||||
movaps 32 * SIZE(AA), %xmm0
|
movaps 32 * SIZE(AA), %xmm0
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
|
@ -1718,7 +1718,7 @@
|
||||||
ALIGN_4
|
ALIGN_4
|
||||||
|
|
||||||
.L62:
|
.L62:
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
@ -2031,7 +2031,7 @@
|
||||||
|
|
||||||
.L72:
|
.L72:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
mulss 4 * SIZE(BB), %xmm0
|
mulss 4 * SIZE(BB), %xmm0
|
||||||
|
@ -2859,7 +2859,7 @@
|
||||||
.L22:
|
.L22:
|
||||||
mulps %xmm0, %xmm2
|
mulps %xmm0, %xmm2
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movaps 4 * SIZE(BB), %xmm2
|
movaps 4 * SIZE(BB), %xmm2
|
||||||
|
@ -3303,7 +3303,7 @@
|
||||||
.L32:
|
.L32:
|
||||||
mulss %xmm0, %xmm2
|
mulss %xmm0, %xmm2
|
||||||
addss %xmm2, %xmm4
|
addss %xmm2, %xmm4
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
prefetcht0 (PREFETCHSIZE + 0) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
movss 4 * SIZE(BB), %xmm2
|
movss 4 * SIZE(BB), %xmm2
|
||||||
|
|
|
@ -75,7 +75,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
#define WPREFETCHSIZE 112
|
#define WPREFETCHSIZE 112
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
|
@ -533,7 +533,7 @@
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movsd 16 * SIZE(AA), %xmm0
|
movsd 16 * SIZE(AA), %xmm0
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
|
|
|
@ -75,7 +75,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
#define WPREFETCHSIZE 112
|
#define WPREFETCHSIZE 112
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
|
@ -994,7 +994,7 @@
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movsd 16 * SIZE(AA), %xmm0
|
movsd 16 * SIZE(AA), %xmm0
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
|
|
|
@ -75,7 +75,7 @@
|
||||||
#define STACK_ALIGN 4096
|
#define STACK_ALIGN 4096
|
||||||
#define STACK_OFFSET 1024
|
#define STACK_OFFSET 1024
|
||||||
|
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
#define PREFETCHSIZE (16 * 10 + 8)
|
#define PREFETCHSIZE (16 * 10 + 8)
|
||||||
#define WPREFETCHSIZE 112
|
#define WPREFETCHSIZE 112
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
|
@ -1820,7 +1820,7 @@
|
||||||
addps %xmm0, %xmm7
|
addps %xmm0, %xmm7
|
||||||
movsd 16 * SIZE(AA), %xmm0
|
movsd 16 * SIZE(AA), %xmm0
|
||||||
mulps %xmm1, %xmm2
|
mulps %xmm1, %xmm2
|
||||||
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCATE)
|
#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT)
|
||||||
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
prefetcht1 (PREFETCHSIZE + 16) * SIZE(AA)
|
||||||
#endif
|
#endif
|
||||||
addps %xmm2, %xmm4
|
addps %xmm2, %xmm4
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlps
|
#define movsd movlps
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlpd
|
#define movsd movlpd
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlps
|
#define movsd movlps
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlpd
|
#define movsd movlpd
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -160,7 +160,7 @@
|
||||||
#define a3 %xmm14
|
#define a3 %xmm14
|
||||||
#define xt1 %xmm15
|
#define xt1 %xmm15
|
||||||
|
|
||||||
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define MOVDDUP(a, b, c) movddup a(b), c
|
#define MOVDDUP(a, b, c) movddup a(b), c
|
||||||
#define MOVDDUP2(a, b, c) movddup a##b, c
|
#define MOVDDUP2(a, b, c) movddup a##b, c
|
||||||
#else
|
#else
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlpd
|
#define movsd movlpd
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlpd
|
#define movsd movlpd
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -76,7 +76,7 @@
|
||||||
#define movsd movlpd
|
#define movsd movlpd
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (16 * 16)
|
#define PREFETCHSIZE (16 * 16)
|
||||||
|
|
|
@ -67,7 +67,7 @@
|
||||||
#define ALIGNED_ACCESS
|
#define ALIGNED_ACCESS
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#ifdef BOBCATE
|
#ifdef BOBCAT
|
||||||
#define PREFETCH prefetch
|
#define PREFETCH prefetch
|
||||||
#define PREFETCHW prefetchw
|
#define PREFETCHW prefetchw
|
||||||
#define PREFETCHSIZE (128 * 5)
|
#define PREFETCHSIZE (128 * 5)
|
||||||
|
|
|
@ -85,7 +85,7 @@
|
||||||
#define movsd movlps
|
#define movsd movlps
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCATE)
|
#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
|
||||||
#define ALIGNED_ACCESS
|
#define ALIGNED_ACCESS
|
||||||
#define MOVUPS_A movaps
|
#define MOVUPS_A movaps
|
||||||
#define MOVUPS_XL movaps
|
#define MOVUPS_XL movaps
|
||||||
|
|
Loading…
Reference in New Issue