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

Merge branch 'sandybridge' into develop 

Just copy the kernel codes from Nehalem. The optimization is ongoing.
This commit is contained in:
Zhang Xianyi 2012-05-31 12:42:32 +08:00
parent a4daa34db7 8cc7f86cf7
commit 90d6ad569d
57 changed files with 311 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
Makefile.system
View File

@ -247,11 +247,11 @@ endif
ifdef DYNAMIC_ARCH
ifeq ($(ARCH), x86)
DYNAMIC_CORE = KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS \
CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA ATOM NANO
CORE2 PENRYN DUNNINGTON NEHALEM SANDYBRIDGE ATHLON OPTERON OPTERON_SSE3 BARCELONA ATOM NANO
endif
ifeq ($(ARCH), x86_64)
DYNAMIC_CORE = PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA ATOM NANO
DYNAMIC_CORE = PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM SANDYBRIDGE OPTERON OPTERON_SSE3 BARCELONA ATOM NANO
endif
ifndef DYNAMIC_CORE
@ -770,6 +770,7 @@ export HAVE_SSE4_1
export HAVE_SSE4_2
export HAVE_SSE4A
export HAVE_SSE5
export HAVE_AVX
export KERNELDIR
export FUNCTION_PROFILE
export TARGET_CORE

2
TargetList.txt
View File

@ -18,6 +18,7 @@ CORE2
PENRYN
DUNNINGTON
NEHALEM
SANDYBRIDGE
ATOM
b)AMD CPU:
@ -47,6 +48,7 @@ CELL
3.MIPS64 CPU:
SICORTEX
LOONGSON3A
LOONGSON3B
4.IA64 CPU:
ITANIUM2

3
cpuid.h
View File

@ -103,6 +103,7 @@
#define CORE_NEHALEM 17
#define CORE_ATOM 18
#define CORE_NANO 19
#define CORE_SANDYBRIDGE 20
#define HAVE_SSE (1 << 0)
#define HAVE_SSE2 (1 << 1)
@ -122,6 +123,7 @@
#define HAVE_MISALIGNSSE (1 << 15)
#define HAVE_128BITFPU (1 << 16)
#define HAVE_FASTMOVU (1 << 17)
#define HAVE_AVX (1 << 18)
#define CACHE_INFO_L1_I 1
#define CACHE_INFO_L1_D 2
@ -188,4 +190,5 @@ typedef struct {
#define CPUTYPE_NSGEODE 41
#define CPUTYPE_VIAC3 42
#define CPUTYPE_NANO 43
#define CPUTYPE_SANDYBRIDGE 44
#endif

15
cpuid_x86.c
View File

@ -189,6 +189,7 @@ int get_cputype(int gettype){
if ((ecx & (1 << 9)) != 0) feature |= HAVE_SSSE3;
if ((ecx & (1 << 19)) != 0) feature |= HAVE_SSE4_1;
if ((ecx & (1 << 20)) != 0) feature |= HAVE_SSE4_2;
if ((ecx & (1 << 28)) != 0) feature |= HAVE_AVX;
if (have_excpuid() >= 0x01) {
cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
@ -983,13 +984,13 @@ int get_cpuname(void){
return CPUTYPE_NEHALEM;
case 10:
//Intel Core i5-2000 /i7-2000 (Sandy Bridge)
return CPUTYPE_NEHALEM;
return CPUTYPE_SANDYBRIDGE;
case 12:
//Xeon Processor 5600 (Westmere-EP)
return CPUTYPE_NEHALEM;
case 13:
//Intel Core i7-3000 / Xeon E5 (Sandy Bridge)
return CPUTYPE_NEHALEM;
return CPUTYPE_SANDYBRIDGE;
case 15:
//Xeon Processor E7 (Westmere-EX)
return CPUTYPE_NEHALEM;
@ -1146,6 +1147,7 @@ static char *cpuname[] = {
"NSGEODE",
"VIAC3",
"NANO",
"SANDYBRIDGE",
};
static char *lowercpuname[] = {
@ -1192,6 +1194,7 @@ static char *lowercpuname[] = {
"tms3x00",
"nsgeode",
"nano",
"sandybridge",
};
static char *corename[] = {
@ -1215,6 +1218,7 @@ static char *corename[] = {
"NEHALEM",
"ATOM",
"NANO",
"SANDYBRIDGE",
};
static char *corename_lower[] = {
@ -1238,6 +1242,7 @@ static char *corename_lower[] = {
"nehalem",
"atom",
"nano",
"sandybridge",
};
@ -1321,13 +1326,13 @@ int get_coretype(void){
return CORE_NEHALEM;
case 10:
//Intel Core i5-2000 /i7-2000 (Sandy Bridge)
return CORE_NEHALEM;
return CORE_SANDYBRIDGE;
case 12:
//Xeon Processor 5600 (Westmere-EP)
return CORE_NEHALEM;
case 13:
//Intel Core i7-3000 / Xeon E5 (Sandy Bridge)
return CORE_NEHALEM;
return CORE_SANDYBRIDGE;
case 15:
//Xeon Processor E7 (Westmere-EX)
return CORE_NEHALEM;
@ -1426,6 +1431,7 @@ void get_cpuconfig(void){
if (features & HAVE_SSE4_2) printf("#define HAVE_SSE4_2\n");
if (features & HAVE_SSE4A) printf("#define HAVE_SSE4A\n");
if (features & HAVE_SSE5 ) printf("#define HAVE_SSSE5\n");
if (features & HAVE_AVX ) printf("#define HAVE_AVX\n");
if (features & HAVE_3DNOWEX) printf("#define HAVE_3DNOWEX\n");
if (features & HAVE_3DNOW) printf("#define HAVE_3DNOW\n");
if (features & HAVE_CFLUSH) printf("#define HAVE_CFLUSH\n");
@ -1491,6 +1497,7 @@ void get_sse(void){
if (features & HAVE_SSE4_2) printf("HAVE_SSE4_2=1\n");
if (features & HAVE_SSE4A) printf("HAVE_SSE4A=1\n");
if (features & HAVE_SSE5 ) printf("HAVE_SSSE5=1\n");
if (features & HAVE_AVX ) printf("HAVE_AVX=1\n");
if (features & HAVE_3DNOWEX) printf("HAVE_3DNOWEX=1\n");
if (features & HAVE_3DNOW) printf("HAVE_3DNOW=1\n");

13
driver/others/parameter.c
View File

@ -165,7 +165,7 @@ int get_L2_size(void){
#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || \
defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
defined(CORE_NEHALEM) || defined(ATOM) || defined(GENERIC)
defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC)
cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
@ -384,6 +384,17 @@ void blas_set_parameter(void){
#endif
#endif
#if defined(SANDYBRIDGE)
sgemm_p = 1024;
dgemm_p = 512;
cgemm_p = 512;
zgemm_p = 256;
#ifdef EXPRECISION
qgemm_p = 256;
xgemm_p = 128;
#endif
#endif
#if defined(CORE_PRESCOTT) || defined(GENERIC)
size >>= 6;

14
getarch.c
View File

@ -278,6 +278,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define CORENAME "NEHALEM"
#endif
#ifdef FORCE_SANDYBRIDGE
#define FORCE
#define FORCE_INTEL
#define ARCHITECTURE "X86"
#define SUBARCHITECTURE "SANDYBRIDGE"
#define ARCHCONFIG "-DSANDYBRIDGE " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2 -DHAVE_AVX"
#define LIBNAME "sandybridge"
#define CORENAME "SANDYBRIDGE"
#endif
#ifdef FORCE_ATOM
#define FORCE
#define FORCE_INTEL

16
kernel/setparam-ref.c
View File

@ -746,6 +746,22 @@ static void init_parameter(void) {
#endif
#endif
#ifdef SANDYBRIDGE
#ifdef DEBUG
fprintf(stderr, "Sandybridge\n");
#endif
TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
#ifdef EXPRECISION
TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
#endif
#endif
#ifdef OPTERON
#ifdef DEBUG

1
kernel/x86/KERNEL.SANDYBRIDGE Normal file
View File

@ -0,0 +1 @@
include $(KERNELDIR)/KERNEL.PENRYN

6
kernel/x86/gemm_kernel_2x4_penryn.S
View File

@ -76,6 +76,12 @@
#define PREFETCHB prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE (8 * 1 - 4)
#define PREFETCHW prefetcht0
#define PREFETCHB prefetcht0
#endif
#ifndef PREFETCH
#define PREFETCH prefetcht0
#endif

8
kernel/x86/gemm_kernel_4x4_penryn.S
View File

@ -69,6 +69,12 @@
#define PREFETCHB prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE (16 * 1 - 8)
#define PREFETCHW prefetcht0
#define PREFETCHB prefetcht0
#endif
#ifndef PREFETCH
#define PREFETCH prefetcht0
#endif
@ -262,7 +268,7 @@
movaps -16 * SIZE(AA), %xmm0
addps %xmm2, %xmm7
#ifndef NEHALEM
#if !(defined(NEHALEM) || defined(SANDYBRIDGE))
PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)
#endif
pshufd $0x93, %xmm1, %xmm2

2
kernel/x86/gemv_n_sse.S
View File

@ -58,7 +58,7 @@
#define PREFETCHSIZE (16 * 4)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 7)

2
kernel/x86/gemv_n_sse2.S
View File

@ -45,7 +45,7 @@
#define PREFETCHSIZE (8 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (8 * 7)

2
kernel/x86/gemv_t_sse.S
View File

@ -58,7 +58,7 @@
#define PREFETCHSIZE (16 * 4)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 7)

2
kernel/x86/gemv_t_sse2.S
View File

@ -45,7 +45,7 @@
#define PREFETCHSIZE (8 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (8 * 7)

2
kernel/x86/trsm_kernel_LN_2x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/trsm_kernel_LN_4x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/trsm_kernel_LT_2x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/trsm_kernel_LT_4x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/trsm_kernel_RT_2x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/trsm_kernel_RT_4x4_penryn.S
View File

@ -62,7 +62,7 @@
#define PREFETCHSIZE (8 * 21 + 4)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHSIZE (8 * 21 + 4)
#endif

2
kernel/x86/zgemm_kernel_1x2_penryn.S
View File

@ -64,7 +64,7 @@
#define PREFETCHB prefetcht0
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCHSIZE (8 * 1 - 4)
#define PREFETCHW prefetcht0
#define PREFETCHB prefetcht0

2
kernel/x86/zgemm_kernel_2x2_penryn.S
View File

@ -64,7 +64,7 @@
#define PREFETCHB prefetcht0
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCHSIZE (16 * 1 + 8)
#define PREFETCHW prefetcht0
#define PREFETCHB prefetcht0

2
kernel/x86/zgemv_n_sse.S
View File

@ -58,7 +58,7 @@
#define PREFETCHSIZE (16 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 7)

2
kernel/x86/zgemv_n_sse2.S
View File

@ -45,7 +45,7 @@
#define PREFETCHSIZE (8 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (8 * 7)

2
kernel/x86/zgemv_t_sse.S
View File

@ -58,7 +58,7 @@
#define PREFETCHSIZE (16 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 7)

2
kernel/x86/zgemv_t_sse2.S
View File

@ -45,7 +45,7 @@
#define PREFETCHSIZE (8 * 2)
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (8 * 7)

4
kernel/x86/zscal_sse.S
View File

@ -55,7 +55,7 @@
#define XX %edi
#define FLAG %ebp
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON)
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(SANDYBRIDGE)
#define USE_PSHUFD
#else
#define USE_PSHUFD_HALF
@ -697,7 +697,7 @@
cmpl $2 * SIZE, INCX
jne .L120
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
PSHUFD2($0, %xmm0, %xmm6)
PSHUFD2($0, %xmm1, %xmm1)

4
kernel/x86/zscal_sse2.S
View File

@ -57,7 +57,7 @@
#include "l1param.h"
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON)
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(SANDYBRIDGE)
#define USE_PSHUFD
#else
#define USE_PSHUFD_HALF
@ -860,7 +860,7 @@
cmpl $2 * SIZE, INCX
jne .L220
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
#ifdef HAVE_SSE3
movddup %xmm0, %xmm6

2
kernel/x86/ztrsm_kernel_LN_2x2_penryn.S
View File

@ -61,7 +61,7 @@
#define PREFETCHSIZE 84
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht1
#define PREFETCHSIZE 84
#endif

2
kernel/x86/ztrsm_kernel_LT_1x2_penryn.S
View File

@ -63,7 +63,7 @@
#define PREFETCHSIZE 84
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht1
#define PREFETCHSIZE 84
#endif

2
kernel/x86/ztrsm_kernel_LT_2x2_penryn.S
View File

@ -61,7 +61,7 @@
#define PREFETCHSIZE 84
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht1
#define PREFETCHSIZE 84
#endif

2
kernel/x86/ztrsm_kernel_RT_1x2_penryn.S
View File

@ -63,7 +63,7 @@
#define PREFETCHSIZE 84
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht1
#define PREFETCHSIZE 84
#endif

2
kernel/x86/ztrsm_kernel_RT_2x2_penryn.S
View File

@ -61,7 +61,7 @@
#define PREFETCHSIZE 84
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht1
#define PREFETCHSIZE 84
#endif

59
kernel/x86_64/KERNEL.SANDYBRIDGE Normal file
View File

@ -0,0 +1,59 @@
SGEMMKERNEL = gemm_kernel_4x8_nehalem.S
SGEMMINCOPY = gemm_ncopy_4.S
SGEMMITCOPY = gemm_tcopy_4.S
SGEMMONCOPY = ../generic/gemm_ncopy_8.c
SGEMMOTCOPY = ../generic/gemm_tcopy_8.c
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
DGEMMKERNEL = gemm_kernel_2x8_nehalem.S
DGEMMINCOPY = dgemm_ncopy_2.S
DGEMMITCOPY = dgemm_tcopy_2.S
DGEMMONCOPY = dgemm_ncopy_8.S
DGEMMOTCOPY = dgemm_tcopy_8.S
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
CGEMMKERNEL = zgemm_kernel_2x4_nehalem.S
CGEMMINCOPY = zgemm_ncopy_2.S
CGEMMITCOPY = zgemm_tcopy_2.S
CGEMMONCOPY = ../generic/zgemm_ncopy_4.c
CGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX)
CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX)
CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX)
CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX)
ZGEMMKERNEL = zgemm_kernel_1x4_nehalem.S
ZGEMMINCOPY = zgemm_ncopy_1.S
ZGEMMITCOPY = zgemm_tcopy_1.S
ZGEMMONCOPY = ../generic/zgemm_ncopy_4.c
ZGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX)
ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX)
ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
STRSMKERNEL_LN = trsm_kernel_LN_4x8_nehalem.S
STRSMKERNEL_LT = trsm_kernel_LT_4x8_nehalem.S
STRSMKERNEL_RN = trsm_kernel_LT_4x8_nehalem.S
STRSMKERNEL_RT = trsm_kernel_RT_4x8_nehalem.S
DTRSMKERNEL_LN = trsm_kernel_LN_2x8_nehalem.S
DTRSMKERNEL_LT = trsm_kernel_LT_2x8_nehalem.S
DTRSMKERNEL_RN = trsm_kernel_LT_2x8_nehalem.S
DTRSMKERNEL_RT = trsm_kernel_RT_2x8_nehalem.S
CTRSMKERNEL_LN = ztrsm_kernel_LN_2x4_nehalem.S
CTRSMKERNEL_LT = ztrsm_kernel_LT_2x4_nehalem.S
CTRSMKERNEL_RN = ztrsm_kernel_LT_2x4_nehalem.S
CTRSMKERNEL_RT = ztrsm_kernel_RT_2x4_nehalem.S
ZTRSMKERNEL_LN = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_LT = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_RN = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_RT = ztrsm_kernel_RT_1x4_nehalem.S
CGEMM3MKERNEL = zgemm3m_kernel_4x8_nehalem.S
ZGEMM3MKERNEL = zgemm3m_kernel_2x8_nehalem.S

6
kernel/x86_64/dgemm_ncopy_2.S
View File

@ -45,6 +45,12 @@
#define PREFETCHW prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef MOVAPS
#define MOVAPS movaps
#endif

2
kernel/x86_64/dgemm_ncopy_4.S
View File

@ -45,7 +45,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0

6
kernel/x86_64/dgemm_ncopy_8.S
View File

@ -45,6 +45,12 @@
#define PREFETCHW prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef MOVAPS
#define MOVAPS movaps
#endif

7
kernel/x86_64/dgemm_tcopy_2.S
View File

@ -52,6 +52,13 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define MOVUPS_A movups
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0

6
kernel/x86_64/dgemm_tcopy_4.S
View File

@ -51,6 +51,12 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define MOVUPS_A movups
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0

7
kernel/x86_64/dgemm_tcopy_8.S
View File

@ -46,6 +46,13 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define MOVUPS_A movups
#endif
#ifdef MOVUPS_A
#define MOVUPS_A1(OFF, ADDR, REGS) MOVUPS_A OFF(ADDR), REGS
#define MOVUPS_A2(OFF, ADDR, BASE, SCALE, REGS) MOVUPS_A OFF(ADDR, BASE, SCALE), REGS

7
kernel/x86_64/gemm_ncopy_2.S
View File

@ -46,6 +46,13 @@
#define PREFETCHW prefetcht0
#endif
#if defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 2)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef WINDOWS_ABI
#define M ARG1 /* rdi */

2
kernel/x86_64/gemm_ncopy_4.S
View File

@ -46,7 +46,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 4)
#define PREFETCH prefetcht0

7
kernel/x86_64/gemm_tcopy_2.S
View File

@ -46,6 +46,13 @@
#define PREFETCHW prefetcht0
#endif
#if defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 2)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef WINDOWS_ABI
#define M ARG1 /* rdi */

2
kernel/x86_64/gemm_tcopy_4.S
View File

@ -46,7 +46,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 4)
#define PREFETCH prefetcht0

2
kernel/x86_64/symv_L_sse.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)

2
kernel/x86_64/symv_L_sse2.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)

2
kernel/x86_64/symv_U_sse.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)

2
kernel/x86_64/symv_U_sse2.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)

2
kernel/x86_64/zscal_sse.S
View File

@ -685,7 +685,7 @@
cmpq $2 * SIZE, INCX
jne .L120
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
pshufd $0, %xmm0, %xmm14
pshufd $0, %xmm1, %xmm1

4
kernel/x86_64/zscal_sse2.S
View File

@ -55,7 +55,7 @@
#include "l1param.h"
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(BARCELONA) || defined(NANO)
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(BARCELONA) || defined(NANO) || defined(SANDYBRIDGE)
#define USE_PSHUFD
#else
#define USE_PSHUFD_HALF
@ -803,7 +803,7 @@
cmpq $2 * SIZE, INCX
jne .L220
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
movddup %xmm0, %xmm14
pxor %xmm15, %xmm15

2
kernel/x86_64/zsymv_L_sse.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)

2
kernel/x86_64/zsymv_L_sse2.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)

2
kernel/x86_64/zsymv_U_sse.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)

2
kernel/x86_64/zsymv_U_sse2.S
View File

@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)

7
l1param.h
View File

@ -9,6 +9,13 @@
#define ALIGNED_ACCESS
#endif
#ifdef SANDYBRIDGE
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (128 * 12)
#define ALIGNED_ACCESS
#endif
#ifdef ATHLON
#define PREFETCH prefetch
#define PREFETCHW prefetchw

11
l2param.h
View File

@ -63,6 +63,17 @@
#define PREFETCHSIZE 64 * 3
#endif
#ifdef SANDYBRIDGE
#define MOVUPS_A movups
#define MOVUPS_XL movups
#define MOVUPS_XS movups
#define MOVUPS_YL movups
#define MOVUPS_YS movups
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE 64 * 3
#endif
#ifdef OPTERON
#define PREFETCH prefetch
#define PREFETCHW prefetchw

74
param.h
View File

@ -913,6 +913,80 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#ifdef SANDYBRIDGE
#define SNUMOPT 8
#define DNUMOPT 4
#define GEMM_DEFAULT_OFFSET_A 32
#define GEMM_DEFAULT_OFFSET_B 0
#define GEMM_DEFAULT_ALIGN 0x03fffUL
#define SYMV_P 8
#define SWITCH_RATIO 4
#ifdef ARCH_X86
#define SGEMM_DEFAULT_UNROLL_M 4
#define DGEMM_DEFAULT_UNROLL_M 2
#define QGEMM_DEFAULT_UNROLL_M 2
#define CGEMM_DEFAULT_UNROLL_M 2
#define ZGEMM_DEFAULT_UNROLL_M 1
#define XGEMM_DEFAULT_UNROLL_M 1
#define SGEMM_DEFAULT_UNROLL_N 4
#define DGEMM_DEFAULT_UNROLL_N 4
#define QGEMM_DEFAULT_UNROLL_N 2
#define CGEMM_DEFAULT_UNROLL_N 2
#define ZGEMM_DEFAULT_UNROLL_N 2
#define XGEMM_DEFAULT_UNROLL_N 1
#else
#define SGEMM_DEFAULT_UNROLL_M 4
#define DGEMM_DEFAULT_UNROLL_M 2
#define QGEMM_DEFAULT_UNROLL_M 2
#define CGEMM_DEFAULT_UNROLL_M 2
#define ZGEMM_DEFAULT_UNROLL_M 1
#define XGEMM_DEFAULT_UNROLL_M 1
#define SGEMM_DEFAULT_UNROLL_N 8
#define DGEMM_DEFAULT_UNROLL_N 8
#define QGEMM_DEFAULT_UNROLL_N 2
#define CGEMM_DEFAULT_UNROLL_N 4
#define ZGEMM_DEFAULT_UNROLL_N 4
#define XGEMM_DEFAULT_UNROLL_N 1
#endif
#define SGEMM_DEFAULT_P 504
#define SGEMM_DEFAULT_R sgemm_r
#define DGEMM_DEFAULT_P 504
#define DGEMM_DEFAULT_R dgemm_r
#define QGEMM_DEFAULT_P 504
#define QGEMM_DEFAULT_R qgemm_r
#define CGEMM_DEFAULT_P 252
#define CGEMM_DEFAULT_R cgemm_r
#define ZGEMM_DEFAULT_P 252
#define ZGEMM_DEFAULT_R zgemm_r
#define XGEMM_DEFAULT_P 252
#define XGEMM_DEFAULT_R xgemm_r
#define SGEMM_DEFAULT_Q 512
#define DGEMM_DEFAULT_Q 256
#define QGEMM_DEFAULT_Q 128
#define CGEMM_DEFAULT_Q 512
#define ZGEMM_DEFAULT_Q 256
#define XGEMM_DEFAULT_Q 128
#define GETRF_FACTOR 0.72
#endif
#ifdef ATOM
#define SNUMOPT 2