Init Sandybridge codes based on Nehalem.

snake9906 · Mar 30, 2012 · 19a48b8 · 19a48b8
1 parent 5cbbc49
commit 19a48b8
Show file tree

Hide file tree

Showing 57 changed files with 309 additions and 45 deletions.
diff --git a/Makefile.system b/Makefile.system
@@ -226,11 +226,11 @@ 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
@@ -740,6 +740,7 @@ 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

diff --git a/TargetList.txt b/TargetList.txt
@@ -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

diff --git a/cpuid.h b/cpuid.h
@@ -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
diff --git a/cpuid_x86.c b/cpuid_x86.c
@@ -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);
@@ -983,7 +984,7 @@ 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;
@@ -1140,6 +1141,7 @@ static char *cpuname[] = {
   "NSGEODE",
   "VIAC3",
   "NANO",
+  "SANDYBRIDGE",
 };
 
 static char *lowercpuname[] = {
@@ -1186,6 +1188,7 @@ static char *lowercpuname[] = {
   "tms3x00",
   "nsgeode",
   "nano",
+  "sandybridge",
 };
 
 static char *corename[] = {
@@ -1209,6 +1212,7 @@ static char *corename[] = {
   "NEHALEM",
   "ATOM",
   "NANO",
+  "SANDYBRIDGE",
 };
 
 static char *corename_lower[] = {
@@ -1232,6 +1236,7 @@ static char *corename_lower[] = {
   "nehalem",
   "atom",
   "nano",
+  "sandybridge",
 };
 
 
@@ -1315,7 +1320,7 @@ 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;
@@ -1414,6 +1419,7 @@ 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");
@@ -1479,6 +1485,7 @@ 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");
 

diff --git a/driver/others/parameter.c b/driver/others/parameter.c
@@ -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;
 

diff --git a/getarch.c b/getarch.c
@@ -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

diff --git a/kernel/setparam-ref.c b/kernel/setparam-ref.c
@@ -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

diff --git a/kernel/x86/KERNEL.SANDYBRIDGE b/kernel/x86/KERNEL.SANDYBRIDGE
@@ -0,0 +1 @@
+include $(KERNELDIR)/KERNEL.PENRYN
diff --git a/kernel/x86/gemm_kernel_2x4_penryn.S b/kernel/x86/gemm_kernel_2x4_penryn.S
@@ -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

diff --git a/kernel/x86/gemm_kernel_4x4_penryn.S b/kernel/x86/gemm_kernel_4x4_penryn.S
@@ -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

diff --git a/kernel/x86/gemv_n_sse.S b/kernel/x86/gemv_n_sse.S
@@ -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)

diff --git a/kernel/x86/gemv_n_sse2.S b/kernel/x86/gemv_n_sse2.S
@@ -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)

diff --git a/kernel/x86/gemv_t_sse.S b/kernel/x86/gemv_t_sse.S
@@ -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)

diff --git a/kernel/x86/gemv_t_sse2.S b/kernel/x86/gemv_t_sse2.S
@@ -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)

diff --git a/kernel/x86/trsm_kernel_LN_2x4_penryn.S b/kernel/x86/trsm_kernel_LN_2x4_penryn.S
@@ -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

diff --git a/kernel/x86/trsm_kernel_LN_4x4_penryn.S b/kernel/x86/trsm_kernel_LN_4x4_penryn.S
@@ -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

diff --git a/kernel/x86/trsm_kernel_LT_2x4_penryn.S b/kernel/x86/trsm_kernel_LT_2x4_penryn.S
@@ -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

diff --git a/kernel/x86/trsm_kernel_LT_4x4_penryn.S b/kernel/x86/trsm_kernel_LT_4x4_penryn.S
@@ -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

diff --git a/kernel/x86/trsm_kernel_RT_2x4_penryn.S b/kernel/x86/trsm_kernel_RT_2x4_penryn.S
@@ -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

diff --git a/kernel/x86/trsm_kernel_RT_4x4_penryn.S b/kernel/x86/trsm_kernel_RT_4x4_penryn.S
@@ -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

diff --git a/kernel/x86/zgemm_kernel_1x2_penryn.S b/kernel/x86/zgemm_kernel_1x2_penryn.S
@@ -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

diff --git a/kernel/x86/zgemm_kernel_2x2_penryn.S b/kernel/x86/zgemm_kernel_2x2_penryn.S
@@ -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

diff --git a/kernel/x86/zgemv_n_sse.S b/kernel/x86/zgemv_n_sse.S
@@ -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)

diff --git a/kernel/x86/zgemv_n_sse2.S b/kernel/x86/zgemv_n_sse2.S
@@ -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)

diff --git a/kernel/x86/zgemv_t_sse.S b/kernel/x86/zgemv_t_sse.S
@@ -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)

diff --git a/kernel/x86/zgemv_t_sse2.S b/kernel/x86/zgemv_t_sse2.S
@@ -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)