Merge branch 'bulldozer' into develop

OpenMathLib · Feb 9, 2013 · 5c8bf6a · 5c8bf6a
2 parents a9500d0 + 13f8fc0
commit 5c8bf6a
Show file tree

Hide file tree

Showing 55 changed files with 2,188 additions and 138 deletions.
diff --git a/Makefile.system b/Makefile.system
@@ -277,14 +277,14 @@ ifeq ($(ARCH), x86)
 DYNAMIC_CORE = KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS \
 	       CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO
 ifneq ($(NO_AVX), 1)
-DYNAMIC_CORE += SANDYBRIDGE 
+DYNAMIC_CORE += SANDYBRIDGE BULLDOZER
 endif
 endif
 
 ifeq ($(ARCH), x86_64)
 DYNAMIC_CORE = PRESCOTT CORE2 PENRYN DUNNINGTON NEHALEM OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO
 ifneq ($(NO_AVX), 1)
-DYNAMIC_CORE += SANDYBRIDGE 
+DYNAMIC_CORE += SANDYBRIDGE BULLDOZER
 endif
 endif
 

diff --git a/TargetList.txt b/TargetList.txt
@@ -29,6 +29,7 @@ BARCELONA
 SHANGHAI
 ISTANBUL
 BOBCAT
+BULLDOZER
 
 c)VIA CPU:
 SSE_GENERIC

diff --git a/cpuid.h b/cpuid.h
@@ -125,7 +125,8 @@
 #define HAVE_MISALIGNSSE (1 << 15)
 #define HAVE_128BITFPU   (1 << 16)
 #define HAVE_FASTMOVU    (1 << 17)
-#define HAVE_AVX     (1 <<  18)
+#define HAVE_AVX      (1 <<  18)
+#define HAVE_FMA4     (1 <<  19)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2

diff --git a/cpuid_x86.c b/cpuid_x86.c
@@ -43,6 +43,8 @@
 #ifdef NO_AVX
 #define CPUTYPE_SANDYBRIDGE CPUTYPE_NEHALEM
 #define CORE_SANDYBRIDGE CORE_NEHALEM
+#define CPUTYPE_BULLDOZER CPUTYPE_BARCELONA
+#define CORE_BULLDOZER CORE_BARCELONA
 #endif
 
 #ifndef CPUIDEMU
@@ -229,6 +231,9 @@ int get_cputype(int gettype){
       cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
       if ((ecx & (1 <<  6)) != 0) feature |= HAVE_SSE4A;
       if ((ecx & (1 <<  7)) != 0) feature |= HAVE_MISALIGNSSE;
+#ifndef NO_AVX
+      if ((ecx & (1 <<  16)) != 0) feature |= HAVE_FMA4;
+#endif
       if ((edx & (1 << 30)) != 0) feature |= HAVE_3DNOWEX;
       if ((edx & (1 << 31)) != 0) feature |= HAVE_3DNOW;
     }
@@ -1078,8 +1083,12 @@ int get_cpuname(void){
 	return CPUTYPE_OPTERON;
       case  1:
       case 10:
-      case  6:   //AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
 	return CPUTYPE_BARCELONA;
+      case  6:   //AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
+	if(support_avx())
+	  return CPUTYPE_BULLDOZER;
+	else
+	  return CPUTYPE_BARCELONA; //OS don't support AVX.
       case  5:
 	return CPUTYPE_BOBCAT;
       }
@@ -1432,8 +1441,13 @@ int get_coretype(void){
     if (family == 0xf){
       if ((exfamily == 0) || (exfamily == 2)) return CORE_OPTERON; 
       else if (exfamily == 5) return CORE_BOBCAT; 
-      else if (exfamily == 6) return CORE_BARCELONA;  //AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
-      else return CORE_BARCELONA;
+      else if (exfamily == 6) {
+	//AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
+	if(support_avx())
+	  return CORE_BULLDOZER;
+	else
+	  return CORE_BARCELONA; //OS don't support AVX. Use old kernels.
+      }else return CORE_BARCELONA;
     }
   }
 
@@ -1519,6 +1533,7 @@ void get_cpuconfig(void){
     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_FMA4 )    printf("#define HAVE_FMA4\n");
     if (features & HAVE_CFLUSH)  printf("#define HAVE_CFLUSH\n");
     if (features & HAVE_HIT)     printf("#define HAVE_HIT 1\n");
     if (features & HAVE_MISALIGNSSE) printf("#define HAVE_MISALIGNSSE\n");
@@ -1585,5 +1600,6 @@ void get_sse(void){
   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");
+  if (features & HAVE_FMA4 )    printf("HAVE_FMA4=1\n");
 
 }
diff --git a/driver/others/dynamic.c b/driver/others/dynamic.c
@@ -63,9 +63,11 @@ extern gotoblas_t  gotoblas_BARCELONA;
 extern gotoblas_t  gotoblas_BOBCAT;
 #ifndef NO_AVX
 extern gotoblas_t  gotoblas_SANDYBRIDGE;
+extern gotoblas_t  gotoblas_BULLDOZER;
 #else
 //Use NEHALEM kernels for sandy bridge
 #define gotoblas_SANDYBRIDGE gotoblas_NEHALEM
+#define gotoblas_BULLDOZER gotoblas_BARCELONA
 #endif
 
 
@@ -204,6 +206,14 @@ static gotoblas_t *get_coretype(void){
 	else return &gotoblas_OPTERON;
       }  else if (exfamily == 5) {
 	return &gotoblas_BOBCAT;
+      } else if (exfamily == 6) {
+	//AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
+	  if(support_avx())
+	    return &gotoblas_BULLDOZER;
+	  else{
+	    fprintf(stderr, "OpenBLAS : Your OS doesn't support AVX. Use Barcelona kernels.\n");
+	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
+	  }	
       } else {
 	return &gotoblas_BARCELONA;
       }
@@ -240,6 +250,7 @@ static char *corename[] = {
     "Nano",
     "Sandybridge",
     "Bobcat",
+    "Bulldozer",
 };
 
 char *gotoblas_corename(void) {
@@ -261,6 +272,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_NANO)         return corename[15];
   if (gotoblas == &gotoblas_SANDYBRIDGE)  return corename[16];
   if (gotoblas == &gotoblas_BOBCAT)       return corename[17];
+  if (gotoblas == &gotoblas_BULLDOZER)    return corename[18];
 
   return corename[0];
 }

diff --git a/driver/others/parameter.c b/driver/others/parameter.c
@@ -163,7 +163,7 @@ int get_L2_size(void){
 
   int eax, ebx, ecx, edx;
 
-#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || \
+#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
     defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
   defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC)
 

diff --git a/getarch.c b/getarch.c
@@ -350,7 +350,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "OPTERON"
 #endif
 
-#if defined(FORCE_BARCELONA) || defined(FORCE_SHANGHAI) || defined(FORCE_ISTANBUL) || defined (FORCE_BULLDOZER)
+#if defined(FORCE_BARCELONA) || defined(FORCE_SHANGHAI) || defined(FORCE_ISTANBUL)
 #define FORCE
 #define FORCE_INTEL
 #define ARCHITECTURE    "X86"
@@ -380,6 +380,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "BOBCAT"
 #endif
 
+#if defined (FORCE_BULLDOZER)
+#define FORCE
+#define FORCE_INTEL
+#define ARCHITECTURE    "X86"
+#define SUBARCHITECTURE "BULLDOZER"
+#define ARCHCONFIG   "-DBULLDOZER " \
+		     "-DL1_DATA_SIZE=49152 -DL1_DATA_LINESIZE=64 " \
+		     "-DL2_SIZE=1024000 -DL2_LINESIZE=64  -DL3_SIZE=16777216 " \
+		     "-DDTB_DEFAULT_ENTRIES=32 -DDTB_SIZE=4096 " \
+		     "-DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 " \
+		     "-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_128BITFPU -DHAVE_FASTMOVU " \
+                     "-DHAVE_AVX -DHAVE_FMA4"
+#define LIBNAME   "bulldozer"
+#define CORENAME  "BULLDOZER"
+#endif
+
 #ifdef FORCE_SSE_GENERIC
 #define FORCE
 #define FORCE_INTEL

diff --git a/kernel/setparam-ref.c b/kernel/setparam-ref.c
@@ -810,6 +810,22 @@ static void init_parameter(void) {
 #endif
 #endif
 
+#ifdef BULLDOZER
+
+#ifdef DEBUG
+  fprintf(stderr, "Bulldozer\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 NANO
 
 #ifdef DEBUG

diff --git a/kernel/x86/KERNEL.BULLDOZER b/kernel/x86/KERNEL.BULLDOZER
@@ -0,0 +1,59 @@
+SGEMMKERNEL    =  gemm_kernel_4x4_barcelona.S
+SGEMMINCOPY    =  
+SGEMMITCOPY    =  
+SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+SGEMMINCOPYOBJ =  
+SGEMMITCOPYOBJ =  
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+DGEMMKERNEL    =  gemm_kernel_2x4_barcelona.S
+DGEMMINCOPY    =  ../generic/gemm_ncopy_2.c
+DGEMMITCOPY    =  ../generic/gemm_tcopy_2.c
+DGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+CGEMMKERNEL    =  zgemm_kernel_2x2_barcelona.S
+CGEMMINCOPY    =  
+CGEMMITCOPY    =  
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_2.c
+CGEMMINCOPYOBJ =
+CGEMMITCOPYOBJ =  
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ZGEMMKERNEL    =  zgemm_kernel_1x2_barcelona.S
+ZGEMMINCOPY    =  ../generic/zgemm_ncopy_1.c
+ZGEMMITCOPY    =  ../generic/zgemm_tcopy_1.c
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_2.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_4x4_sse.S
+STRSMKERNEL_LT	=  trsm_kernel_LT_4x4_sse.S
+STRSMKERNEL_RN	=  trsm_kernel_LT_4x4_sse.S
+STRSMKERNEL_RT	=  trsm_kernel_RT_4x4_sse.S
+
+DTRSMKERNEL_LN	=  trsm_kernel_LN_2x4_sse2.S
+DTRSMKERNEL_LT	=  trsm_kernel_LT_2x4_sse2.S
+DTRSMKERNEL_RN	=  trsm_kernel_LT_2x4_sse2.S
+DTRSMKERNEL_RT	=  trsm_kernel_RT_2x4_sse2.S
+
+CTRSMKERNEL_LN	=  ztrsm_kernel_LN_2x2_sse.S
+CTRSMKERNEL_LT	=  ztrsm_kernel_LT_2x2_sse.S
+CTRSMKERNEL_RN	=  ztrsm_kernel_LT_2x2_sse.S
+CTRSMKERNEL_RT	=  ztrsm_kernel_RT_2x2_sse.S
+
+ZTRSMKERNEL_LN	=  ztrsm_kernel_LT_1x2_sse2.S
+ZTRSMKERNEL_LT	=  ztrsm_kernel_LT_1x2_sse2.S
+ZTRSMKERNEL_RN	=  ztrsm_kernel_LT_1x2_sse2.S
+ZTRSMKERNEL_RT	=  ztrsm_kernel_RT_1x2_sse2.S
+
+CGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
+ZGEMM3MKERNEL    =  zgemm3m_kernel_2x4_barcelona.S
diff --git a/kernel/x86/gemm_kernel_4x4_barcelona.S b/kernel/x86/gemm_kernel_4x4_barcelona.S
@@ -596,7 +596,7 @@
 .L22:
 	mulps	%xmm0, %xmm2
 	addps	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movsd	 4 * SIZE(BB), %xmm2
@@ -842,7 +842,7 @@
 .L32:
 	mulss	%xmm0, %xmm2
 	addss	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 4 * SIZE(BB), %xmm2
@@ -1168,7 +1168,7 @@
 
 .L52:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulps	 4 * SIZE(BB), %xmm0
@@ -1198,7 +1198,7 @@
 	addps	%xmm0, %xmm5
 	movaps	32 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm2
@@ -1347,7 +1347,7 @@
 	ALIGN_4
 
 .L62:
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 
@@ -1531,7 +1531,7 @@
 
 .L72:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulss	 4 * SIZE(BB), %xmm0
@@ -1778,7 +1778,7 @@
 
 .L92:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(AA), %xmm0
@@ -1793,7 +1793,7 @@
 	mulps	12 * SIZE(BB), %xmm0
 	addps	%xmm0, %xmm7
 	movaps	32 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm3
@@ -1924,7 +1924,7 @@
 
 .L102:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movsd	 2 * SIZE(AA), %xmm0
@@ -2069,7 +2069,7 @@
 
 .L112:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BULLDOZER)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 1 * SIZE(AA), %xmm0

diff --git a/kernel/x86/scal_sse.S b/kernel/x86/scal_sse.S
@@ -269,7 +269,7 @@
 	sarl	$5, I
 	jle	.L113
 
-#if defined(BARCELONA)
+#if defined(BARCELONA) || defined(BULLDOZER)
 
 	movaps	%xmm0, %xmm1
 	mulps	-32 * SIZE(X), %xmm1

diff --git a/kernel/x86/scal_sse2.S b/kernel/x86/scal_sse2.S
@@ -253,7 +253,7 @@
 	sarl	$4, I
 	jle	.L113
 
-#if defined(BARCELONA)
+#if defined(BARCELONA) || defined(BULLDOZER)
 
 	movaps  %xmm0, %xmm1
 	mulpd	-16 * SIZE(X), %xmm1