Support AMD Piledriver by bulldozer kernels.

Makefile.system

@@ -311,14 +311,14 @@ 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 BULLDOZER
+DYNAMIC_CORE += SANDYBRIDGE BULLDOZER PILEDRIVER
 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 BULLDOZER
+DYNAMIC_CORE += SANDYBRIDGE BULLDOZER PILEDRIVER
 endif
 endif
 

README.md

@@ -48,6 +48,7 @@ Please read GotoBLAS_01Readme.txt
 - **Intel Haswell**: Optimized Level-3 BLAS with AVX on x86-64 (identical to Sandy Bridge).
 - **AMD Bobcat**: Used GotoBLAS2 Barcelona codes.
 - **AMD Bulldozer**: x86-64 S/DGEMM AVX kernels. (Thank Werner Saar)
+- **AMD PILEDRIVER**: Used Bulldozer codes.
 
 #### MIPS64:
 - **ICT Loongson 3A**: Optimized Level-3 BLAS and the part of Level-1,2.

common_x86.h

@@ -171,6 +171,11 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 #define MMXSTORE	movd
 #endif
 
+#if defined(PILEDRIVER) || defined(BULLDOZER)
+//Enable some optimazation for barcelona.
+#define BARCELONA_OPTIMIZATION
+#endif
+
 #if defined(HAVE_3DNOW)
 #define EMMS	femms
 #elif defined(HAVE_MMX)

common_x86_64.h

@@ -218,6 +218,11 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 
 #ifdef ASSEMBLER
 
+#if defined(PILEDRIVER) || defined(BULLDOZER)
+//Enable some optimazation for barcelona.
+#define BARCELONA_OPTIMIZATION
+#endif
+
 #if defined(HAVE_3DNOW)
 #define EMMS	femms
 #elif defined(HAVE_MMX)

cpuid.h

@@ -106,6 +106,7 @@
 #define CORE_SANDYBRIDGE 20
 #define CORE_BOBCAT     21
 #define CORE_BULLDOZER  22
+#define CORE_PILEDRIVER  23
 #define CORE_HASWELL CORE_SANDYBRIDGE
 
 #define HAVE_SSE      (1 <<  0)
@@ -128,6 +129,7 @@
 #define HAVE_FASTMOVU    (1 << 17)
 #define HAVE_AVX      (1 <<  18)
 #define HAVE_FMA4     (1 <<  19)
+#define HAVE_FMA3     (1 <<  20)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2
@@ -197,6 +199,7 @@ typedef struct {
 #define CPUTYPE_SANDYBRIDGE             44
 #define CPUTYPE_BOBCAT                  45
 #define CPUTYPE_BULLDOZER               46
+#define CPUTYPE_PILEDRIVER              47
 // this define is because BLAS doesn't have haswell specific optimizations yet
 #define CPUTYPE_HASWELL CPUTYPE_SANDYBRIDGE 
 

cpuid_x86.c

@@ -47,6 +47,8 @@
 #define CORE_SANDYBRIDGE CORE_NEHALEM
 #define CPUTYPE_BULLDOZER CPUTYPE_BARCELONA
 #define CORE_BULLDOZER CORE_BARCELONA
+#define CPUTYPE_PILEDRIVER CPUTYPE_BARCELONA
+#define CORE_PILEDRIVER CORE_BARCELONA
 #endif
 
 #ifndef CPUIDEMU
@@ -228,6 +230,7 @@ int get_cputype(int gettype){
 #ifndef NO_AVX
     if (support_avx()) feature |= HAVE_AVX;
 #endif
+    if ((ecx & (1 << 20)) != 0) feature |= HAVE_FMA3;
 
     if (have_excpuid() >= 0x01) {
       cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
@@ -1100,11 +1103,21 @@ int get_cpuname(void){
       case  1:
       case 10:
 	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  6:
+	switch (model) {
+	case 1:
+	  //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 2:
+	  if(support_avx())
+	    return CPUTYPE_PILEDRIVER;
+	  else
+	    return CPUTYPE_BARCELONA; //OS don't support AVX.	  
+	}
+	break;
       case  5:
 	return CPUTYPE_BOBCAT;
       }
@@ -1229,6 +1242,7 @@ static char *cpuname[] = {
   "SANDYBRIDGE",
   "BOBCAT",
   "BULLDOZER",
+  "PILEDRIVER",
 };
 
 static char *lowercpuname[] = {
@@ -1278,6 +1292,7 @@ static char *lowercpuname[] = {
   "sandybridge",
   "bobcat",
   "bulldozer",
+  "piledriver",
 };
 
 static char *corename[] = {
@@ -1304,6 +1319,7 @@ static char *corename[] = {
   "SANDYBRIDGE",
   "BOBCAT",
   "BULLDOZER",
+  "PILEDRIVER",
 };
 
 static char *corename_lower[] = {
@@ -1330,6 +1346,7 @@ static char *corename_lower[] = {
   "sandybridge",
   "bobcat",
   "bulldozer",
+  "piledriver",
 };
 
 
@@ -1472,11 +1489,19 @@ int get_coretype(void){
       if ((exfamily == 0) || (exfamily == 2)) return CORE_OPTERON; 
       else if (exfamily == 5) return CORE_BOBCAT; 
       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.
+	switch (model) {
+	case 1:
+	  //AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
+	  if(support_avx())
+	    return CORE_BULLDOZER;
+	  else
+	    return CORE_BARCELONA; //OS don't support AVX.
+	case 2:
+	  if(support_avx())
+	    return CORE_PILEDRIVER;
+	  else
+	    return CORE_BARCELONA; //OS don't support AVX.	  
+	}
       }else return CORE_BARCELONA;
     }
   }
@@ -1564,6 +1589,7 @@ void get_cpuconfig(void){
     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_FMA3 )    printf("#define HAVE_FMA3\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");
@@ -1631,5 +1657,6 @@ void get_sse(void){
   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");
+  if (features & HAVE_FMA3 )    printf("HAVE_FMA3=1\n");
 
 }

driver/others/dynamic.c

@@ -64,10 +64,12 @@ extern gotoblas_t  gotoblas_BOBCAT;
 #ifndef NO_AVX
 extern gotoblas_t  gotoblas_SANDYBRIDGE;
 extern gotoblas_t  gotoblas_BULLDOZER;
+extern gotoblas_t  gotoblas_PILEDRIVER;
 #else
 //Use NEHALEM kernels for sandy bridge
 #define gotoblas_SANDYBRIDGE gotoblas_NEHALEM
 #define gotoblas_BULLDOZER gotoblas_BARCELONA
+#define gotoblas_PILEDRIVER gotoblas_BARCELONA
 #endif
 //Use sandy bridge kernels for haswell.
 #define gotoblas_HASWELL gotoblas_SANDYBRIDGE
@@ -228,13 +230,23 @@ static gotoblas_t *get_coretype(void){
       }  else if (exfamily == 5) {
 	return &gotoblas_BOBCAT;
       } else if (exfamily == 6) {
-	//AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
+	if(model == 1){
+	  //AMD Bulldozer Opteron 6200 / Opteron 4200 / AMD FX-Series
 	  if(support_avx())
 	    return &gotoblas_BULLDOZER;
 	  else{
 	    fprintf(stderr, "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Barcelona kernels as a fallback, which may give poorer performance.\n");
 	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
-	  }	
+	  }
+	}else if(model == 2){
+	  //AMD Bulldozer Opteron 6300 / Opteron 4300 / Opteron 3300
+	  if(support_avx())
+	    return &gotoblas_PILEDRIVER;
+	  else{
+	    fprintf(stderr, "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Barcelona kernels as a fallback, which may give poorer performance.\n");
+	    return &gotoblas_BARCELONA; //OS doesn't support AVX. Use old kernels.
+	  }
+	}
       } else {
 	return &gotoblas_BARCELONA;
       }
@@ -272,6 +284,7 @@ static char *corename[] = {
     "Sandybridge",
     "Bobcat",
     "Bulldozer",
+    "Piledriver",
 };
 
 char *gotoblas_corename(void) {
@@ -294,6 +307,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_SANDYBRIDGE)  return corename[16];
   if (gotoblas == &gotoblas_BOBCAT)       return corename[17];
   if (gotoblas == &gotoblas_BULLDOZER)    return corename[18];
+  if (gotoblas == &gotoblas_PILEDRIVER)    return corename[19];
 
   return corename[0];
 }

getarch.c

@@ -106,6 +106,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /* #define FORCE_ISTANBUL	*/
 /* #define FORCE_BOBCAT		*/
 /* #define FORCE_BULLDOZER	*/
+/* #define FORCE_PILEDRIVER     */
 /* #define FORCE_SSE_GENERIC	*/
 /* #define FORCE_VIAC3		*/
 /* #define FORCE_NANO		*/
@@ -398,6 +399,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "BULLDOZER"
 #endif
 
+#if defined (FORCE_PILEDRIVER)
+#define FORCE
+#define FORCE_INTEL
+#define ARCHITECTURE    "X86"
+#define SUBARCHITECTURE "PILEDRIVER"
+#define ARCHCONFIG   "-DPILEDRIVER " \
+		     "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=64 " \
+		     "-DL2_SIZE=2097152 -DL2_LINESIZE=64  -DL3_SIZE=12582912 " \
+		     "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+		     "-DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2" \
+		     "-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_128BITFPU -DHAVE_FASTMOVU -DHAVE_CFLUSH" \
+                     "-DHAVE_AVX -DHAVE_FMA4 -DHAVE_FMA3"
+#define LIBNAME   "piledriver"
+#define CORENAME  "PILEDRIVER"
+#endif
+
 #ifdef FORCE_SSE_GENERIC
 #define FORCE
 #define FORCE_INTEL

kernel/setparam-ref.c

@@ -826,6 +826,22 @@ static void init_parameter(void) {
 #endif
 #endif
 
+#ifdef PILEDRIVER
+
+#ifdef DEBUG
+  fprintf(stderr, "Piledriver\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

kernel/x86/KERNEL.PILEDRIVER

@@ -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

kernel/x86/trsm_kernel_LN_2x4_sse2.S

@@ -69,7 +69,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHSIZE (8 * 10 + 4)
 #endif
@@ -439,7 +439,7 @@
 .L22:
 	mulsd	%xmm0, %xmm2
 	addsd	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movlpd	 2 * SIZE(BB), %xmm2
@@ -488,7 +488,7 @@
 	movlpd	40 * SIZE(BB), %xmm3
 	addsd	%xmm0, %xmm7
 	movlpd	 8 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulsd	%xmm1, %xmm2
@@ -1697,7 +1697,7 @@
 
 .L42:
 	mulpd	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulpd	 2 * SIZE(BB), %xmm0
@@ -1727,7 +1727,7 @@
 	addpd	%xmm0, %xmm7
 	movapd	16 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulpd	%xmm1, %xmm2

kernel/x86/trsm_kernel_LN_4x4_sse.S

@@ -64,7 +64,7 @@
 #define BORIG	60(%esp)
 #define BUFFER 128(%esp)
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHW    prefetchw
 #define PREFETCHSIZE (16 * 10 + 8)
@@ -437,7 +437,7 @@
 .L32:
 	mulss	%xmm0, %xmm2
 	addss	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 4 * SIZE(BB), %xmm2
@@ -833,7 +833,7 @@
 .L22:
 	mulps	%xmm0, %xmm2
 	addps	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(BB), %xmm2
@@ -1848,7 +1848,7 @@
 
 .L72:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulss	 4 * SIZE(BB), %xmm0
@@ -2109,7 +2109,7 @@
 	ALIGN_4
 
 .L62:
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 
@@ -2429,7 +2429,7 @@
 
 .L52:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulps	 4 * SIZE(BB), %xmm0
@@ -2459,7 +2459,7 @@
 	addps	%xmm0, %xmm5
 	movaps	32 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm2
@@ -2952,7 +2952,7 @@
 
 .L112:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 1 * SIZE(AA), %xmm0
@@ -3148,7 +3148,7 @@
 
 .L102:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movsd	 2 * SIZE(AA), %xmm0
@@ -3389,7 +3389,7 @@
 
 .L92:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(AA), %xmm0
@@ -3404,7 +3404,7 @@
 	mulps	12 * SIZE(BB), %xmm0
 	addps	%xmm0, %xmm7
 	movaps	32 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm3

kernel/x86/trsm_kernel_LT_2x4_sse2.S

@@ -69,7 +69,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHSIZE (8 * 10 + 4)
 #endif
@@ -910,7 +910,7 @@
 .L22:
 	mulsd	%xmm0, %xmm2
 	addsd	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movlpd	 2 * SIZE(BB), %xmm2
@@ -959,7 +959,7 @@
 	movlpd	40 * SIZE(BB), %xmm3
 	addsd	%xmm0, %xmm7
 	movlpd	 8 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulsd	%xmm1, %xmm2
@@ -1439,7 +1439,7 @@
 
 .L42:
 	mulpd	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulpd	 2 * SIZE(BB), %xmm0
@@ -1469,7 +1469,7 @@
 	addpd	%xmm0, %xmm7
 	movapd	16 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulpd	%xmm1, %xmm2

kernel/x86/trsm_kernel_LT_4x4_sse.S

@@ -64,7 +64,7 @@
 #define BORIG	60(%esp)
 #define BUFFER 128(%esp)
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHW    prefetchw
 #define PREFETCHSIZE (16 * 10 + 8)
@@ -872,7 +872,7 @@
 .L22:
 	mulps	%xmm0, %xmm2
 	addps	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(BB), %xmm2
@@ -1316,7 +1316,7 @@
 .L32:
 	mulss	%xmm0, %xmm2
 	addss	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 4 * SIZE(BB), %xmm2
@@ -1855,7 +1855,7 @@
 
 .L52:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulps	 4 * SIZE(BB), %xmm0
@@ -1885,7 +1885,7 @@
 	addps	%xmm0, %xmm5
 	movaps	32 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm2
@@ -2249,7 +2249,7 @@
 	ALIGN_4
 
 .L62:
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 
@@ -2562,7 +2562,7 @@
 
 .L72:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulss	 4 * SIZE(BB), %xmm0
@@ -2957,7 +2957,7 @@
 
 .L92:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(AA), %xmm0
@@ -2972,7 +2972,7 @@
 	mulps	12 * SIZE(BB), %xmm0
 	addps	%xmm0, %xmm7
 	movaps	32 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm3
@@ -3280,7 +3280,7 @@
 
 .L102:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movsd	 2 * SIZE(AA), %xmm0
@@ -3515,7 +3515,7 @@
 
 .L112:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 1 * SIZE(AA), %xmm0

kernel/x86/trsm_kernel_RT_2x4_sse2.S

@@ -69,7 +69,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHSIZE (8 * 10 + 4)
 #endif
@@ -1036,7 +1036,7 @@
 
 .L42:
 	mulpd	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulpd	 2 * SIZE(BB), %xmm0
@@ -1066,7 +1066,7 @@
 	addpd	%xmm0, %xmm7
 	movapd	16 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulpd	%xmm1, %xmm2
@@ -2224,7 +2224,7 @@
 .L22:
 	mulsd	%xmm0, %xmm2
 	addsd	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movlpd	 2 * SIZE(BB), %xmm2
@@ -2273,7 +2273,7 @@
 	movlpd	40 * SIZE(BB), %xmm3
 	addsd	%xmm0, %xmm7
 	movlpd	 8 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	PREFETCH (PREFETCHSIZE  + 8) * SIZE(AA)
 #endif
 	mulsd	%xmm1, %xmm2

kernel/x86/trsm_kernel_RT_4x4_sse.S

@@ -64,7 +64,7 @@
 #define BORIG	60(%esp)
 #define BUFFER 128(%esp)
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH     prefetch
 #define PREFETCHW    prefetchw
 #define PREFETCHSIZE (16 * 10 + 8)
@@ -439,7 +439,7 @@
 
 .L92:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(AA), %xmm0
@@ -454,7 +454,7 @@
 	mulps	12 * SIZE(BB), %xmm0
 	addps	%xmm0, %xmm7
 	movaps	32 * SIZE(AA), %xmm0
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm3
@@ -758,7 +758,7 @@
 
 .L102:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movsd	 2 * SIZE(AA), %xmm0
@@ -993,7 +993,7 @@
 
 .L112:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 1 * SIZE(AA), %xmm0
@@ -1324,7 +1324,7 @@
 
 .L52:
 	mulps	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulps	 4 * SIZE(BB), %xmm0
@@ -1354,7 +1354,7 @@
 	addps	%xmm0, %xmm5
 	movaps	32 * SIZE(AA), %xmm0
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
 #endif
 	mulps	%xmm1, %xmm2
@@ -1718,7 +1718,7 @@
 	ALIGN_4
 
 .L62:
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 
@@ -2031,7 +2031,7 @@
 
 .L72:
 	mulss	%xmm0, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	mulss	 4 * SIZE(BB), %xmm0
@@ -2859,7 +2859,7 @@
 .L22:
 	mulps	%xmm0, %xmm2
 	addps	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movaps	 4 * SIZE(BB), %xmm2
@@ -3303,7 +3303,7 @@
 .L32:
 	mulss	%xmm0, %xmm2
 	addss	%xmm2, %xmm4
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
 #endif
 	movss	 4 * SIZE(BB), %xmm2

kernel/x86/ztrsm_kernel_LN_2x2_sse.S

@@ -75,7 +75,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCHSIZE (16 * 10 + 8)
 #define WPREFETCHSIZE 112
 #define PREFETCH      prefetch
@@ -533,7 +533,7 @@
 	addps	%xmm0, %xmm7
 	movsd	16 * SIZE(AA), %xmm0
 	mulps	%xmm1, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht1     (PREFETCHSIZE + 16) * SIZE(AA)
 #endif
 	addps	%xmm2, %xmm4

kernel/x86/ztrsm_kernel_LT_2x2_sse.S

@@ -75,7 +75,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCHSIZE (16 * 10 + 8)
 #define WPREFETCHSIZE 112
 #define PREFETCH      prefetch
@@ -994,7 +994,7 @@
 	addps	%xmm0, %xmm7
 	movsd	16 * SIZE(AA), %xmm0
 	mulps	%xmm1, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht1     (PREFETCHSIZE + 16) * SIZE(AA)
 #endif
 	addps	%xmm2, %xmm4

kernel/x86/ztrsm_kernel_RT_2x2_sse.S

@@ -75,7 +75,7 @@
 #define STACK_ALIGN	4096
 #define STACK_OFFSET	1024
 
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCHSIZE (16 * 10 + 8)
 #define WPREFETCHSIZE 112
 #define PREFETCH      prefetch
@@ -1820,7 +1820,7 @@
 	addps	%xmm0, %xmm7
 	movsd	16 * SIZE(AA), %xmm0
 	mulps	%xmm1, %xmm2
-#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 	prefetcht1     (PREFETCHSIZE + 16) * SIZE(AA)
 #endif
 	addps	%xmm2, %xmm4

kernel/x86_64/KERNEL.PILEDRIVER

@@ -0,0 +1,70 @@
+ZGEMVNKERNEL = zgemv_n_dup.S
+ZGEMVTKERNEL = zgemv_t_dup.S
+
+DGEMVNKERNEL = dgemv_n_bulldozer.S
+DGEMVTKERNEL = dgemv_t_bulldozer.S
+DAXPYKERNEL  = daxpy_bulldozer.S
+DDOTKERNEL   = ddot_bulldozer.S
+DCOPYKERNEL  = dcopy_bulldozer.S
+
+SGEMMKERNEL    =  sgemm_kernel_16x2_bulldozer.S
+SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
+SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
+SGEMMONCOPY    =  gemm_ncopy_2_bulldozer.S
+SGEMMOTCOPY    =  gemm_tcopy_2_bulldozer.S
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX) 
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX) 
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+DGEMMKERNEL    =  dgemm_kernel_8x2_bulldozer.S
+DGEMMINCOPY    =  dgemm_ncopy_8_bulldozer.S
+DGEMMITCOPY    =  dgemm_tcopy_8_bulldozer.S
+DGEMMONCOPY    =  gemm_ncopy_2_bulldozer.S
+DGEMMOTCOPY    =  gemm_tcopy_2_bulldozer.S
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+CGEMMKERNEL    =  cgemm_kernel_4x2_bulldozer.S
+CGEMMINCOPY    =  ../generic/zgemm_ncopy_4.c
+CGEMMITCOPY    =  ../generic/zgemm_tcopy_4.c
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_2.c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+ZGEMMKERNEL    =  zgemm_kernel_2x2_bulldozer.S
+ZGEMMINCOPY    =
+ZGEMMITCOPY    =
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_2.c
+ZGEMMINCOPYOBJ =
+ZGEMMITCOPYOBJ =
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
+ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
+
+STRSMKERNEL_LN  =  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT  =  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN  =  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT  =  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+

kernel/x86_64/symv_L_sse.S

@@ -76,7 +76,7 @@
 #define movsd		movlps
 #endif
 
-#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH	prefetch
 #define PREFETCHW	prefetchw
 #define PREFETCHSIZE	(16 * 16)

kernel/x86_64/symv_L_sse2.S

@@ -76,7 +76,7 @@
 #define movsd		movlpd
 #endif
 
-#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH	prefetch
 #define PREFETCHW	prefetchw
 #define PREFETCHSIZE	(16 * 16)

kernel/x86_64/symv_U_sse.S

@@ -76,7 +76,7 @@
 #define movsd		movlps
 #endif
 
-#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH	prefetch
 #define PREFETCHW	prefetchw
 #define PREFETCHSIZE	(16 * 16)

kernel/x86_64/symv_U_sse2.S

@@ -76,7 +76,7 @@
 #define movsd		movlpd
 #endif
 
-#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define PREFETCH	prefetch
 #define PREFETCHW	prefetchw
 #define PREFETCHSIZE	(16 * 16)

kernel/x86_64/zsymv_L_sse.S

@@ -160,7 +160,7 @@
 #define a3     %xmm14
 #define	xt1    %xmm15
 
-#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #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

@@ -76,7 +76,7 @@
 #define movsd		movlpd
 #endif
 
-#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #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) || defined(BULLDOZER)
+#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BARCELONA_OPTIMIZATION)
 #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

@@ -76,7 +76,7 @@
 #define movsd		movlpd
 #endif
 
-#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #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) || defined(BULLDOZER)
+#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BARCELONA_OPTIMIZATION)
 #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

@@ -76,7 +76,7 @@
 #define movsd		movlpd
 #endif
 
-#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA)  || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #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) || defined(BULLDOZER)
+#if (defined(HAVE_SSE3) && !defined(CORE_OPTERON)) || defined(BARCELONA) || defined(SHANGHAI) || defined(BARCELONA_OPTIMIZATION)
 #define MOVDDUP(a, b, c)	movddup	a(b), c
 #define MOVDDUP2(a, b, c)	movddup	a##b, c
 #else

l2param.h

@@ -85,7 +85,7 @@
 #define movsd		movlps
 #endif
 
-#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BULLDOZER)
+#if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT) || defined(BARCELONA_OPTIMIZATION)
 #define ALIGNED_ACCESS
 #define MOVUPS_A	movaps
 #define MOVUPS_XL	movaps

param.h

@@ -234,6 +234,86 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 
+#define QGEMM_DEFAULT_UNROLL_N 2
+#define CGEMM_DEFAULT_UNROLL_N 2
+#define ZGEMM_DEFAULT_UNROLL_N 2
+#define XGEMM_DEFAULT_UNROLL_N 1
+
+#ifdef ARCH_X86
+#define SGEMM_DEFAULT_UNROLL_N 4
+#define DGEMM_DEFAULT_UNROLL_N 4
+#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
+#else
+#define SGEMM_DEFAULT_UNROLL_N 2
+#define DGEMM_DEFAULT_UNROLL_N 2
+#define SGEMM_DEFAULT_UNROLL_M 16
+#define DGEMM_DEFAULT_UNROLL_M 8
+#define QGEMM_DEFAULT_UNROLL_M 2
+#define CGEMM_DEFAULT_UNROLL_M 4
+#define ZGEMM_DEFAULT_UNROLL_M 2
+#define XGEMM_DEFAULT_UNROLL_M 1
+#define CGEMM3M_DEFAULT_UNROLL_N 4
+#define CGEMM3M_DEFAULT_UNROLL_M 8
+#define ZGEMM3M_DEFAULT_UNROLL_N 4
+#define ZGEMM3M_DEFAULT_UNROLL_M 4
+#define GEMV_UNROLL 8
+#endif
+
+
+#if defined(ARCH_X86_64)
+#define SGEMM_DEFAULT_P 768
+#define DGEMM_DEFAULT_P 384
+#else
+#define SGEMM_DEFAULT_P 448
+#define DGEMM_DEFAULT_P 224
+#endif
+#define QGEMM_DEFAULT_P 112
+#define CGEMM_DEFAULT_P 224
+#define ZGEMM_DEFAULT_P 112 
+#define XGEMM_DEFAULT_P  56
+
+#if defined(ARCH_X86_64)
+#define SGEMM_DEFAULT_Q 168
+#define DGEMM_DEFAULT_Q 168
+#else
+#define SGEMM_DEFAULT_Q 224
+#define DGEMM_DEFAULT_Q 224
+#endif
+#define QGEMM_DEFAULT_Q 224
+#define CGEMM_DEFAULT_Q 224
+#define ZGEMM_DEFAULT_Q 224
+#define XGEMM_DEFAULT_Q 224
+
+#define SGEMM_DEFAULT_R sgemm_r
+#define QGEMM_DEFAULT_R qgemm_r
+#define DGEMM_DEFAULT_R dgemm_r
+#define CGEMM_DEFAULT_R cgemm_r
+#define ZGEMM_DEFAULT_R zgemm_r
+#define XGEMM_DEFAULT_R xgemm_r
+
+#define SYMV_P	16
+#define HAVE_EXCLUSIVE_CACHE
+
+#define GEMM_THREAD gemm_thread_mn
+
+#endif
+
+#ifdef PILEDRIVER
+
+#define SNUMOPT		8
+#define DNUMOPT		4
+
+#define GEMM_DEFAULT_OFFSET_A  64
+#define GEMM_DEFAULT_OFFSET_B 832
+#define GEMM_DEFAULT_ALIGN 0x0fffUL
+
+
+
 #define QGEMM_DEFAULT_UNROLL_N 2
 #define CGEMM_DEFAULT_UNROLL_N 2
 #define ZGEMM_DEFAULT_UNROLL_N 2