Merge pull request #2 from xianyi/develop

update

.travis.yml

@@ -25,14 +25,14 @@ matrix:
         - TARGET_BOX=LINUX64
         - BTYPE="BINARY=64"
 
-    - <<: *test-ubuntu
-      os: linux-ppc64le
-      before_script:
-        - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=POWER8 NUM_THREADS=32"
-      env:
-        # for matrix annotation only
-        - TARGET_BOX=PPC64LE_LINUX
-        - BTYPE="BINARY=64 USE_OPENMP=1"
+        #    - <<: *test-ubuntu
+        #      os: linux-ppc64le
+        #      before_script:
+        #        - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=POWER8 NUM_THREADS=32"
+        #      env:
+        #        # for matrix annotation only
+        #        - TARGET_BOX=PPC64LE_LINUX
+        #        - BTYPE="BINARY=64 USE_OPENMP=1"
 
     - <<: *test-ubuntu
       env:

Makefile.system

@@ -699,7 +699,7 @@ endif
 
 ifeq ($(C_COMPILER), PGI)
 ifdef BINARY64
-CCOMMON_OPT += -tp p7-64
+CCOMMON_OPT += -tp p7-64 -D__MMX__ -Mnollvm
 else
 CCOMMON_OPT += -tp p7
 endif

cmake/prebuild.cmake

@@ -106,7 +106,29 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
   file(APPEND ${TARGET_CONF_TEMP}
     "#define ${TCORE}\n"
     "#define CHAR_CORENAME \"${TCORE}\"\n")
-  if ("${TCORE}" STREQUAL "ARMV7")
+  if ("${TCORE}" STREQUAL "CORE2")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L2_SIZE\t1048576\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define DTB_DEFAULT_ENTRIES\t256\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_CMOV\n"
+      "#define HAVE_MMX\n"
+      "#define HAVE_SSE\n"
+      "#define HAVE_SSE2\n"
+      "#define HAVE_SSE3\n"
+      "#define HAVE_SSSE3\n")
+      set(SGEMM_UNROLL_M 8)
+      set(SGEMM_UNROLL_N 4)
+      set(DGEMM_UNROLL_M 4)
+      set(DGEMM_UNROLL_N 4)
+      set(CGEMM_DEFAULT_UNROLL_M 4)
+      set(CGEMM_DEFAULT_UNROLL_N 2)
+      set(ZGEMM_DEFAULT_UNROLL_M 2)
+      set(ZGEMM_DEFAULT_UNROLL_N 2)      
+  elseif ("${TCORE}" STREQUAL "ARMV7")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_DATA_SIZE\t65536\n"
       "#define L1_DATA_LINESIZE\t32\n"

kernel/power/caxpy.c

@@ -24,12 +24,21 @@ CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
-
 #include "common.h"
- 
-
 #ifndef HAVE_ASM_KERNEL
 #include <altivec.h> 
+
+#define  offset_0 0
+#define  offset_1 16
+#define  offset_2 32
+#define  offset_3 48
+#define  offset_4 64
+#define  offset_5 80
+#define  offset_6 96
+#define  offset_7 112
+
+static const unsigned char __attribute__((aligned(16))) swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+
 static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i)
 {
 
@@ -43,28 +52,29 @@ static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT
     register __vector float valpha_i = {alpha_i, alpha_i,alpha_i, alpha_i};
 #endif
 
-    __vector unsigned char swap_mask = { 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
-    register __vector float *vy = (__vector float *) y;
-    register __vector float *vx = (__vector float *) x;
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+    register __vector float *vptr_y = (__vector float *) y;
+    register __vector float *vptr_x = (__vector float *) x; 
     BLASLONG i=0;
-    for (; i < n/2; i += 8) {
+    for(;i<n/2;i+=8){ 
+
+        register __vector float vy_0 = vec_vsx_ld( offset_0 ,vptr_y ) ;
+        register __vector float vy_1 = vec_vsx_ld( offset_1 ,vptr_y ) ;
+        register __vector float vy_2 = vec_vsx_ld( offset_2 ,vptr_y ) ;
+        register __vector float vy_3 = vec_vsx_ld( offset_3 ,vptr_y ) ;
+        register __vector float vy_4 = vec_vsx_ld( offset_4 ,vptr_y ) ;
+        register __vector float vy_5 = vec_vsx_ld( offset_5 ,vptr_y ) ;
+        register __vector float vy_6 = vec_vsx_ld( offset_6 ,vptr_y ) ;
+        register __vector float vy_7 = vec_vsx_ld( offset_7 ,vptr_y ) ;
 
-        register __vector float vy_0 = vy[i];
-        register __vector float vy_1 = vy[i + 1];
-        register __vector float vy_2 = vy[i + 2];
-        register __vector float vy_3 = vy[i + 3];
-        register __vector float vy_4 = vy[i + 4];
-        register __vector float vy_5 = vy[i + 5];
-        register __vector float vy_6 = vy[i + 6];
-        register __vector float vy_7 = vy[i + 7];
-        register __vector float vx_0 = vx[i];
-        register __vector float vx_1 = vx[i + 1];
-        register __vector float vx_2 = vx[i + 2];
-        register __vector float vx_3 = vx[i + 3];
-        register __vector float vx_4 = vx[i + 4];
-        register __vector float vx_5 = vx[i + 5];
-        register __vector float vx_6 = vx[i + 6];
-        register __vector float vx_7 = vx[i + 7];
+        register __vector float vx_0 = vec_vsx_ld( offset_0 ,vptr_x ) ;
+        register __vector float vx_1 = vec_vsx_ld( offset_1 ,vptr_x ) ;
+        register __vector float vx_2 = vec_vsx_ld( offset_2 ,vptr_x ) ;
+        register __vector float vx_3 = vec_vsx_ld( offset_3 ,vptr_x ) ;
+        register __vector float vx_4 = vec_vsx_ld( offset_4 ,vptr_x ) ;
+        register __vector float vx_5 = vec_vsx_ld( offset_5 ,vptr_x ) ;
+        register __vector float vx_6 = vec_vsx_ld( offset_6 ,vptr_x ) ;
+        register __vector float vx_7 = vec_vsx_ld( offset_7 ,vptr_x ) ;
         vy_0 += vx_0*valpha_r;
         vy_1 += vx_1*valpha_r;
         vy_2 += vx_2*valpha_r;
@@ -89,15 +99,17 @@ static void caxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT
         vy_5 += vx_5*valpha_i;
         vy_6 += vx_6*valpha_i;
         vy_7 += vx_7*valpha_i;
-        vy[i] = vy_0;
-        vy[i + 1] = vy_1;
-        vy[i + 2] = vy_2;
-        vy[i + 3] = vy_3;
-        vy[i + 4] = vy_4;
-        vy[i + 5] = vy_5 ;
-        vy[i + 6] = vy_6 ;
-        vy[i + 7] = vy_7 ;        
+        vec_vsx_st( vy_0, offset_0 ,vptr_y ) ;
+        vec_vsx_st( vy_1, offset_1 ,vptr_y ) ;
+        vec_vsx_st( vy_2, offset_2 ,vptr_y ) ;
+        vec_vsx_st( vy_3, offset_3 ,vptr_y ) ;
+        vec_vsx_st( vy_4, offset_4 ,vptr_y ) ;
+        vec_vsx_st( vy_5, offset_5 ,vptr_y ) ;
+        vec_vsx_st( vy_6, offset_6 ,vptr_y ) ;
+        vec_vsx_st( vy_7, offset_7 ,vptr_y ) ;   
 
+        vptr_x+=8;
+        vptr_y+=8;   
     }
 }
 #endif

kernel/power/cdot.c

@@ -25,15 +25,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  *****************************************************************************/
 
 #include "common.h"
-
 #ifndef HAVE_KERNEL_8
 #include <altivec.h> 
+
+#define  offset_0 0
+#define  offset_1 16
+#define  offset_2 32
+#define  offset_3 48
+
+
+
+static const unsigned char __attribute__((aligned(16))) swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
 static void cdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, float *dot)
 {
-    __vector unsigned char swap_mask = { 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
-    register __vector float *vy = (__vector float *) y;
-    register __vector float *vx = (__vector float *) x;
-    BLASLONG i = 0;
+    __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
+    register __vector float *vptr_y = (__vector float *) y;
+    register __vector float *vptr_x = (__vector float *) x;
     register __vector float vd_0  = { 0 };
     register __vector float vd_1  = { 0 };
     register __vector float vd_2  = { 0 };
@@ -41,26 +48,23 @@ static void cdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, float *dot)
     register __vector float vdd_0 = { 0 };
     register __vector float vdd_1 = { 0 };
     register __vector float vdd_2 = { 0 };
-    register __vector float vdd_3 = { 0 };
-    for (; i < n/2; i += 4) {
-
-        register __vector float vyy_0 ;
-        register __vector float vyy_1 ;
-        register __vector float vyy_2 ;
-        register __vector float vyy_3 ;
-
-        register __vector float vy_0 = vy[i];
-        register __vector float vy_1 = vy[i + 1];
-        register __vector float vy_2 = vy[i + 2];
-        register __vector float vy_3 = vy[i + 3]; 
-        register __vector float vx_0= vx[i];
-        register __vector float vx_1 = vx[i + 1];
-        register __vector float vx_2 = vx[i + 2];
-        register __vector float vx_3 = vx[i + 3]; 
-        vyy_0 = vec_perm(vy_0, vy_0, swap_mask);
-        vyy_1 = vec_perm(vy_1, vy_1, swap_mask);
-        vyy_2 = vec_perm(vy_2, vy_2, swap_mask);
-        vyy_3 = vec_perm(vy_3, vy_3, swap_mask);  
+    register __vector float vdd_3 = { 0 };     
+    BLASLONG i=0;
+    for(;i<n/2;i+=4){ 
+
+        register __vector float vy_0 = vec_vsx_ld( offset_0 ,vptr_y ) ;
+        register __vector float vy_1 = vec_vsx_ld( offset_1 ,vptr_y ) ;
+        register __vector float vy_2 = vec_vsx_ld( offset_2 ,vptr_y ) ;
+        register __vector float vy_3 = vec_vsx_ld( offset_3 ,vptr_y ) ; 
+
+        register __vector float vx_0 = vec_vsx_ld( offset_0 ,vptr_x ) ;
+        register __vector float vx_1 = vec_vsx_ld( offset_1 ,vptr_x ) ;
+        register __vector float vyy_0 = vec_perm(vy_0, vy_0, swap_mask);
+        register __vector float vyy_1 = vec_perm(vy_1, vy_1, swap_mask);
+        register __vector float vx_2 = vec_vsx_ld( offset_2 ,vptr_x ) ;
+        register __vector float vx_3 = vec_vsx_ld( offset_3 ,vptr_x ) ;  
+        register __vector float vyy_2 = vec_perm(vy_2, vy_2, swap_mask);
+        register __vector float vyy_3 = vec_perm(vy_3, vy_3, swap_mask);  
 
         vd_0 += vx_0 * vy_0;
         vd_1 += vx_1 * vy_1;
@@ -72,6 +76,8 @@ static void cdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, float *dot)
         vdd_2 += vx_2 * vyy_2;
         vdd_3 += vx_3 * vyy_3;       
        
+       vptr_x+=4;
+       vptr_y+=4;
 
     }
     //aggregate
@@ -96,7 +102,7 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
     BLASLONG i = 0;
     BLASLONG ix=0, iy=0;
     OPENBLAS_COMPLEX_FLOAT result;
-    FLOAT dot[4] __attribute__ ((aligned(16))) = {0.0, 0.0, 0.0, 0.0};
+    FLOAT dot[4] __attribute__((aligned(16))) = {0.0, 0.0, 0.0, 0.0};
 
     if (n <= 0) {
         CREAL(result) = 0.0;

kernel/power/cgemv_n.c

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define NBMAX 1024
 
 
-static const unsigned char swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+static const unsigned char __attribute__((aligned(16))) swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
 
  
 static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
@@ -62,23 +62,24 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
     register __vector float vx3_r = {x[6], -x[6],x[6], -x[6]};
     register __vector float vx3_i = {x[7], x[7],x[7], x[7]};
 #endif
-    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_y = (__vector float *) y;
     register __vector float *vptr_a0 = (__vector float *) a0;
     register __vector float *vptr_a1 = (__vector float *) a1;
     register __vector float *vptr_a2 = (__vector float *) a2;
     register __vector float *vptr_a3 = (__vector float *) a3; 
     BLASLONG  i = 0; 
-    for (;i< n / 2; i+=2) {
-        register __vector float vy_0  = vy[i];
-        register __vector float vy_1  = vy[i + 1];
-        register __vector float va0   = vptr_a0[i];
-        register __vector float va1   = vptr_a1[i];
-        register __vector float va2   = vptr_a2[i];
-        register __vector float va3   = vptr_a3[i];
-        register __vector float va0_1 = vptr_a0[i + 1];
-        register __vector float va1_1 = vptr_a1[i + 1];
-        register __vector float va2_1 = vptr_a2[i + 1];
-        register __vector float va3_1 = vptr_a3[i + 1];
+    BLASLONG i2=16;
+    for (;i< n * 8; i+=32,i2+=32) {
+        register __vector float vy_0  = vec_vsx_ld(i,vptr_y);
+        register __vector float vy_1  = vec_vsx_ld(i2,vptr_y);
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0);
+        register __vector float va1   = vec_vsx_ld(i, vptr_a1);
+        register __vector float va2   = vec_vsx_ld(i ,vptr_a2);
+        register __vector float va3   = vec_vsx_ld(i ,vptr_a3);
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0);
+        register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1);
+        register __vector float va2_1 = vec_vsx_ld(i2 ,vptr_a2);
+        register __vector float va3_1 = vec_vsx_ld(i2 ,vptr_a3);
 
         vy_0 += va0*vx0_r + va1*vx1_r + va2*vx2_r + va3*vx3_r;
         vy_1 += va0_1*vx0_r + va1_1*vx1_r + va2_1*vx2_r + va3_1*vx3_r;
@@ -93,8 +94,8 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
         vy_0 += va0*vx0_i + va1*vx1_i + va2*vx2_i + va3*vx3_i;
         vy_1 += va0_1*vx0_i + va1_1*vx1_i + va2_1*vx2_i + va3_1*vx3_i;
 
-        vy[i] = vy_0;
-        vy[i + 1] = vy_1;
+        vec_vsx_st(vy_0 ,i,  vptr_y);
+        vec_vsx_st(vy_1,i2,vptr_y);
     }
 
 }	
@@ -118,17 +119,19 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
     register __vector float vx1_r = {x[2], -x[2],x[2], -x[2]};
     register __vector float vx1_i = {x[3], x[3],x[3], x[3]}; 
 #endif
-    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_y = (__vector float *) y;
     register __vector float *vptr_a0 = (__vector float *) a0;
     register __vector float *vptr_a1 = (__vector float *) a1; 
-    BLASLONG  i = 0; 
-    for (;i< n / 2; i+=2) {
-        register __vector float vy_0  = vy[i];
-        register __vector float vy_1  = vy[i + 1];
-        register __vector float va0   = vptr_a0[i];
-        register __vector float va1   = vptr_a1[i]; 
-        register __vector float va0_1 = vptr_a0[i + 1];
-        register __vector float va1_1 = vptr_a1[i + 1]; 
+    BLASLONG  i = 0;
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vy_0  = vec_vsx_ld(i,vptr_y);
+        register __vector float vy_1  = vec_vsx_ld(i2,vptr_y);
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0);
+        register __vector float va1   = vec_vsx_ld(i, vptr_a1); 
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0);
+        register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); 
+
         register __vector float va0x   = vec_perm(va0, va0,swap_mask);
         register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask);
         register __vector float va1x   = vec_perm(va1, va1,swap_mask);
@@ -136,8 +139,8 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
         vy_0 += va0*vx0_r + va1*vx1_r + va0x*vx0_i + va1x*vx1_i;
         vy_1 += va0_1*vx0_r + va1_1*vx1_r + va0x_1*vx0_i + va1x_1*vx1_i; 
 
-        vy[i] = vy_0;
-        vy[i + 1] = vy_1;
+        vec_vsx_st(vy_0 ,i,  vptr_y);
+        vec_vsx_st(vy_1,i2,vptr_y);
     }
 
 }
@@ -154,21 +157,23 @@ static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) {
     register __vector float vx0_r = {x[0], -x[0],x[0], -x[0]};
     register __vector float vx0_i = {x[1], x[1],x[1], x[1]}; 
 #endif
-    register __vector float *vy = (__vector float *) y;
+    register __vector float *vptr_y = (__vector float *) y;
     register __vector float *vptr_a0 = (__vector float *) ap; 
     BLASLONG  i = 0; 
-    for (;i< n / 2; i+=2) {
-        register __vector float vy_0  = vy[i];
-        register __vector float vy_1  = vy[i + 1];
-        register __vector float va0   = vptr_a0[i];
-        register __vector float va0_1 = vptr_a0[i + 1]; 
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vy_0  = vec_vsx_ld(i,vptr_y);
+        register __vector float vy_1  = vec_vsx_ld(i2,vptr_y);
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0); 
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); 
+
         register __vector float va0x   = vec_perm(va0, va0,swap_mask);
         register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask);
         vy_0 += va0*vx0_r + va0x*vx0_i;
         vy_1 += va0_1*vx0_r + va0x_1*vx0_i; 
 
-        vy[i] = vy_0;
-        vy[i + 1] = vy_1;
+        vec_vsx_st(vy_0 ,i,  vptr_y);
+        vec_vsx_st(vy_1,i2,vptr_y);
     }
 }
 
@@ -176,7 +181,7 @@ static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) {
 
 
 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT alpha_r, FLOAT alpha_i) {
-    BLASLONG i;
+    BLASLONG i=0;
 
 
     if (inc_dest != 2) {
@@ -213,20 +218,24 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT
 
         register __vector float *vptr_src = (__vector float *) src;
         register __vector float *vptr_y = (__vector float *) dest; 
-        for (i = 0; i < n/2; i += 2 ){
 
-            register __vector float vy_0 = vptr_y[i];
-            register __vector float vy_1 = vptr_y[i +1]; 
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vy_0  = vec_vsx_ld(i,vptr_y);
+        register __vector float vy_1  = vec_vsx_ld(i2,vptr_y);
+
 
-            register __vector float vsrc = vptr_src[i];
-            register __vector float vsrc_1 = vptr_src[i + 1]; 
-            register __vector float vsrcx = vec_perm(vsrc, vsrc, swap_mask);
-            register __vector float vsrcx_1 = vec_perm(vsrc_1, vsrc_1, swap_mask);
+        register __vector float vsrc  = vec_vsx_ld(i,vptr_src);
+        register __vector float vsrc_1  = vec_vsx_ld(i2,vptr_src);
 
-            vy_0 += vsrc*valpha_r + vsrcx*valpha_i;
-            vy_1 += vsrc_1*valpha_r +  vsrcx_1*valpha_i;  
-            vptr_y[i] = vy_0;
-            vptr_y[i+1 ] = vy_1;  
+        register __vector float vsrcx = vec_perm(vsrc, vsrc, swap_mask);
+        register __vector float vsrcx_1 = vec_perm(vsrc_1, vsrc_1, swap_mask);
+
+        vy_0 += vsrc*valpha_r + vsrcx*valpha_i;
+        vy_1 += vsrc_1*valpha_r +  vsrcx_1*valpha_i;  
+
+        vec_vsx_st(vy_0 ,i,  vptr_y);
+        vec_vsx_st(vy_1,i2,vptr_y);
 
         }
  
@@ -237,7 +246,7 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT
 
 
 int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT * buffer) {
-    BLASLONG i;
+    BLASLONG i=0;
     FLOAT *a_ptr;
     FLOAT *x_ptr;
     FLOAT *y_ptr;
@@ -247,8 +256,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
     BLASLONG m2;
     BLASLONG m3;
     BLASLONG n2;
-
-    FLOAT xbuffer[8], *ybuffer;
+    FLOAT xbuffer[8] __attribute__((aligned(16)));
+    FLOAT *ybuffer;
 
     if (m < 1) return (0);
     if (n < 1) return (0);

kernel/power/cgemv_t.c

@@ -29,10 +29,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define NBMAX 1024 
 #include <altivec.h> 
-static const unsigned char swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
+static const unsigned char __attribute__((aligned(16))) swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11};
 
 static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
-    BLASLONG i;
+
     FLOAT *a0, *a1, *a2, *a3;
     a0 = ap;
     a1 = ap + lda;
@@ -48,26 +48,39 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
     register __vector float vtemp2_r = {0.0, 0.0,0.0,0.0};
     register __vector float vtemp3_p = {0.0, 0.0,0.0,0.0};
     register __vector float vtemp3_r = {0.0, 0.0,0.0,0.0};
-    __vector float* va0 = (__vector float*) a0;
-    __vector float* va1 = (__vector float*) a1;
-    __vector float* va2 = (__vector float*) a2;
-    __vector float* va3 = (__vector float*) a3;
+    __vector float* vptr_a0 = (__vector float*) a0;
+    __vector float* vptr_a1 = (__vector float*) a1;
+    __vector float* vptr_a2 = (__vector float*) a2;
+    __vector float* vptr_a3 = (__vector float*) a3;
     __vector float* v_x = (__vector float*) x;
 
-    for (i = 0; i < n / 2; i+=2) {
-        register __vector float vx_0  = v_x[i]; 
-        register __vector float vx_1  = v_x[i+1];         
+    BLASLONG  i = 0;
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vx_0  = vec_vsx_ld( i,v_x) ; 
+        register __vector float vx_1  = vec_vsx_ld(i2, v_x); 
+                
         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
 
-        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
-        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 
-        vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];
-        vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1]; 
-        vtemp2_p += vx_0*va2[i] + vx_1*va2[i+1];
-        vtemp2_r += vxr_0*va2[i] + vxr_1*va2[i+1]; 
-        vtemp3_p += vx_0*va3[i] + vx_1*va3[i+1];
-        vtemp3_r += vxr_0*va3[i] + vxr_1*va3[i+1]; 
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0);
+        register __vector float va1   = vec_vsx_ld(i, vptr_a1);
+        register __vector float va2   = vec_vsx_ld(i ,vptr_a2);
+        register __vector float va3   = vec_vsx_ld(i ,vptr_a3);
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0);
+        register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1);
+        register __vector float va2_1 = vec_vsx_ld(i2 ,vptr_a2);
+        register __vector float va3_1 = vec_vsx_ld(i2 ,vptr_a3);
+
+
+        vtemp0_p += vx_0*va0 + vx_1*va0_1 ;
+        vtemp0_r += vxr_0*va0 + vxr_1*va0_1; 
+        vtemp1_p += vx_0*va1 + vx_1*va1_1;
+        vtemp1_r += vxr_0*va1 + vxr_1*va1_1; 
+        vtemp2_p += vx_0*va2 + vx_1*va2_1;
+        vtemp2_r += vxr_0*va2 + vxr_1*va2_1; 
+        vtemp3_p += vx_0*va3 + vx_1*va3_1;
+        vtemp3_r += vxr_0*va3 + vxr_1*va3_1; 
 
     }
 
@@ -128,7 +141,7 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
  
 
 static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
-    BLASLONG i;
+
     FLOAT *a0, *a1;
     a0 = ap;
     a1 = ap + lda; 
@@ -138,23 +151,33 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
     register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0};
     register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0};
     register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0}; 
-    __vector float* va0 = (__vector float*) a0;
-    __vector float* va1 = (__vector float*) a1; 
+
+
+    __vector float* vptr_a0 = (__vector float*) a0;
+    __vector float* vptr_a1 = (__vector float*) a1; 
     __vector float* v_x = (__vector float*) x;
 
-    for (i = 0; i < n / 2; i+=2) {
-        register __vector float vx_0  = v_x[i]; 
-        register __vector float vx_1  = v_x[i+1];         
+    BLASLONG  i = 0;
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vx_0  = vec_vsx_ld( i,v_x) ; 
+        register __vector float vx_1  = vec_vsx_ld(i2, v_x); 
+                
         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
 
-        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
-        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; 
-        vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1];
-        vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1];  
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0);
+        register __vector float va1   = vec_vsx_ld(i, vptr_a1); 
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0);
+        register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); 
 
-    }
 
+        vtemp0_p += vx_0*va0 + vx_1*va0_1 ;
+        vtemp0_r += vxr_0*va0 + vxr_1*va0_1; 
+        vtemp1_p += vx_0*va1 + vx_1*va1_1;
+        vtemp1_r += vxr_0*va1 + vxr_1*va1_1;  
+
+    }
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
 
     register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3];
@@ -193,23 +216,27 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
  
 
 static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
-   BLASLONG i;  
+ 
     __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr);
     //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real)
     register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0};
     register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0}; 
-    __vector float* va0 = (__vector float*) ap; 
+    __vector float* vptr_a0 = (__vector float*) ap; 
     __vector float* v_x = (__vector float*) x;
-
-    for (i = 0; i < n / 2; i+=2) {
-        register __vector float vx_0  = v_x[i]; 
-        register __vector float vx_1  = v_x[i+1];         
+    BLASLONG  i = 0;
+    BLASLONG  i2 = 16;  
+    for (;i< n * 8; i+=32, i2+=32) { 
+        register __vector float vx_0  = vec_vsx_ld( i,v_x) ; 
+        register __vector float vx_1  = vec_vsx_ld(i2, v_x); 
+                
         register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask);
         register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask);
 
-        vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ;
-        vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1];  
+        register __vector float va0   = vec_vsx_ld(i,vptr_a0); 
+        register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0);  
 
+        vtemp0_p += vx_0*va0 + vx_1*va0_1 ;
+        vtemp0_r += vxr_0*va0 + vxr_1*va0_1;  
     }
 
 #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
@@ -249,8 +276,8 @@ static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {
 }
 
 int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer) {
-    BLASLONG i;
-    BLASLONG j;
+    BLASLONG i=0;
+    BLASLONG j=0;
     FLOAT *a_ptr;
     FLOAT *x_ptr;
     FLOAT *y_ptr;
@@ -260,8 +287,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
     BLASLONG m2;
     BLASLONG m3;
     BLASLONG n2;
-
-    FLOAT ybuffer[8], *xbuffer;
+    FLOAT ybuffer[8] __attribute__((aligned(16)));
+    FLOAT *xbuffer;
 
     if (m < 1) return (0);
     if (n < 1) return (0);

kernel/power/dgemv_n.c

@@ -145,7 +145,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG m3;
 	BLASLONG n2;
 	BLASLONG lda4 =  lda << 2;
-	FLOAT xbuffer[8] __attribute__ ((aligned (16)));;
+	FLOAT xbuffer[8] __attribute__ ((aligned (16)));
 	FLOAT *ybuffer;
 
         if ( m < 1 ) return(0);

kernel/power/dgemv_t.c

@@ -581,9 +581,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
     BLASLONG m1;
     BLASLONG m2;
     BLASLONG m3;
-    BLASLONG n2;
-
-    FLOAT ybuffer[8], *xbuffer;
+    BLASLONG n2; 
+    FLOAT ybuffer[8] __attribute__((aligned(16)));
+    FLOAT *xbuffer;
 
     if (m < 1) return (0);
     if (n < 1) return (0);

kernel/power/sgemv_n.c

@@ -174,7 +174,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	BLASLONG lda4 =  lda << 2;
 	BLASLONG lda8 =  lda << 3;
-	FLOAT xbuffer[8],*ybuffer;
+	FLOAT xbuffer[8] __attribute__((aligned(16)));
+	FLOAT *ybuffer;
 
         if ( m < 1 ) return(0);
         if ( n < 1 ) return(0);

kernel/power/sgemv_n_8.c

@@ -213,7 +213,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
 	BLASLONG n2;
 	BLASLONG lda4 =  lda << 2;
 	BLASLONG lda8 =  lda << 3;
-	FLOAT xbuffer[8],*ybuffer;
+	FLOAT xbuffer[8] __attribute__((aligned(16)));
+	FLOAT *ybuffer;
 
         if ( m < 1 ) return(0);
         if ( n < 1 ) return(0);

kernel/power/sgemv_t.c

@@ -177,10 +177,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
     BLASLONG m1;
     BLASLONG m2;
     BLASLONG m3;
-    BLASLONG n2;
-
-    FLOAT ybuffer[8], *xbuffer;
-
+    BLASLONG n2; 
+    FLOAT ybuffer[8] __attribute__((aligned(16)));
+    FLOAT *xbuffer; 
     if (m < 1) return (0);
     if (n < 1) return (0);
 

kernel/power/sgemv_t_8.c

@@ -204,8 +204,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
     BLASLONG m3;
     BLASLONG n2;
 
-    FLOAT ybuffer[8], *xbuffer;
-
+    FLOAT ybuffer[8] __attribute__((aligned(16)));
+    FLOAT *xbuffer; 
     if (m < 1) return (0);
     if (n < 1) return (0);
 

kernel/power/zgemv_n_4.c

@@ -614,8 +614,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
     BLASLONG m2;
     BLASLONG m3;
     BLASLONG n2;
-
-    FLOAT xbuffer[8], *ybuffer;
+    FLOAT xbuffer[8] __attribute__((aligned(16)));
+    FLOAT *ybuffer;
 
     if (m < 1) return (0);
     if (n < 1) return (0);

kernel/power/zgemv_t_4.c

@@ -532,8 +532,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
     BLASLONG m2;
     BLASLONG m3;
     BLASLONG n2;
-
-    FLOAT ybuffer[8], *xbuffer;
+    FLOAT ybuffer[8] __attribute__((aligned(16)));
+    FLOAT *xbuffer;
 
     if (m < 1) return (0);
     if (n < 1) return (0);

kernel/x86_64/zdot.c

@@ -181,11 +181,19 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
 #if defined(SMP)
 	int nthreads;
 	FLOAT dummy_alpha;
+#if defined(C_PGI) || defined(C_SUN)	
+	FLOAT zdotr=0., zdoti=0.;
+#endif	
 #endif
+	
 	OPENBLAS_COMPLEX_FLOAT zdot;
-       CREAL(zdot) = 0.0;
-       CIMAG(zdot) = 0.0;
-
+#if defined(C_PGI) || defined(C_SUN)	
+        zdot=OPENBLAS_MAKE_COMPLEX_FLOAT(0.0,0.0);
+#else
+	CREAL(zdot) = 0.0;
+	CIMAG(zdot) = 0.0;
+#endif
+	
 #if defined(SMP)
 	if (inc_x == 0 || inc_y == 0 || n <= 10000)
 		nthreads = 1;
@@ -211,15 +219,23 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
 
 		ptr = (OPENBLAS_COMPLEX_FLOAT *)result;
 		for (i = 0; i < nthreads; i++) {
+#if defined(C_PGI) || defined(C_SUN)			
+			zdotr += CREAL(*ptr);
+			zdoti += CIMAG(*ptr);
+#else			
 			CREAL(zdot) = CREAL(zdot) + CREAL(*ptr);
 			CIMAG(zdot) = CIMAG(zdot) + CIMAG(*ptr);
+#endif
 			ptr = (void *)(((char *)ptr) + sizeof(double) * 2);
 		}
+#if defined(C_PGI) || defined(C_SUN)		
+	zdot = OPENBLAS_MAKE_COMPLEX_FLOAT(zdotr,zdoti);
+#endif
 	}
 #else
 	zdot_compute(n, x, inc_x, y, inc_y, &zdot);
 #endif
-
+	
 	return zdot;
 }