small fixes, some (i(dz)amin,i(dz)amax,(dz)dot,(dz)asum) mikrokernels can be inlined

diff --git a/kernel/zarch/dasum.c b/kernel/zarch/dasum.c
index c0b5ab9..46dfcc0 100644 (file)
--- a/kernel/zarch/dasum.c
+++ b/kernel/zarch/dasum.c
@@ -36,12 +36,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 
-static FLOAT __attribute__ ((noinline)) dasum_kernel_32(BLASLONG n, FLOAT *x) {
-
-    __asm__ (
-            "pfd     1, 0(%1)   \n\t"
-            "sllg    %%r0,%0,3  \n\t"
-            "agr     %%r0,%1    \n\t"   
+ 
+
+static   FLOAT  dasum_kernel_32(BLASLONG n, FLOAT *x) {
+    FLOAT asum   ; 
+    __asm__  (
+            "pfd     1, 0(%3)   \n\t"
+            "sllg    %%r0,%2,3  \n\t"
+            "agr     %%r0,%3    \n\t"   
             "vzero   %%v0       \n\t"
             "vzero   %%v1       \n\t"
             "vzero   %%v2       \n\t"
@@ -95,15 +97,18 @@ static FLOAT __attribute__ ((noinline)) dasum_kernel_32(BLASLONG n, FLOAT *x) {
             "vfadb   %%v0,%%v25,%%v24   \n\t"
             "vrepg   %%v1,%%v0,1        \n\t"
             "adbr    %%f0,%%f1          \n\t"
-            :
-            : "r"(n), "a"(x)
-            : "cc", "memory","r0","f0","f1","v0","v1","v2","v3","v24","v25","v26","v27","v28","v29","v30","v31"
+            "ldr     %0,%%f0  \n\t" 
+            : "=f"(asum),"+&a"(x)
+            : "r"(n), "1"(x)
+            : "cc",   "r0" ,"f0","f1","v0","v1","v2","v3","v24","v25","v26","v27","v28","v29","v30","v31"
             );
+      return asum;
 
 }
 
 
 
+
 FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
     BLASLONG i = 0;
     BLASLONG j = 0;

diff --git a/kernel/zarch/ddot.c b/kernel/zarch/ddot.c
index e4a2bd9..f557466 100644 (file)
--- a/kernel/zarch/ddot.c
+++ b/kernel/zarch/ddot.c
@@ -30,75 +30,76 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #if  defined(Z13)
-static void __attribute__ ((noinline))   ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+static  FLOAT  ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y)
 {
-   
-                 __asm__ volatile( 
-                    "pfd 1, 0(%1) \n\t"
-                    "pfd 1, 0(%2) \n\t"      
-                    "vzero %%v24  \n\t"
-                    "vzero %%v25  \n\t" 
-                    "vzero %%v26  \n\t"
-                    "vzero %%v27  \n\t"                  
-                    "srlg  %%r0,%0,4    \n\t" 
-                    "xgr   %%r1,%%r1    \n\t"
-                    ".align 16 \n\t"    
-                    "1: \n\t"
-                    "pfd 1, 256(%%r1,%1) \n\t"
-                    "pfd 1, 256(%%r1,%2) \n\t"                
-                    "vl  %%v16, 0(%%r1,%1)  \n\t"
-                    "vl  %%v17, 16(%%r1,%1) \n\t"
-                    "vl  %%v18, 32(%%r1,%1) \n\t"
-                    "vl  %%v19, 48(%%r1,%1) \n\t"
-
-                    "vl  %%v28, 0(%%r1,%2) \n\t"
-                    "vfmadb    %%v24,%%v16,%%v28,%%v24  \n\t"  
-                    "vl  %%v29, 16(%%r1,%2) \n\t"
-                    "vfmadb    %%v25,%%v17,%%v29,%%v25  \n\t"   
+    FLOAT dot;
+         __asm__ volatile( 
+            "pfd 1, 0(%2) \n\t"
+            "pfd 1, 0(%3) \n\t"      
+            "vzero %%v24  \n\t"
+            "vzero %%v25  \n\t" 
+            "vzero %%v26  \n\t"
+            "vzero %%v27  \n\t"                  
+            "srlg  %1,%1,4    \n\t" 
+            "xgr   %%r1,%%r1    \n\t"
+            ".align 16 \n\t"    
+            "1: \n\t"
+            "pfd 1, 256(%%r1,%2) \n\t"
+            "pfd 1, 256(%%r1,%3) \n\t"                
+            "vl  %%v16, 0(%%r1,%2)  \n\t"
+            "vl  %%v17, 16(%%r1,%2) \n\t"
+            "vl  %%v18, 32(%%r1,%2) \n\t"
+            "vl  %%v19, 48(%%r1,%2) \n\t"
+
+            "vl  %%v28, 0(%%r1,%3) \n\t"
+            "vfmadb    %%v24,%%v16,%%v28,%%v24  \n\t"  
+            "vl  %%v29, 16(%%r1,%3) \n\t"
+            "vfmadb    %%v25,%%v17,%%v29,%%v25  \n\t"   
+     
+            "vl  %%v30, 32(%%r1,%3) \n\t"
+            "vfmadb    %%v26,%%v18,%%v30,%%v26  \n\t"      
+            "vl  %%v31, 48(%%r1,%3) \n\t" 
+            "vfmadb    %%v27,%%v19,%%v31,%%v27  \n\t"   
+ 
+            "vl  %%v16, 64(%%r1,%2)  \n\t"
+            "vl  %%v17, 80(%%r1,%2)  \n\t"
+            "vl  %%v18, 96(%%r1,%2)  \n\t"
+            "vl  %%v19, 112(%%r1,%2) \n\t"
+
+            "vl  %%v28, 64(%%r1,%3) \n\t"
+            "vfmadb    %%v24,%%v16,%%v28,%%v24  \n\t"  
+            "vl  %%v29, 80(%%r1,%3) \n\t"
+            "vfmadb    %%v25,%%v17,%%v29,%%v25  \n\t"  
+          
+     
+            "vl  %%v30, 96(%%r1,%3)  \n\t"
+            "vfmadb    %%v26,%%v18,%%v30,%%v26  \n\t" 
+            "vl  %%v31, 112(%%r1,%3) \n\t" 
+            "vfmadb    %%v27,%%v19,%%v31,%%v27  \n\t"  
              
-                    "vl  %%v30, 32(%%r1,%2) \n\t"
-                    "vfmadb    %%v26,%%v18,%%v30,%%v26  \n\t"      
-                    "vl  %%v31, 48(%%r1,%2) \n\t" 
-                    "vfmadb    %%v27,%%v19,%%v31,%%v27  \n\t"   
-         
-                    "vl  %%v16, 64(%%r1,%1)  \n\t"
-                    "vl  %%v17, 80(%%r1,%1)  \n\t"
-                    "vl  %%v18, 96(%%r1,%1)  \n\t"
-                    "vl  %%v19, 112(%%r1,%1) \n\t"
-
-                    "vl  %%v28, 64(%%r1,%2) \n\t"
-                    "vfmadb    %%v24,%%v16,%%v28,%%v24  \n\t"  
-                    "vl  %%v29, 80(%%r1,%2) \n\t"
-                    "vfmadb    %%v25,%%v17,%%v29,%%v25  \n\t"  
-                  
-             
-                    "vl  %%v30, 96(%%r1,%2)  \n\t"
-                    "vfmadb    %%v26,%%v18,%%v30,%%v26  \n\t" 
-                    "vl  %%v31, 112(%%r1,%2) \n\t" 
-                    "vfmadb    %%v27,%%v19,%%v31,%%v27  \n\t"  
-                     
-                    
-                    "la %%r1,128(%%r1) \n\t"
-                    "brctg %%r0,1b \n\t"
-                    "vfadb   %%v24,%%v25,%%v24    \n\t"
-                    "vfadb   %%v24,%%v26,%%v24    \n\t"
-                    "vfadb   %%v24,%%v27,%%v24    \n\t"                 
-                    "vrepg   %%v1,%%v24,1         \n\t"
-                    "vfadb   %%v1,%%v24,%%v1      \n\t"  
-                    " std    %%f1,0(%3)           \n\t"  
-                    :
-                    :"r"(n),"a"(x),"a"(y),"a"(d)
-                    :"cc" , "memory" ,"r0","r1","v16", "v17","v18","v19","v20","v21","v22","v23",
-                    "v24","v25","v26","v27","v28","v29","v30","v31"
-
-                 );
+            
+            "la %%r1,128(%%r1) \n\t"
+            "brctg %1,1b \n\t"
+            "vfadb   %%v24,%%v25,%%v24    \n\t"
+            "vfadb   %%v24,%%v26,%%v24    \n\t"
+            "vfadb   %%v24,%%v27,%%v24    \n\t"                 
+            "vrepg   %%v1,%%v24,1         \n\t"
+            "vfadb   %%v1,%%v24,%%v1      \n\t"  
+            "ldr %0,  %%f1     \n\t"  
+            : "=f"(dot) ,"+&r"(n)
+            : "a"(x),"a"(y) 
+            :"cc" , "r1","v16", "v17","v18","v19","v20","v21","v22","v23",
+            "v24","v25","v26","v27","v28","v29","v30","v31"
+
+         );
+    return dot;        
 
 }
 
 
 #else
 
-static void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
+static FLOAT ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y )
 {
     BLASLONG register i = 0;
     FLOAT dot = 0.0;
@@ -117,8 +118,8 @@ static void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
               i+=8 ;
 
        }
-       *d += dot;
-
+    return dot;
+    
 }
 
 #endif
@@ -136,9 +137,9 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
     {
 
         BLASLONG n1 = n & -16;
-
+        
         if ( n1 )
-            ddot_kernel_8(n1, x, y , &dot );
+            dot = ddot_kernel_8(n1, x, y  );
 
         i = n1;
         while(i < n)

diff --git a/kernel/zarch/dgemv_n_4.c b/kernel/zarch/dgemv_n_4.c
index 5897b2c..bb202e7 100644 (file)
--- a/kernel/zarch/dgemv_n_4.c
+++ b/kernel/zarch/dgemv_n_4.c
@@ -186,9 +186,7 @@ static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *xo, FLOAT *y, FLOAT *
 
 #endif
 
-
  
-static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest) __attribute__ ((noinline));
 
 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
 {

diff --git a/kernel/zarch/idamax.c b/kernel/zarch/idamax.c
index 7926adf..b8a22aa 100644 (file)
--- a/kernel/zarch/idamax.c
+++ b/kernel/zarch/idamax.c
@@ -37,29 +37,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
-static BLASLONG __attribute__((noinline)) diamax_kernel_32_TUNED(BLASLONG n, FLOAT *x, FLOAT *maxf) {
-
-    __asm__(
-            "pfd 1, 0(%1) \n\t"
-            "sllg   %%r0,%0,3  \n\t"
-            "agr    %%r0,%1    \n\t"
-            "VLEIG  %%v20,0,0  \n\t"
-            "VLEIG  %%v20,1,1  \n\t"
-            "VLEIG  %%v21,2,0  \n\t"
-            "VLEIG  %%v21,3,1  \n\t"
-            "VLEIG  %%v22,4,0  \n\t"
-            "VLEIG  %%v22,5,1  \n\t"
-            "VLEIG  %%v23,6,0  \n\t"
-            "VLEIG  %%v23,7,1  \n\t"
-            "VREPIG %%v4,8     \n\t"
+ 
+/**
+ * Find  maximum index 
+ * Warning: requirements n>0  and n % 32 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) maximum absolute value .( only for output )
+ * @return  index 
+ */
+static BLASLONG diamax_kernel_32_TUNED(BLASLONG n, FLOAT *x, FLOAT *maxf) {
+    BLASLONG index;
+   __asm__(
+            "pfd 1, 0(%4) \n\t"
+            "sllg   %%r0,%3,3  \n\t"
+            "agr    %%r0,%4    \n\t"
+            "vleig  %%v20,0,0  \n\t"
+            "vleig  %%v20,1,1  \n\t"
+            "vleig  %%v21,2,0  \n\t"
+            "vleig  %%v21,3,1  \n\t"
+            "vleig  %%v22,4,0  \n\t"
+            "vleig  %%v22,5,1  \n\t"
+            "vleig  %%v23,6,0  \n\t"
+            "vleig  %%v23,7,1  \n\t"
+            "vrepig %%v4,8     \n\t"
             "vzero %%v5        \n\t"
             "vzero %%v18       \n\t"
             "vzero %%v19       \n\t"
             ".align 16 \n\t"
             "1: \n\t"
-            "pfd 1, 256(%1 ) \n\t"
-            "vlm  %%v24,%%v31, 0(%1 ) \n\t"
-
+            "pfd 1, 256(%2 ) \n\t"
+            "vlm  %%v24,%%v31, 0(%2 ) \n\t"
             "vflpdb  %%v24, %%v24 \n\t"
             "vflpdb  %%v25, %%v25 \n\t"
             "vflpdb  %%v26, %%v26 \n\t"
@@ -68,7 +76,6 @@ static BLASLONG __attribute__((noinline)) diamax_kernel_32_TUNED(BLASLONG n, FLO
             "vflpdb  %%v29, %%v29 \n\t"
             "vflpdb  %%v30, %%v30 \n\t"
             "vflpdb  %%v31, %%v31 \n\t"
-
             "vfchdb  %%v16,%%v25,%%v24  \n\t "
             "vfchdb  %%v17,%%v27,%%v26  \n\t "
             "vsel    %%v1,%%v21,%%v20,%%v16 \n\t"
@@ -82,30 +89,24 @@ static BLASLONG __attribute__((noinline)) diamax_kernel_32_TUNED(BLASLONG n, FLO
             "vsel    %%v26,%%v23,%%v22,%%v17 \n\t"
             "vsel    %%v27,%%v31,%%v30,%%v17 \n\t"
 
-
             "vfchdb  %%v28, %%v3,%%v0         \n\t"
             "vfchdb  %%v29,%%v27, %%v25       \n\t"
             "vsel    %%v1,%%v2,%%v1,%%v28     \n\t"
             "vsel    %%v0,%%v3,%%v0,%%v28     \n\t"
             "vsel    %%v24,%%v26,%%v24,%%v29  \n\t"
             "vsel    %%v25,%%v27,%%v25,%%v29  \n\t"
-
-            "VAG %%v1,%%v1,%%v5    \n\t"
-            "VAG %%v24,%%v24,%%v5  \n\t"
-            "VAG %%v24,%%v24,%%v4  \n\t"
-
+            "vag %%v1,%%v1,%%v5    \n\t"
+            "vag %%v24,%%v24,%%v5  \n\t"
+            "vag %%v24,%%v24,%%v4  \n\t"
             "vfchdb %%v16,%%v25 , %%v0     \n\t"
-            "VAG    %%v5,%%v5,%%v4         \n\t"
+            "vag    %%v5,%%v5,%%v4         \n\t"
             "vsel   %%v29,%%v25,%%v0,%%v16 \n\t"
             "vsel   %%v28,%%v24,%%v1,%%v16 \n\t"
-
             "vfchdb %%v17, %%v29,%%v18   \n\t"
             "vsel  %%v19,%%v28,%%v19,%%v17 \n\t"
             "vsel    %%v18,%%v29,%%v18,%%v17 \n\t"
-
-            "VAG %%v5,%%v5,%%v4        \n\t"
-            "vlm  %%v24,%%v31,128(%1 ) \n\t"
-
+            "vag %%v5,%%v5,%%v4        \n\t"
+            "vlm  %%v24,%%v31,128(%2 ) \n\t"
             "vflpdb  %%v24, %%v24 \n\t"
             "vflpdb  %%v25, %%v25 \n\t"
             "vflpdb  %%v26, %%v26 \n\t"
@@ -114,7 +115,6 @@ static BLASLONG __attribute__((noinline)) diamax_kernel_32_TUNED(BLASLONG n, FLO
             "vflpdb  %%v29, %%v29 \n\t"
             "vflpdb  %%v30, %%v30 \n\t"
             "vflpdb  %%v31, %%v31 \n\t"
-
             "vfchdb  %%v16,%%v25,%%v24  \n\t "
             "vfchdb  %%v17,%%v27,%%v26  \n\t "
             "vsel    %%v1,%%v21,%%v20,%%v16 \n\t"
@@ -128,56 +128,53 @@ static BLASLONG __attribute__((noinline)) diamax_kernel_32_TUNED(BLASLONG n, FLO
             "vsel    %%v26,%%v23,%%v22,%%v17 \n\t"
             "vsel    %%v27,%%v31,%%v30,%%v17 \n\t"
 
-
             "vfchdb  %%v28, %%v3,%%v0     \n\t"
             "vfchdb  %%v29,%%v27, %%v25   \n\t"
             "vsel    %%v1,%%v2,%%v1,%%v28 \n\t"
             "vsel    %%v0,%%v3,%%v0,%%v28 \n\t"
             "vsel    %%v24,%%v26,%%v24,%%v29 \n\t"
             "vsel    %%v25,%%v27,%%v25,%%v29 \n\t"
-
-            "VAG %%v1,%%v1,%%v5     \n\t"
-            "VAG %%v24,%%v24,%%v5   \n\t"
-            "la %1,256(%1)          \n\t"
-            "VAG %%v24,%%v24,%%v4   \n\t"
-
+            "vag %%v1,%%v1,%%v5     \n\t"
+            "vag %%v24,%%v24,%%v5   \n\t"
+            "la %2,256(%2)          \n\t"
+            "vag %%v24,%%v24,%%v4   \n\t"
             "vfchdb  %%v16,%%v25 , %%v0  \n\t"
-            "VAG     %%v5,%%v5,%%v4      \n\t"
+            "vag     %%v5,%%v5,%%v4      \n\t"
             "vsel    %%v29,%%v25,%%v0,%%v16 \n\t"
             "vsel    %%v28,%%v24,%%v1,%%v16 \n\t"
-
             "vfchdb  %%v17, %%v29,%%v18      \n\t"
             "vsel    %%v19,%%v28,%%v19,%%v17 \n\t"
             "vsel    %%v18,%%v29,%%v18,%%v17 \n\t"
-
-            "VAG %%v5,%%v5,%%v4 \n\t"
-
-            "clgrjl %1,%%r0,1b  \n\t"
-
+            "vag %%v5,%%v5,%%v4 \n\t"
+            "clgrjl %2,%%r0,1b  \n\t"
 
             "vrepg %%v26,%%v18,1     \n\t"
             "vrepg %%v5,%%v19,1      \n\t"
             "wfcdb  %%v26,%%v18      \n\t"
             "jne 2f  \n\t"
-            "VSTEG   %%v18,0(%2),0   \n\t"
-            "VMNLG   %%v1,%%v5,%%v19 \n\t"
-            "VLGVG   %%r2,%%v1,0     \n\t"
-            "br %%r14  \n\t"
+            "vsteg   %%v18,%1,0   \n\t"
+            "vmnlg   %%v1,%%v5,%%v19 \n\t"
+            "vlgvg   %0,%%v1,0     \n\t"
+            "j 3f  \n\t"
             "2: \n\t"
             "wfchdb  %%v16,%%v26,%%v18      \n\t"
             "vsel    %%v1,%%v5,%%v19,%%v16  \n\t"
             "vsel    %%v0,%%v26,%%v18,%%v16 \n\t"
-            "VLGVG  %%r2,%%v1,0  \n\t"
-            "std %%f0,0(%2)    \n\t"
-
-            :
-            : "r"(n), "a"(x), "a"(maxf)
-            : "cc", "memory","r0","r1","r2","f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            "vlgvg  %0,%%v1,0  \n\t"
+            "std %%f0,%1    \n\t"
+            "3: "
+            : "=r"(index) ,"=m"(*maxf) , "+&a"(x)
+            : "r"(n), "2"(x) 
+            : "cc", "r0", "f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
             "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
-
             );
+    
+    return index;
 
 }
+
+
+ 
  
 BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
     BLASLONG i = 0;

diff --git a/kernel/zarch/idamin.c b/kernel/zarch/idamin.c
index 27dbe58..555864a 100644 (file)
--- a/kernel/zarch/idamin.c
+++ b/kernel/zarch/idamin.c
@@ -37,28 +37,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
-static BLASLONG __attribute__((noinline)) diamin_kernel_32(BLASLONG n, FLOAT *x, FLOAT *maxf) {
-
-    __asm__(
-            "pfd 1, 0(%1) \n\t"
-            "sllg %%r0,%0,3    \n\t"
-            "agr %%r0,%1  \n\t"
-            "VLEIG  %%v20,0,0  \n\t"
-            "VLEIG  %%v20,1,1  \n\t"
-            "VLEIG  %%v21,2,0  \n\t"
-            "VLEIG  %%v21,3,1  \n\t"
-            "VLEIG  %%v22,4,0  \n\t"
-            "VLEIG  %%v22,5,1  \n\t"
-            "VLEIG  %%v23,6,0  \n\t"
-            "VLEIG  %%v23,7,1  \n\t"
-            "VREPIG %%v4,8     \n\t"
-            "vzero %%v5        \n\t"
-            "vlrepg %%v18,0(%1)   \n\t"
+/**
+ * Find  minimum index 
+ * Warning: requirements n>0  and n % 32 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) minimum absolute value .( only for output )
+ * @return minimum index 
+ */
+static BLASLONG diamin_kernel_32(BLASLONG n, FLOAT *x, FLOAT *minf) {
+     BLASLONG index;
+    __asm__( 
+            "pfd 1, 0(%4) \n\t"
+            "sllg %%r0,%3,3    \n\t" 
+            "agr %%r0,%4  \n\t"
+            "vleig  %%v20,0,0  \n\t"
+            "vleig  %%v20,1,1  \n\t"
+            "vleig  %%v21,2,0  \n\t"
+            "vleig  %%v21,3,1  \n\t"
+            "vleig  %%v22,4,0  \n\t"
+            "vleig  %%v22,5,1  \n\t"
+            "vleig  %%v23,6,0  \n\t"
+            "vleig  %%v23,7,1  \n\t"
+            "vrepig %%v4,8     \n\t"
+            "vlrepg %%v18,0(%4)   \n\t"
+            "vzero %%v5        \n\t" 
+            "vflpdb  %%v18, %%v18 \n\t"
             "vzero %%v19          \n\t"
             ".align 16 \n\t"
             "1: \n\t"
-            "pfd 1, 256(%1 ) \n\t"
-            "vlm  %%v24,%%v31, 0(%1 ) \n\t"
+            "pfd 1, 256(%2 ) \n\t"
+            "vlm  %%v24,%%v31, 0(%2 ) \n\t"
 
             "vflpdb  %%v24, %%v24 \n\t"
             "vflpdb  %%v25, %%v25 \n\t"
@@ -90,12 +99,12 @@ static BLASLONG __attribute__((noinline)) diamin_kernel_32(BLASLONG n, FLOAT *x,
             "vsel    %%v24,%%v26,%%v24,%%v29 \n\t"
             "vsel    %%v25,%%v27,%%v25,%%v29 \n\t"
 
-            "VAG %%v1,%%v1,%%v5   \n\t"
-            "VAG %%v24,%%v24,%%v5   \n\t"
-            "VAG %%v24,%%v24,%%v4   \n\t"
+            "vag %%v1,%%v1,%%v5   \n\t"
+            "vag %%v24,%%v24,%%v5   \n\t"
+            "vag %%v24,%%v24,%%v4   \n\t"
 
             "vfchdb %%v16, %%v0,%%v25   \n\t"
-            "VAG %%v5,%%v5,%%v4 \n\t"
+            "vag %%v5,%%v5,%%v4 \n\t"
             "vsel    %%v29,%%v25,%%v0,%%v16 \n\t"
             "vsel  %%v28,%%v24,%%v1,%%v16 \n\t"
 
@@ -103,9 +112,9 @@ static BLASLONG __attribute__((noinline)) diamin_kernel_32(BLASLONG n, FLOAT *x,
             "vsel  %%v19,%%v28,%%v19,%%v17 \n\t"
             "vsel    %%v18,%%v29,%%v18,%%v17 \n\t"
 
-            "VAG %%v5,%%v5,%%v4 \n\t"
+            "vag %%v5,%%v5,%%v4 \n\t"
 
-            "vlm  %%v24,%%v31,128(%1 ) \n\t"
+            "vlm  %%v24,%%v31,128(%2 ) \n\t"
             "vflpdb  %%v24, %%v24 \n\t"
             "vflpdb  %%v25, %%v25 \n\t"
             "vflpdb  %%v26, %%v26 \n\t"
@@ -136,13 +145,13 @@ static BLASLONG __attribute__((noinline)) diamin_kernel_32(BLASLONG n, FLOAT *x,
             "vsel    %%v24,%%v26,%%v24,%%v29 \n\t"
             "vsel    %%v25,%%v27,%%v25,%%v29 \n\t"
 
-            "VAG     %%v1,%%v1,%%v5     \n\t"
-            "VAG     %%v24,%%v24,%%v5   \n\t"
-            "la %1,256(%1)   \n\t"
-            "VAG     %%v24,%%v24,%%v4   \n\t"
+            "vag     %%v1,%%v1,%%v5     \n\t"
+            "vag     %%v24,%%v24,%%v5   \n\t"
+            "la %2,256(%2)   \n\t"
+            "vag     %%v24,%%v24,%%v4   \n\t"
 
             "vfchdb  %%v16, %%v0,%%v25      \n\t"
-            "VAG     %%v5,%%v5,%%v4         \n\t"
+            "vag     %%v5,%%v5,%%v4         \n\t"
             "vsel    %%v29,%%v25,%%v0,%%v16 \n\t"
             "vsel    %%v28,%%v24,%%v1,%%v16 \n\t"
 
@@ -150,51 +159,55 @@ static BLASLONG __attribute__((noinline)) diamin_kernel_32(BLASLONG n, FLOAT *x,
             "vsel    %%v19,%%v28,%%v19,%%v17 \n\t"
             "vsel    %%v18,%%v29,%%v18,%%v17 \n\t"
 
-            "VAG     %%v5,%%v5,%%v4 \n\t"
+            "vag     %%v5,%%v5,%%v4 \n\t"
 
-            "clgrjl %1,%%r0,1b \n\t"
+            "clgrjl %2,%%r0,1b \n\t"
 
 
             "vrepg   %%v26,%%v18,1   \n\t"
             "vrepg   %%v5,%%v19,1    \n\t"
             "wfcdb   %%v26,%%v18     \n\t"
             "jne 2f  \n\t"
-            "VSTEG   %%v18,0(%2),0  \n\t"
-            "VMNLG   %%v1,%%v5,%%v19 \n\t"
-            "VLGVG   %%r2,%%v1,0  \n\t"
-            "br %%r14  \n\t"
+            "vsteg   %%v18,%1,0  \n\t"
+            "vmnlg   %%v1,%%v5,%%v19 \n\t"
+            "vlgvg   %0,%%v1,0  \n\t"
+            "j 3f  \n\t"
             "2: \n\t"
             "wfchdb  %%v16,%%v18 ,%%v26     \n\t "
             "vsel    %%v1,%%v5,%%v19,%%v16  \n\t"
             "vsel    %%v0,%%v26,%%v18,%%v16 \n\t"
-            "VLGVG   %%r2,%%v1,0  \n\t"
-            "std %%f0,0(%2)    \n\t"
+            "vlgvg   %0,%%v1,0  \n\t"
+            "std %%f0,%1    \n\t"
+             "3:"
 
-            :
-            : "r"(n), "a"(x), "a"(maxf)
-            : "cc", "memory","r0","r1","r2","f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            : "+r"(index) ,"=m"(*minf),"+&a"(x)
+            : "r"(n), "2"(x) 
+            : "cc","r0", "f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
             "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
 
             );
+    
+    return index;
 
 }
 
+
+
 BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
     BLASLONG i = 0;
     BLASLONG j = 0;
     BLASLONG ix = 0;
-    FLOAT minf = 0.0;
     BLASLONG min = 0;
-
+    FLOAT minf = 0.0;
+    
     if (n <= 0 || inc_x <= 0) return (min);
-
+    minf = ABS(x[0]); //index's not incremented,though it will make first comparision redundant
     if (inc_x == 1) {
 
         BLASLONG n1 = n & -32;
         if (n1 > 0) {
 
             min = diamin_kernel_32(n1, x, &minf);
-
             i = n1;
         }
 

diff --git a/kernel/zarch/izamax.c b/kernel/zarch/izamax.c
index 3966dfa..53b0400 100644 (file)
--- a/kernel/zarch/izamax.c
+++ b/kernel/zarch/izamax.c
@@ -32,56 +32,65 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ABS fabs 
 #define CABS1(x,i)    ABS(x[i])+ABS(x[i+1])
 
- 
 
-static BLASLONG __attribute__((noinline)) ziamax_kernel_8_TUNED(BLASLONG n, FLOAT *x, FLOAT *maxf) { 
 
+ 
+/**
+ * Find  maximum index 
+ * Warning: requirements n>0  and n % 8 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) maximum absolute value .( only for output )
+ * @return  index 
+ */
+static BLASLONG   ziamax_kernel_8_TUNED(BLASLONG n, FLOAT *x, FLOAT *maxf) { 
+    BLASLONG index;
     __asm__(
-            "pfd 1, 0(%1) \n\t" 
-            "VLEIG  %%v16,0,0  \n\t"
-            "VLEIG  %%v16,1,1  \n\t"
-            "VLEIG  %%v17,2,0  \n\t"
-            "VLEIG  %%v17,3,1  \n\t"
-            "VLEIG  %%v18,4,0  \n\t"
-            "VLEIG  %%v18,5,1  \n\t"
-            "VLEIG  %%v19,6,0  \n\t"
-            "VLEIG  %%v19,7,1  \n\t" 
-            "VLEIG  %%v20,8,0  \n\t"
-            "VLEIG  %%v20,9,1  \n\t"
-            "VLEIG  %%v21,10,0 \n\t"
-            "VLEIG  %%v21,11,1 \n\t"
-            "VLEIG  %%v22,12,0 \n\t"
-            "VLEIG  %%v22,13,1 \n\t"
-            "VLEIG  %%v23,14,0 \n\t"
-            "VLEIG  %%v23,15,1 \n\t" 
+            "pfd 1, 0(%4) \n\t" 
+            "vleig  %%v16,0,0  \n\t"
+            "vleig  %%v16,1,1  \n\t"
+            "vleig  %%v17,2,0  \n\t"
+            "vleig  %%v17,3,1  \n\t"
+            "vleig  %%v18,4,0  \n\t"
+            "vleig  %%v18,5,1  \n\t"
+            "vleig  %%v19,6,0  \n\t"
+            "vleig  %%v19,7,1  \n\t" 
+            "vleig  %%v20,8,0  \n\t"
+            "vleig  %%v20,9,1  \n\t"
+            "vleig  %%v21,10,0 \n\t"
+            "vleig  %%v21,11,1 \n\t"
+            "vleig  %%v22,12,0 \n\t"
+            "vleig  %%v22,13,1 \n\t"
+            "vleig  %%v23,14,0 \n\t"
+            "vleig  %%v23,15,1 \n\t" 
     
     
-            "sllg %%r0,%0,4 \n\t"
-            "agr %%r0,%1    \n\t" 
+            "sllg %%r0,%3,4 \n\t"
+            "agr %%r0,%4    \n\t" 
             "vzero %%v6     \n\t"
             "vzero %%v7     \n\t"
-            "VREPIG %%v4,16 \n\t"
+            "vrepig %%v4,16 \n\t"
             "vzero %%v5     \n\t"
             ".align 16 \n\t"
             "1: \n\t"
-            "pfd 1, 256(%1 ) \n\t"
+            "pfd 1, 256(%2 ) \n\t"
         
-            "vleg %%v24 ,  0( %1),0 \n\t" 
-            "vleg %%v25 ,  8( %1),0 \n\t"
-            "vleg %%v24 , 16( %1),1 \n\t" 
-            "vleg %%v25 , 24( %1),1 \n\t"
-            "vleg %%v26 , 32( %1),0 \n\t"  
-            "vleg %%v27 , 40( %1),0 \n\t"
-            "vleg %%v26 , 48( %1),1 \n\t" 
-            "vleg %%v27 , 56( %1),1 \n\t" 
-            "vleg %%v28 , 64( %1),0 \n\t" 
-            "vleg %%v29 , 72( %1),0 \n\t" 
-            "vleg %%v28 , 80( %1),1 \n\t"
-            "vleg %%v29 , 88( %1),1 \n\t" 
-            "vleg %%v30 , 96( %1),0 \n\t"  
-            "vleg %%v31 ,104( %1),0 \n\t"
-            "vleg %%v30 ,112( %1),1 \n\t"
-            "vleg %%v31 ,120( %1),1 \n\t"  
+            "vleg %%v24 ,  0( %2),0 \n\t" 
+            "vleg %%v25 ,  8( %2),0 \n\t"
+            "vleg %%v24 , 16( %2),1 \n\t" 
+            "vleg %%v25 , 24( %2),1 \n\t"
+            "vleg %%v26 , 32( %2),0 \n\t"  
+            "vleg %%v27 , 40( %2),0 \n\t"
+            "vleg %%v26 , 48( %2),1 \n\t" 
+            "vleg %%v27 , 56( %2),1 \n\t" 
+            "vleg %%v28 , 64( %2),0 \n\t" 
+            "vleg %%v29 , 72( %2),0 \n\t" 
+            "vleg %%v28 , 80( %2),1 \n\t"
+            "vleg %%v29 , 88( %2),1 \n\t" 
+            "vleg %%v30 , 96( %2),0 \n\t"  
+            "vleg %%v31 ,104( %2),0 \n\t"
+            "vleg %%v30 ,112( %2),1 \n\t"
+            "vleg %%v31 ,120( %2),1 \n\t"  
             "vflpdb  %%v24, %%v24   \n\t" 
             "vflpdb  %%v25, %%v25   \n\t" 
             "vflpdb  %%v26, %%v26   \n\t" 
@@ -97,22 +106,22 @@ static BLASLONG __attribute__((noinline)) ziamax_kernel_8_TUNED(BLASLONG n, FLOA
             "vfadb %%v3,%%v30,%%v31 \n\t"
      
     
-            "vleg %%v24 , 128( %1),0  \n\t" 
-            "vleg %%v25 , 136( %1),0  \n\t"
-            "vleg %%v24 , 144( %1),1  \n\t" 
-            "vleg %%v25 , 152( %1),1  \n\t"
-            "vleg %%v26 , 160( %1),0  \n\t"  
-            "vleg %%v27 , 168( %1),0  \n\t"
-            "vleg %%v26 , 176( %1),1  \n\t" 
-            "vleg %%v27 , 184( %1),1  \n\t" 
-            "vleg %%v28 , 192( %1),0  \n\t" 
-            "vleg %%v29 , 200( %1),0  \n\t" 
-            "vleg %%v28 , 208( %1),1  \n\t"
-            "vleg %%v29 , 216( %1),1  \n\t" 
-            "vleg %%v30 , 224( %1),0  \n\t"  
-            "vleg %%v31 , 232( %1),0  \n\t"
-            "vleg %%v30 , 240( %1),1  \n\t"
-            "vleg %%v31 , 248( %1),1  \n\t"  
+            "vleg %%v24 , 128( %2),0  \n\t" 
+            "vleg %%v25 , 136( %2),0  \n\t"
+            "vleg %%v24 , 144( %2),1  \n\t" 
+            "vleg %%v25 , 152( %2),1  \n\t"
+            "vleg %%v26 , 160( %2),0  \n\t"  
+            "vleg %%v27 , 168( %2),0  \n\t"
+            "vleg %%v26 , 176( %2),1  \n\t" 
+            "vleg %%v27 , 184( %2),1  \n\t" 
+            "vleg %%v28 , 192( %2),0  \n\t" 
+            "vleg %%v29 , 200( %2),0  \n\t" 
+            "vleg %%v28 , 208( %2),1  \n\t"
+            "vleg %%v29 , 216( %2),1  \n\t" 
+            "vleg %%v30 , 224( %2),0  \n\t"  
+            "vleg %%v31 , 232( %2),0  \n\t"
+            "vleg %%v30 , 240( %2),1  \n\t"
+            "vleg %%v31 , 248( %2),1  \n\t"  
             "vflpdb  %%v24, %%v24 \n\t" 
             "vflpdb  %%v25, %%v25 \n\t" 
             "vflpdb  %%v26, %%v26 \n\t" 
@@ -151,49 +160,49 @@ static BLASLONG __attribute__((noinline)) ziamax_kernel_8_TUNED(BLASLONG n, FLOA
             "vsel  %%v29,%%v25,%%v2,%%v30  \n\t" 
             "vsel  %%v31,%%v27,%%v3 ,%%v30 \n\t" 
     
-            "la %1,256(%1)      \n\t"  
+            "la %2,256(%2)      \n\t"  
            
             "vfchdb %%v0, %%v31,%%v28     \n\t" 
             "vsel  %%v25,%%v29,%%v26,%%v0 \n\t" 
             "vsel  %%v27,%%v31,%%v28,%%v0 \n\t"
     
-            "VAG %%v25,%%v25,%%v5 \n\t"
+            "vag %%v25,%%v25,%%v5 \n\t"
     
              //cmp with previous
             "vfchdb %%v30, %%v27,%%v6    \n\t"
             "vsel  %%v7,%%v25,%%v7,%%v30 \n\t" 
             "vsel  %%v6,%%v27,%%v6,%%v30 \n\t"
     
-            "VAG %%v5,%%v5,%%v4 \n\t" 
+            "vag %%v5,%%v5,%%v4 \n\t" 
     
-            "clgrjl %1,%%r0,1b \n\t"
+            "clgrjl %2,%%r0,1b \n\t"
 
             //xtract index
             "vrepg %%v26,%%v6,1      \n\t"
             "vrepg %%v5,%%v7,1       \n\t"
             "wfcdb  %%v26,%%v6       \n\t"
             "jne 2f  \n\t"
-            "VSTEG   %%v6,0(%2),0  \n\t"
-            "VMNLG  %%v1,%%v5,%%v7 \n\t"
-            "VLGVG  %%r2,%%v1,0  \n\t"
-            "br %%r14  \n\t"
+            "vsteg   %%v6,%1,0  \n\t"
+            "vmnlg  %%v1,%%v5,%%v7 \n\t"
+            "vlgvg  %0,%%v1,0  \n\t"
+            "j 3  \n\t"
             "2: \n\t"
             "wfchdb  %%v16,%%v26,%%v6      \n\t"
             "vsel    %%v1,%%v5,%%v7,%%v16  \n\t"
             "vsel    %%v0,%%v26,%%v6,%%v16 \n\t"
-            "VLGVG  %%r2,%%v1,0  \n\t"
-            "std %%f0,0(%2)    \n\t"
-
-            :
-            : "r"(n), "a"(x), "a"(maxf)
-            : "cc", "memory","r0","r1","r2","f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            "vlgvg  %0,%%v1,0  \n\t"
+            "std %%f0,%1    \n\t"
+            "3: \n\t"
+            : "=r"(index),"=m"(*maxf),"+&a"(x)
+            : "r"(n), "2"(x)
+            : "cc","r0", "f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
             "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
 
             );
+    return index;
 
 }
 
-
   
 
  

diff --git a/kernel/zarch/izamin.c b/kernel/zarch/izamin.c
index 4af1410..77ecf67 100644 (file)
--- a/kernel/zarch/izamin.c
+++ b/kernel/zarch/izamin.c
@@ -33,58 +33,65 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CABS1(x,i)    ABS(x[i])+ABS(x[i+1])
 
  
-
-static BLASLONG __attribute__((noinline)) ziamin_kernel_8_TUNED(BLASLONG n, FLOAT *x, FLOAT *minf) { 
-
+/**
+ * Find  minimum index 
+ * Warning: requirements n>0  and n % 8 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) minimum absolute value .( only for output )
+ * @return minimum index 
+ */
+static BLASLONG ziamin_kernel_8_TUNED(BLASLONG n, FLOAT *x, FLOAT *minf) { 
+    BLASLONG index ;
     __asm__(
-            "pfd 1, 0(%1) \n\t" 
-            "VLEIG  %%v16,0,0  \n\t"
-            "VLEIG  %%v16,1,1  \n\t"
-            "VLEIG  %%v17,2,0  \n\t"
-            "VLEIG  %%v17,3,1  \n\t"
-            "VLEIG  %%v18,4,0  \n\t"
-            "VLEIG  %%v18,5,1  \n\t"
-            "VLEIG  %%v19,6,0  \n\t"
-            "VLEIG  %%v19,7,1  \n\t" 
-            "VLEIG  %%v20,8,0  \n\t"
-            "VLEIG  %%v20,9,1  \n\t"
-            "VLEIG  %%v21,10,0 \n\t"
-            "VLEIG  %%v21,11,1 \n\t"
-            "VLEIG  %%v22,12,0 \n\t"
-            "VLEIG  %%v22,13,1 \n\t"
-            "VLEIG  %%v23,14,0 \n\t"
-            "VLEIG  %%v23,15,1 \n\t" 
-            "ld %%f6,0(%1)     \n\t"
+            "pfd 1, 0(%4) \n\t" 
+            "vleig  %%v16,0,0  \n\t"
+            "vleig  %%v16,1,1  \n\t"
+            "vleig  %%v17,2,0  \n\t"
+            "vleig  %%v17,3,1  \n\t"
+            "vleig  %%v18,4,0  \n\t"
+            "vleig  %%v18,5,1  \n\t"
+            "vleig  %%v19,6,0  \n\t"
+            "vleig  %%v19,7,1  \n\t" 
+            "vleig  %%v20,8,0  \n\t"
+            "vleig  %%v20,9,1  \n\t"
+            "vleig  %%v21,10,0 \n\t"
+            "vleig  %%v21,11,1 \n\t"
+            "vleig  %%v22,12,0 \n\t"
+            "vleig  %%v22,13,1 \n\t"
+            "vleig  %%v23,14,0 \n\t"
+            "vleig  %%v23,15,1 \n\t" 
+            "ld %%f6,0(%4)     \n\t"
             "lpdbr %%f6,%%f6   \n\t" 
-            "ld %%f7,8(%1)     \n\t"
+            "ld %%f7,8(%4)     \n\t"
             "lpdbr %%f7,%%f7   \n\t"   
             "adbr %%f6,%%f7    \n\t"      
-            "sllg %%r0,%0,4    \n\t"
-            "agr %%r0,%1       \n\t" 
+            "sllg %%r0,%3,4    \n\t"
+            "agr %%r0,%4       \n\t" 
             "vrepg %%v6,%%v6,0 \n\t"
             "vzero %%v7        \n\t"
-            "VREPIG %%v4,16    \n\t"
+            "vrepig %%v4,16    \n\t"
             "vzero %%v5        \n\t"
             ".align 16 \n\t"
             "1: \n\t"
-            "pfd 1, 256(%1 ) \n\t"
+            "pfd 1, 256(%2 ) \n\t"
         
-            "vleg %%v24 ,  0( %1),0 \n\t" 
-            "vleg %%v25 ,  8( %1),0 \n\t"
-            "vleg %%v24 , 16( %1),1 \n\t" 
-            "vleg %%v25 , 24( %1),1 \n\t"
-            "vleg %%v26 , 32( %1),0 \n\t"  
-            "vleg %%v27 , 40( %1),0 \n\t"
-            "vleg %%v26 , 48( %1),1 \n\t" 
-            "vleg %%v27 , 56( %1),1 \n\t" 
-            "vleg %%v28 , 64( %1),0 \n\t" 
-            "vleg %%v29 , 72( %1),0 \n\t" 
-            "vleg %%v28 , 80( %1),1 \n\t"
-            "vleg %%v29 , 88( %1),1 \n\t" 
-            "vleg %%v30 , 96( %1),0 \n\t"  
-            "vleg %%v31 ,104( %1),0 \n\t"
-            "vleg %%v30 ,112( %1),1 \n\t"
-            "vleg %%v31 ,120( %1),1 \n\t"  
+            "vleg %%v24 ,  0( %2),0 \n\t" 
+            "vleg %%v25 ,  8( %2),0 \n\t"
+            "vleg %%v24 , 16( %2),1 \n\t" 
+            "vleg %%v25 , 24( %2),1 \n\t"
+            "vleg %%v26 , 32( %2),0 \n\t"  
+            "vleg %%v27 , 40( %2),0 \n\t"
+            "vleg %%v26 , 48( %2),1 \n\t" 
+            "vleg %%v27 , 56( %2),1 \n\t" 
+            "vleg %%v28 , 64( %2),0 \n\t" 
+            "vleg %%v29 , 72( %2),0 \n\t" 
+            "vleg %%v28 , 80( %2),1 \n\t"
+            "vleg %%v29 , 88( %2),1 \n\t" 
+            "vleg %%v30 , 96( %2),0 \n\t"  
+            "vleg %%v31 ,104( %2),0 \n\t"
+            "vleg %%v30 ,112( %2),1 \n\t"
+            "vleg %%v31 ,120( %2),1 \n\t"  
             "vflpdb  %%v24, %%v24   \n\t" 
             "vflpdb  %%v25, %%v25   \n\t" 
             "vflpdb  %%v26, %%v26   \n\t" 
@@ -100,22 +107,22 @@ static BLASLONG __attribute__((noinline)) ziamin_kernel_8_TUNED(BLASLONG n, FLOA
             "vfadb %%v3,%%v30,%%v31 \n\t"
      
     
-            "vleg %%v24 ,128( %1),0 \n\t" 
-            "vleg %%v25 ,136( %1),0 \n\t"
-            "vleg %%v24 ,144( %1),1 \n\t" 
-            "vleg %%v25 ,152( %1),1 \n\t"
-            "vleg %%v26 ,160( %1),0 \n\t"  
-            "vleg %%v27 ,168( %1),0 \n\t"
-            "vleg %%v26 ,176( %1),1 \n\t" 
-            "vleg %%v27 ,184( %1),1 \n\t" 
-            "vleg %%v28 ,192( %1),0 \n\t" 
-            "vleg %%v29 ,200( %1),0 \n\t" 
-            "vleg %%v28 ,208( %1),1 \n\t"
-            "vleg %%v29 ,216( %1),1 \n\t" 
-            "vleg %%v30 ,224( %1),0 \n\t"  
-            "vleg %%v31 ,232( %1),0 \n\t"
-            "vleg %%v30 ,240( %1),1 \n\t"
-            "vleg %%v31 ,248( %1),1 \n\t"  
+            "vleg %%v24 ,128( %2),0 \n\t" 
+            "vleg %%v25 ,136( %2),0 \n\t"
+            "vleg %%v24 ,144( %2),1 \n\t" 
+            "vleg %%v25 ,152( %2),1 \n\t"
+            "vleg %%v26 ,160( %2),0 \n\t"  
+            "vleg %%v27 ,168( %2),0 \n\t"
+            "vleg %%v26 ,176( %2),1 \n\t" 
+            "vleg %%v27 ,184( %2),1 \n\t" 
+            "vleg %%v28 ,192( %2),0 \n\t" 
+            "vleg %%v29 ,200( %2),0 \n\t" 
+            "vleg %%v28 ,208( %2),1 \n\t"
+            "vleg %%v29 ,216( %2),1 \n\t" 
+            "vleg %%v30 ,224( %2),0 \n\t"  
+            "vleg %%v31 ,232( %2),0 \n\t"
+            "vleg %%v30 ,240( %2),1 \n\t"
+            "vleg %%v31 ,248( %2),1 \n\t"  
             "vflpdb  %%v24, %%v24   \n\t" 
             "vflpdb  %%v25, %%v25   \n\t" 
             "vflpdb  %%v26, %%v26   \n\t" 
@@ -155,49 +162,53 @@ static BLASLONG __attribute__((noinline)) ziamin_kernel_8_TUNED(BLASLONG n, FLOA
             "vsel  %%v29,%%v25,%%v2,%%v30  \n\t" 
             "vsel  %%v31,%%v27,%%v3 ,%%v30 \n\t" 
     
-            "la %1,256(%1)      \n\t"  
+            "la %2,256(%2)      \n\t"  
            
             "vfchdb %%v0,%%v28, %%v31      \n\t" 
             "vsel  %%v25,%%v29,%%v26,%%v0  \n\t" 
             "vsel  %%v27,%%v31,%%v28,%%v0  \n\t"
     
-            "VAG %%v25,%%v25,%%v5 \n\t"
+            "vag %%v25,%%v25,%%v5 \n\t"
     
              //cmp with previous
             "vfchdb %%v30,%%v6 , %%v27     \n\t"
             "vsel  %%v7,%%v25,%%v7,%%v30   \n\t" 
             "vsel  %%v6,%%v27,%%v6,%%v30   \n\t"
     
-            "VAG %%v5,%%v5,%%v4 \n\t" 
+            "vag %%v5,%%v5,%%v4 \n\t" 
     
-            "clgrjl %1,%%r0,1b  \n\t"
+            "clgrjl %2,%%r0,1b  \n\t"
 
             //xtract index
             "vrepg %%v26,%%v6,1      \n\t"
             "vrepg %%v5,%%v7,1       \n\t"
             "wfcdb  %%v26,%%v6       \n\t"
             "jne 2f  \n\t"
-            "VSTEG   %%v6,0(%2),0    \n\t"
-            "VMNLG  %%v1,%%v5,%%v7   \n\t"
-            "VLGVG  %%r2,%%v1,0      \n\t"
-            "br %%r14  \n\t"
+            "vsteg   %%v6,%1,0    \n\t"
+            "vmnlg  %%v1,%%v5,%%v7   \n\t"
+            "vlgvg  %0,%%v1,0      \n\t"
+            "j 3f  \n\t"
             "2: \n\t"
             "wfchdb  %%v16,%%v6 ,%%v26     \n\t"
             "vsel    %%v1,%%v5,%%v7,%%v16  \n\t"
             "vsel    %%v0,%%v26,%%v6,%%v16 \n\t"
-            "VLGVG   %%r2,%%v1,0  \n\t"
-            "std %%f0,0(%2)       \n\t"
+            "vlgvg   %0,%%v1,0  \n\t"
+            "std %%f0,%1       \n\t"
+            "3: \n\t"
 
-            :
-            : "r"(n), "a"(x), "a"(minf)
-            : "cc", "memory","r0","r1","r2","f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            : "+r"(index) ,"=m"(*minf), "+&a"(x)
+            : "r"(n), "2"(x) 
+            : "cc","r0","f0","v0","v1","v2","v3","v4","v5","v6","v7","v16",
             "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
 
             );
 
+    return index;
 }
 
  
+
+ 
  
 
 BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
@@ -220,6 +231,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 
             i = n1;
         }
+        else {
+            //assign minf
+             minf = CABS1(x,0);
+             ix += 2;
+             i++;
+         }
 
         while(i < n)
         {

diff --git a/kernel/zarch/zasum.c b/kernel/zarch/zasum.c
index a9504f2..e01f465 100644 (file)
--- a/kernel/zarch/zasum.c
+++ b/kernel/zarch/zasum.c
@@ -40,12 +40,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 
-static FLOAT __attribute__ ((noinline)) zasum_kernel_16(BLASLONG n, FLOAT *x) {
-
+static FLOAT zasum_kernel_16(BLASLONG n, FLOAT *x) {
+    
+    FLOAT asum;
     __asm__ (
-            "pfd 1, 0(%1) \n\t"
-            "sllg  %%r0,%0,4 \n\t"
-            "agr   %%r0,%1   \n\t"   
+            "pfd 1, 0(%3) \n\t"
+            "sllg  %%r0,%2,4 \n\t"
+            "agr   %%r0,%3   \n\t"   
             "vzero %%v0      \n\t"
             "vzero %%v1      \n\t"
             "vzero %%v22     \n\t"
@@ -99,10 +100,12 @@ static FLOAT __attribute__ ((noinline)) zasum_kernel_16(BLASLONG n, FLOAT *x) {
             "vfadb   %%v0,%%v25,%%v24  \n\t"
             "vrepg %%v1,%%v0,1         \n\t"
             "adbr %%f0,%%f1    \n\t"
-            :
-            : "r"(n), "a"(x)
-            : "cc", "memory","r0","f0","f1","v0","v1","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
+            "ldr %0 ,%%f0"
+            : "=f"(asum),"+&a"(x)
+            : "r"(n), "1"(x)
+            : "cc",  "r0","f0","f1","v0","v1","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
             );
+    return asum;
 
 }
 

diff --git a/kernel/zarch/zdot.c b/kernel/zarch/zdot.c
index 478cf25..cfa4549 100644 (file)
--- a/kernel/zarch/zdot.c
+++ b/kernel/zarch/zdot.c
@@ -27,30 +27,31 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #include "common.h"
+#if defined(Z13)
 
-static void __attribute__ ((noinline)) zdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d) {
+static void zdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d) {
 
     __asm__ volatile(
-            "pfd 1, 0(%1) \n\t"
             "pfd 1, 0(%2) \n\t"
+            "pfd 1, 0(%3) \n\t"
             "vzero %%v24  \n\t"
             "vzero %%v25  \n\t"
             "vzero %%v26  \n\t"
             "vzero %%v27  \n\t"
-            "srlg %%r0,%0,3      \n\t"
+            "srlg %1,%1,3      \n\t"
             "xgr %%r1,%%r1       \n\t"
             ".align 16 \n\t"
             "1: \n\t"
-            "pfd 1, 256(%%r1,%1)     \n\t"
             "pfd 1, 256(%%r1,%2)     \n\t"
-            "vl  %%v16,  0(%%r1,%1)  \n\t"
-            "vl  %%v17, 16(%%r1,%1)  \n\t"
-            "vl  %%v18, 32(%%r1,%1)  \n\t"
-            "vl  %%v19, 48(%%r1,%1)  \n\t"
-            "vl  %%v28,  0(%%r1,%2)  \n\t"
-            "vl  %%v29, 16(%%r1,%2)  \n\t"
-            "vl  %%v30, 32(%%r1,%2)  \n\t"
-            "vl  %%v31, 48(%%r1,%2)  \n\t"
+            "pfd 1, 256(%%r1,%3)     \n\t"
+            "vl  %%v16,  0(%%r1,%2)  \n\t"
+            "vl  %%v17, 16(%%r1,%2)  \n\t"
+            "vl  %%v18, 32(%%r1,%2)  \n\t"
+            "vl  %%v19, 48(%%r1,%2)  \n\t"
+            "vl  %%v28,  0(%%r1,%3)  \n\t"
+            "vl  %%v29, 16(%%r1,%3)  \n\t"
+            "vl  %%v30, 32(%%r1,%3)  \n\t"
+            "vl  %%v31, 48(%%r1,%3)  \n\t"
             "vpdi %%v20,%%v16,%%v16,4 \n\t"
             "vpdi %%v21,%%v17,%%v17,4 \n\t"
             "vpdi %%v22,%%v18,%%v18,4 \n\t"
@@ -68,14 +69,14 @@ static void __attribute__ ((noinline)) zdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT
 
 
 
-            "vl  %%v16, 64(%%r1,%1) \n\t"
-            "vl  %%v17, 80(%%r1,%1) \n\t"
-            "vl  %%v18, 96(%%r1,%1) \n\t"
-            "vl  %%v19,112(%%r1,%1) \n\t"
-            "vl  %%v28, 64(%%r1,%2) \n\t"
-            "vl  %%v29, 80(%%r1,%2) \n\t"
-            "vl  %%v30, 96(%%r1,%2) \n\t"
-            "vl  %%v31,112(%%r1,%2) \n\t"
+            "vl  %%v16, 64(%%r1,%2) \n\t"
+            "vl  %%v17, 80(%%r1,%2) \n\t"
+            "vl  %%v18, 96(%%r1,%2) \n\t"
+            "vl  %%v19,112(%%r1,%2) \n\t"
+            "vl  %%v28, 64(%%r1,%3) \n\t"
+            "vl  %%v29, 80(%%r1,%3) \n\t"
+            "vl  %%v30, 96(%%r1,%3) \n\t"
+            "vl  %%v31,112(%%r1,%3) \n\t"
             "vpdi %%v20,%%v16,%%v16,4 \n\t"
             "vpdi %%v21,%%v17,%%v17,4 \n\t"
             "vpdi %%v22,%%v18,%%v18,4 \n\t"
@@ -91,22 +92,24 @@ static void __attribute__ ((noinline)) zdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT
 
 
             "la %%r1,128(%%r1) \n\t"
-            "brctg %%r0,1b     \n\t"
+            "brctg %1,1b     \n\t"
             "vfadb %%v24,%%v26,%%v24 \n\t"
             "vfadb %%v25,%%v25,%%v27 \n\t"
-            "vsteg %%v24,0(%3),0     \n\t"
-            "vsteg %%v24,8(%3),1     \n\t"
-            "vsteg %%v25,16(%3),1    \n\t"
-            "vsteg %%v25,24(%3),0    \n\t"
-            :
-            : "r"(n), "a"(x), "a"(y), "a"(d)
-            : "cc", "memory","r0","r1","v16",
+            "vsteg %%v24,0(%4),0     \n\t"
+            "vsteg %%v24,8(%4),1     \n\t"
+            "vsteg %%v25,16(%4),1    \n\t"
+            "vsteg %%v25,24(%4),0    \n\t"
+            : "=m"(*d) ,"+&r"(n) 
+            :  "a"(x), "a"(y), "a"(d)
+            : "cc", "r1","v16",
             "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31" 
             );
 
 }
 
-static __attribute__ ((noinline)) void zdot_kernel_8n(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d) {
+#else
+
+static  void zdot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d) {
     BLASLONG register i = 0;
     FLOAT dot[4] = {0.0, 0.0, 0.0, 0.0};
     BLASLONG j = 0;
@@ -144,6 +147,8 @@ static __attribute__ ((noinline)) void zdot_kernel_8n(BLASLONG n, FLOAT *x, FLOA
 
 }
 
+#endif
+
 OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) {
     BLASLONG i;
     BLASLONG ix, iy;