zscal (case: real alpha=0 ) mikrokernel shift&mem fix , da_i as input reg. small typo fixes

kernel/zarch/dswap.c

@@ -139,7 +139,7 @@ static void   dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y)
 
 #else
 
-static void  __attribute__ ((noinline))  dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y)
+static void   dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y)
 {
          __asm__ volatile(
             "pfd 2, 0(%[ptr_x]) \n\t"
@@ -227,7 +227,7 @@ static void  __attribute__ ((noinline))  dswap_kernel_32(BLASLONG n, FLOAT *x, F
               [mem_y] "+m" (*(double (*)[n])y),
               [n_tmp] "+&r"(n)
             : [ptr_x] "a"(x), [ptr_y] "a"(y) 
-            : "cc", "memory","r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            : "cc", "r1", "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;

kernel/zarch/zscal.c

@@ -96,7 +96,7 @@ static void   zscal_kernel_8(BLASLONG n, FLOAT da_r,FLOAT da_i, FLOAT *x) {
             "clgrjl %[x_ptr],%%r0,1b        \n\t"
             : [mem] "+m" (*(double (*)[2*n])x) ,[x_ptr] "+&a"(x)
             : [n] "r"(n), [alpha_r] "f"(da_r),[alpha_i] "f"(da_i)
-            : "cc", "memory","r0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25"
+            : "cc", "r0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25"
             );
 
 
@@ -106,10 +106,9 @@ static void   zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
  
         __asm__ (   "pfd    2, 0(%1)          \n\t" 
                     "lgdr   %%r0,%[alpha]     \n\t"
-                    "vlvgg  %%v16,%%r0,0      \n\t" 
-                    "lcdbr  %[alpha],%[alpha] \n\t"                                       
-                    "lgdr   %%r0,%[alpha]     \n\t"
-                    "vlvgg  %%v16,%%r0,1      \n\t"
+                    "vlvgp  %%v16,%%r0,%%r0   \n\t" //load both from disjoint
+                    "vflcdb %%v16,%%v16       \n\t" //complement both
+                    "vlvgg  %%v16,%%r0,0      \n\t" //restore 1st                   
                     "vlr    %%v17 ,%%v16      \n\t" 
                     "sllg   %%r0,%[n],4       \n\t"  
                     "agr    %%r0,%[x_ptr]     \n\t"
@@ -129,8 +128,8 @@ static void   zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
                     "vsteg  %%v26, 40(%[x_ptr]),0   \n\t"   
                     "vl     %%v27, 48(%[x_ptr])     \n\t" 
                     "vfmdb  %%v27,%%v27,%%v17 \n\t"  
-                    "vsteg  %%v27, 40(%[x_ptr]),1   \n\t" 
-                    "vsteg  %%v27, 48(%[x_ptr]),0   \n\t" 
+                    "vsteg  %%v27, 48(%[x_ptr]),1   \n\t" 
+                    "vsteg  %%v27, 56(%[x_ptr]),0   \n\t" 
                     "vl     %%v28, 64(%[x_ptr])     \n\t"
                     "vfmdb  %%v28,%%v28,%%v16        \n\t"
                     "vsteg  %%v28, 64(%[x_ptr]),1   \n\t" 
@@ -141,8 +140,8 @@ static void   zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
                     "vsteg  %%v29, 88(%[x_ptr]),0   \n\t" 
                     "vl     %%v30, 96(%[x_ptr])     \n\t"
                     "vfmdb  %%v30,%%v30,%%v16       \n\t"
-                    "vsteg  %%v27,  96(%[x_ptr]),1  \n\t" 
-                    "vsteg  %%v27, 104(%[x_ptr]),0  \n\t"  
+                    "vsteg  %%v30,  96(%[x_ptr]),1  \n\t" 
+                    "vsteg  %%v30, 104(%[x_ptr]),0  \n\t"  
                     "vl     %%v31, 112(%[x_ptr])    \n\t" 
                     "vfmdb  %%v31,%%v31,%%v17 \n\t"  
                     "vsteg  %%v31, 112(%[x_ptr]),1  \n\t" 

kernel/zarch/zswap.c

@@ -137,7 +137,7 @@ static void   zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y)
 
 #else
 
-static void  __attribute__ ((noinline))  zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y)
+static void   zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y)
 {
          __asm__ volatile(
             "pfd 2, 0(%[ptr_x]) \n\t"
@@ -225,7 +225,7 @@ static void  __attribute__ ((noinline))  zswap_kernel_16(BLASLONG n, FLOAT *x, F
               [mem_y] "+m" (*(double (*)[2*n])y),
               [n_tmp] "+&r"(n)
             : [ptr_x] "a"(x), [ptr_y] "a"(y) 
-            : "cc", "memory", "r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16",
+            : "cc",  "r1", "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;