Optimize gemv_n_sve kernel

Author: manaalmj

kernel/arm64/gemv_n_sve.c

@@ -58,6 +58,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
   ix = 0;
   a_ptr = a;
   if (inc_y == 1) {
+    BLASLONG width = n / 3;
     BLASLONG width = n / 3;
     uint64_t sve_size = SV_COUNT();
     svbool_t pg_true = SV_TRUE();
@@ -68,8 +69,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
     FLOAT *a2_ptr = a + lda * width * 2;
 
     for (j = 0; j < width; j++) {
-      i = 0;
-      while ((i + sve_size * 1 - 1) < m) {
+
+      for (i = 0; (i + sve_size - 1) < m; i += sve_size) {
         ix = j * inc_x;
 
         SV_TYPE x0_vec = SV_DUP(alpha * x[ix + (inc_x * width * 0)]);
@@ -86,8 +87,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
         y_vec = svmla_lane(y_vec, a02_vec, x2_vec, 0);
 
         svst1(pg_true, y + i, y_vec);
-
-        i += sve_size * 1;
       }
 
       if (i < m) {
@@ -117,29 +116,31 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
     a_ptr = a2_ptr;
     for (j = width * 3; j < n; j++) {
       ix = j * inc_x;
-      i = 0;
-      while ((i + sve_size * 1 - 1) < m) {
+      for (i = 0; (i + sve_size - 1) < m; i += sve_size) {
         SV_TYPE y_vec = svld1(pg_true, y + i);
         SV_TYPE x_vec = SV_DUP(alpha * x[(ix)]);
         SV_TYPE a_vec = svld1(pg_true, a_ptr + i);
         y_vec = svmla_x(pg_true, y_vec, a_vec, x_vec);
         svst1(pg_true, y + i, y_vec);
-        i += sve_size * 1;
       }
 
       if (i < m) {
         SV_TYPE y_vec = svld1(pg, y + i);
         SV_TYPE x_vec = SV_DUP(alpha * x[(ix)]);
         SV_TYPE a_vec = svld1(pg, a_ptr + i);
         y_vec = svmla_m(pg, y_vec, a_vec, x_vec);
+        y_vec = svmla_m(pg, y_vec, a_vec, x_vec);
         svst1(pg, y + i, y_vec);
       }
 
+
       a_ptr += lda;
       ix += inc_x;
     }
 
     return (0);
+
+    return (0);
   }
 
   for (j = 0; j < n; j++) {