1

2026-03-06 22:37:52 +07:00
parent cb7863cf42
commit 8fb610e69c
407 changed files with 899 additions and 898 deletions
@@ -45,14 +45,14 @@ __kernel void MatrixMult_GPU2(__global double *matrix_a,
   int offset_a_start=cols_a*BLOCK_SIZE*group_j;
   double sum=(float)0.0;

+   __local double submatrix_a[BLOCK_SIZE][BLOCK_SIZE];
+   __local double submatrix_b[BLOCK_SIZE][BLOCK_SIZE];
+
   for(int offset_a=offset_a_start;
       offset_a<offset_a_start+cols_a;
       offset_a+=BLOCK_SIZE,
       offset_b+=BLOCK_SIZE*cols_b)
     {
-      __local double submatrix_a[BLOCK_SIZE][BLOCK_SIZE];
-      __local double submatrix_b[BLOCK_SIZE][BLOCK_SIZE];
-
      submatrix_a[i][j]=matrix_a[offset_a+cols_a*i+j];
      submatrix_b[i][j]=matrix_b[offset_b+cols_b*i+j];