Update matrix_opencl.cpp

matrix_opencl.cpp

@@ -98,7 +98,7 @@ const std::string kernel_source_matrix_mul = R"(
     int nloc = get_local_size(0);
 
     float tmp;
-    float Awrk[1024];
+    float Awrk[10000];
 
     for (k = 0; k < K; k++)
         Awrk[k] = A[i * K + k];