// super super slow sgemm kernel by heehoon

#define TS 32
#define WPT 8

__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {

 const int row = get_local_id(0);
 const int col = get_local_id(1);
 const int global_row = TS*get_group_id(0) + row;
 const int global_col = TS*get_group_id(1) + col;

 __local float Asub[TS][TS];
 __local float Bsub[TS][TS];

 float intermediate_val[WPT];
 for(int w=0; w<WPT; w++) {
    intermediate_val[w] = 0.0f;
 }
 const int RTS = TS/WPT;
 const int num_tiles = (K % TS == 0)? K / TS : K / TS + 1;
 
 for(int t = 0; t < num_tiles; t++){
   for(int w = 0; w < WPT; w++){
     const int t_row = TS * t + row;
     const int t_col = TS * t + col;

     if( t_col >= K ) Asub[row + w*RTS][col] = 0;
      else Asub[row + w*RTS][col] = A[(global_row + w*RTS) * K + t_col];
     
     if( (t_row >= K) || (global_col >= N) ) Bsub[row + w*RTS][col] = 0;
      else Bsub[row + w*RTS][col] = B[(t_row + w*RTS)*N + global_col];
     }

   barrier(CLK_LOCAL_MEM_FENCE);

   for(int k = 0; k < TS; k++){
     for(int w = 0; w < WPT; w++){
      intermediate_val[w] += Asub[row + w*RTS][k] * Bsub[k][col];
     }
   }

   barrier(CLK_LOCAL_MEM_FENCE);

 }

 for(int w = 0; w<WPT; w++){
  if( global_col < N )
   C[(global_row + w*RTS)*N + global_col] = intermediate_val[w];
 }

}