#define TS 32              // tile size
#define WPT 8              // work per thread
#define RTS (TS/WPT)       // 

#define WPT_NR 16          // work per thread in case no remainder
#define RTS_NR (TS/WPT_NR)  

#define PAD_X 16       
#define PAD_Y 16       

// single precision matrix multiplication
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {

  // thread index
  const int row = get_local_id(0);                    // local row index of C
  const int col = get_local_id(1);                    // local column index of C
  const int global_row = TS * get_group_id(0) + row;  // global row index of C
  const int global_col = TS * get_group_id(1) + col;  // global column index of C  
  __local float Asub[TS][TS];
  __local float Bsub[TS][TS];

  float intermediate_val[WPT];
  
  //printf("row, col: %d, %d  // global row, global col: %d, %d \n", row, col, global_row, global_col);
  
  for(int w = 0; w < WPT; w++) {
    intermediate_val[w] = 0.0f; 
  }
  
  // remainder
  const int num_tiles = (K + TS - 1) / TS;
  //printf("K = %d, K % TS = %d, numtile = %d\n", K, (K % TS), num_tiles);
  
  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(global_row + w * RTS >= M || t_col >= K)
        Asub[row + w * RTS][col] = 0.0f;
      else
        Asub[row + w * RTS][col] = A[(global_row + w * RTS) * K + t_col];
      
      if(t_row + w * RTS >= K || global_col >= N) 
        Bsub[row + w * RTS][col] = 0.0f;
      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_row + w * RTS >= M || global_col >= N)
  	  continue;
  	else
  	  C[(global_row + w * RTS) * N + global_col] = intermediate_val[w];
  }
}

__kernel void sgemmNoRemainder(__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_NR];
  
  for(int w = 0; w < WPT_NR; w++) {
	  intermediate_val[w] = 0.0f; 
  }

  // No remainder
  const int num_tiles = K / TS;

  for(int t = 0; t < num_tiles; t++){
    for(int w = 0; w < WPT_NR; w++){
      const int t_row = TS * t + row;
      const int t_col = TS * t + col;
      
      Asub[row + w * RTS_NR][col] = A[(global_row + w * RTS_NR) * K + t_col];
      Bsub[row + w * RTS_NR][col] = B[(t_row + w * RTS_NR) * N + global_col];
    }
    barrier(CLK_LOCAL_MEM_FENCE);

    for (int k = 0; k < TS; k++) {
      for(int w = 0; w < WPT_NR; w++) { 
        intermediate_val[w] += Asub[row + w * RTS_NR][k] * Bsub[k][col];
      }
    }
    barrier(CLK_LOCAL_MEM_FENCE);
  }

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


















// Pad the P * Q matrix with zeroes to form a P_XL * Q_XL matrix
__kernel void paddingAddZeroes(const int P, const int Q, const __global float* input, const int P_XL, const int Q_XL, __global float* output) {
    
    // thread index
    const int tx = get_group_id(0)*PAD_X + get_local_id(0); // 0..P_XL in blocks of PAD_X
    const int ty = get_group_id(1)*PAD_Y + get_local_id(1); // 0..Q_XL in blocks of PAD_Y

    // Check whether we are within bounds of the XL matrix
    if (tx < P_XL && ty < Q_XL) {

        // Copy the input or pad a zero
        float value;
        if (tx < P && ty < Q) { value = input[ty*P + tx]; }
        else { value = 0.0f; }

        // Store the result
        output[ty*P_XL + tx] = value;
    }
}

// Remove padded values from a P_XL * Q_XL matrix to form a P * Q matrix
__kernel void paddingRemoveZeroes(const int P_XL, const int Q_XL, const __global float* input, const int P, const int Q, __global float* output) {
    
    // Thread identifiers
    const int tx = get_group_id(0)*PAD_X + get_local_id(0); // 0..P in blocks of PAD_X
    const int ty = get_group_id(1)*PAD_Y + get_local_id(1); // 0..Q in blocks of PAD_Y

    // Only store the result if within P * Q bounds
    if (tx < P && ty < Q) { output[ty*P + tx] = input[ty*P_XL + tx]; }
}