/* 3 */
#define TS  32
#define WPT 8
#define RTS 4   // (TS/WPT)


// super super slow sgemm kernel by heehoon
// SGEMM(single-precision matrix-multiplication)
// /* 2 */__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K, int TS) {
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {
/* 3 */
    // Thread identifiers
    const int row = get_local_id(0);    // local row (max = TS)
    const int col = get_local_id(1);    // local col (max = TS/WPT = RTS)
    const int global_row = TS * get_group_id(0) + row;  // row ID of C (0~M)
    const int global_col = TS * get_group_id(1) + col;  // col ID of C (0~N)

    __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 num_tiles = K / TS;
    const int num_tiles = (K + TS - 1) / TS;
    //printf("\nnum_tilesi: (K + TS -1)/TS = %d\n", 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;

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

    // Store
    for (int w=0; w<WPT; w++) {
        if (global_row + w*RTS >= M || global_col >= N)
            continue;
        C[(global_row + w*RTS) * N + global_col] = intermediate_val[w];
    }

/* 1. original code
  int i = get_global_id(0); // row index of C (max = TS)
  int j = get_global_id(1); // column index of C (max = TS/WPT = RTS)
  if (i >= M || j >= N) return; // boundary check

  C[i * N + j] = 0;
  for (int k = 0; k < K; k++) {
    C[i * N + j] += A[i * K + k] * B[k * N + j];
  }
*/

/* 2. TS = 8 -> run_performance.sh: 922.67 GFLOPS but run_validate.sh -> INVALID
    //printf("\n[hong] TS = %d\n", TS);
    // Thread identifiers
    const int row = get_local_id(0);    // local row
    const int col = get_local_id(1);    // local col
    //const int global_row = 32 * get_group_id(0) + row;  // row ID of C
    //const int global_col = 32 * get_group_id(1) + col;  // col ID of C
    const int global_row = TS * get_group_id(0) + row;  // row ID of C
    const int global_col = TS * get_group_id(1) + col;  // col ID of C

    __local float Asub[32][32];
    __local float Bsub[32][32];
    //__local float Asub[8][8];
    //__local float Bsub[8][8];
    // __local float Asub[TS][TS]; // variable length arrays are not supported in OpenCL
    // __local float Bsub[TS][TS]; // variable length arrays are not supported in OpenCL

    float intermediate_val = 0.0f;
    //const int num_tiles = K / 32;
    const int num_tiles = K / TS;
    for (int t = 0; t < num_tiles; t++) {
        //const int t_row = 32 * t + row;
        //const int t_col = 32 * t + col;
        const int t_row = TS * t + row;
        const int t_col = TS * t + col;
        Asub[row][col] = A[global_row * K + t_col];
        Bsub[row][col] = B[t_row * N + global_col];

        barrier(CLK_LOCAL_MEM_FENCE);

        //for (int k = 0; k < 32; k++) {
        for (int k = 0; k < TS; k++) {
            intermediate_val += Asub[row][k] * Bsub[k][col];
        }

        barrier(CLK_LOCAL_MEM_FENCE);
    }

    C[global_row * N + global_col] = intermediate_val;
*/
  
}