#include "mat_mul.h"

#include <cstdlib>
#include <cstdio>
#include <pthread.h>

static float *A, *B, *C;
static int M, N, K;
static int num_threads;


static void* mat_mul_thread(void *data) {
  int thread_num = *((int*)(&data));
  int size_i = 4096;
  int size_j = 4096;
  int size_k = 64;
  int ii,jj,kk,i,j,k;
  for (ii = 0; ii < M; ii += size_i) {
    int min_ii;
    if(ii+size_i < M) min_ii = ii+size_i; else min_ii = M;
    for (jj = 0; jj < N; jj += size_j) {
      int min_jj;
      if(jj+size_j < N) min_jj = jj+size_j; else min_jj = N;
      for (kk = 0; kk < K; kk += size_k) {
        int min_kk;
        if(kk+size_k < K) min_kk = kk+size_k; else min_kk = K;
        for (i = ii; i < min_ii; i++) {
          int iN = i*N;
          int iK = i*K;
          if( (i%num_threads) == thread_num) {
            for (k = kk; k < min_kk; k++) {
              int kN = k*N;
              for (j = jj; j < min_jj; j++) {
                C[iN + j + 0] += A[iK+ k] * B[kN+ j + 0];
              }
            }
          }
        }
      }
    }
  }

  return NULL;
}


void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K, int _num_threads) {
  A = _A, B = _B, C = _C;
  M = _M, N = _N, K = _K;
  num_threads = _num_threads;

  // TODO: create '_num_threads' pthreads
  pthread_t thread[80]; //maximum number of threads is 40
  int new_threads;
  int idx;
  new_threads = num_threads;
  for(idx=0; idx<new_threads; idx++) {
    pthread_create(&thread[idx], NULL, mat_mul_thread, (void *)idx);
  }
  for(idx=0; idx<new_threads; idx++) {
    pthread_join(thread[idx], NULL);
  }
}