#include "mat_mul.h"

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

#define bs 32
#define min(a, b) (((a) < (b)) ? (a) : (b))
#define max(a, b) (((a) > (b)) ? (a) : (b))

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

static void* mat_mul_thread(void *data) {

  int pid = *(int *) data;

  int chunk = M / num_threads;
  int start = pid * chunk;
  int end = pid == num_threads - 1 ? M : (pid + 1) * chunk;
  
  float tmp;

  for (int kk = 0; kk < K; kk += bs) {
    for (int i = start; i < end; ++i) {
      for (int k = kk; k < min(kk + bs, K); ++k) {
        tmp = A[i * K + k];
          for (int j = 0; j < N; ++j) {
            C[i * N + j] += tmp * B[k * N + j];
        }
      }
    }
  }

  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;
  
  pthread_t thread[num_threads];
  int pid[num_threads];

  for (int i = 0; i < num_threads; i++) {
    pid[i] = i;
    pthread_create(&thread[i], NULL, mat_mul_thread, (void*) &pid[i]);
  }

  for (int i = 0; i < num_threads; i++) {
    pthread_join(thread[i], NULL);
  }

}