#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;

#define BS	40	
#define MIN(a,b) ((a>b) ? b : a)


static void* mat_mul_thread(void *data) {
  int pid = * (int *) data;

  int slice = M / num_threads;
  int start = pid * slice;
  int end = pid == num_threads - 1 ? M : (pid + 1) * slice;
  
  int kk,i,j,k;

  float Aik;
  
	for (kk = 0; kk < K; kk += BS){
		for (i = start; i < end; i++){
			for (k = kk; k < MIN(kk+BS, K); k++)
			{
				Aik = A[i * K + k];
				for (j = 0; j < N; j++){
					C[i * N + j] += Aik * 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;

  // TODO: create '_num_threads' pthreads
  pthread_t thread[40];
  int pid[40];
  
	for (int i = 0; i < num_threads; i++) {
		pid[i] = i;
		pthread_create(&thread[i], NULL, mat_mul_thread, &pid[i]);
	}

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