#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 int num_threads_my;

/*
//----------------------------------------------------------------------------------------------
// ORIGINAL CODE
static void* mat_mul_thread(void *data) {
  // TODO: parallelize & optimize matrix multiplication
  for (int i = 0; i < M; ++i) {
      for (int j = 0; j < N; ++j) {
        for (int k = 0; k < K; ++k) {
          C[i * N + j] += A[i * K + k] * B[k * N + j];
        }
    }
  }

  return NULL;
}
//----------------------------------------------------------------------------------------------
*/


/*
//----------------------------------------------------------------------------------------------
// 1. Use multiple Threading 
//    About 30 sec for one multiplication.
//    Run time: 30.xx secs --> STOPPED BY LIMITED RUN TIME 

static void* mat_mul_thread(void *data) {
  // TODO: parallelize & optimize matrix multiplication
  // Divide-and-Conquer with threading: Divide jobs by num_threads 

  int pid = * (int *) data;		// pthread ID

  int i_slice = M / num_threads;	// separate jobs by num_threads
  int i_start = pid * i_slice;		// divide by row of A(i) as starting row
  int i_end = (pid == num_threads - 1) ? M : i_start + i_slice;		// ending row will be (slice size or M (max row of A)

  for (int i = i_start; i < i_end; i++) {	// Threading by pid and its row location calculation
	for(int j=0 ; j < N; j++) {
	  for(int k =0; k < K; k++) {
		C[ i*N + j] += A [i* K + k] * B[ k * N + j];
	  }	// for k
	} // for j
  } // for i

  return NULL;
}
//----------------------------------------------------------------------------------------------
*/


/*
//----------------------------------------------------------------------------------------------
// 1. Use Multiple Threading 
// 2. Tiling for column: bs(block size). bs selection affects its performance
//    Run time: 4.xx sec, around 30 GFLOPS

//#define MIN(x,y)  ((x) <= (y) ? (x) : (y))

static void* mat_mul_thread(void *data) {
  // TODO: parallelize & optimize matrix multiplication
  // Divide-and-Conquer with threading: Divide jobs by num_threads 
  // Tiling for kk (column of A or Row of B)

  int pid = * (int *) data;		// pthread ID

  int i_slice = M / num_threads;	// separate jobs by num_threads
  int i_start = pid * i_slice;		// divide by row of A(i) as starting row
  int i_end = (pid == num_threads - 1) ? M : i_start + i_slice;		// ending row will be (slice size or M (max row of A)
				  
  int idx_iNj;
  int idx_iKkk;
	
  int bs = 32;	// BLOCKSIZE: 32 --> around 30 GFLOPS, run time: 4.53 sec.(avg.)
  //int bs = 64;	// BLOCKSIZE: 64 --> around 26 GFLOPS, run time: 5.23 sec.(avg.)
  //int bs = 96;	// BLOCKSIZE: 96 --> around 24 GFLOPS, run time: 5.71 sec.(avg.)
  //int bs = 128;	// BLOCKSIZE: 128 --> around 24 GFLOPS, run time: 5.71 sec.(avg.)

  int min_kk;
  float sum;

  for (int kk = 0; kk < K; kk += bs) {	// Tiling according to column of A or Row of B
	min_kk = ((kk+bs) <= K) ? (kk+bs) : K;

    for (int i = i_start; i < i_end; i++) {	// Threading by pid and its row location calculation
		for(int j=0 ; j < N; j++) {
		    idx_iKkk = i * K + kk;
		    idx_iNj = i * N + j;
			sum = 0;
			//for(int k = kk; k < min_kk; k++) {
			for(int k = kk; k < MIN(kk+bs, K); k++) {
				sum += A [idx_iKkk++] * B[ k * N + j];
				//C[ idx_iNj] += A [idx_iKkk++] * B[ k * N + j];
				//C[ i*N + j] += A [i * K + k] * B[ k * N + j];
			}
		 C[idx_iNj] = sum;
		}
    }
  }

  return NULL;
  
}
//----------------------------------------------------------------------------------------------
*/


//----------------------------------------------------------------------------------------------
// 1. Use Multiple Threading 
// 2. Tiling for column of A or row of B: bs(block size). bs selection affects its performance
// 3. To use locality, change the order of (loop j) and (loop k), 
//     since B[k][j] for j=0,1,2,... is better than B[k][j] for k=0,1,2, ... 
//     (Occurred address jumping in every iteration, less data locality)
//     ==> 1D locality is larger than 2D locality for matrix multiplication in this case.
//         After trials for 2D tiling, as a result, 1D tiling has driven better performance than 2D tiling (around 150 ~ 170 GFLOPS)

//#define MIN(x,y)  ((x) <= (y) ? (x) : (y))

static void* mat_mul_thread(void *data) {
  // TODO: parallelize & optimize matrix multiplication
  // Divide-and-Conquer with threading: Divide jobs by num_threads 
  // Tiling for kk (column of A or Row of B)
  // Use data locality: Cange the order of (loop j) and (loop k), 
  //		                since B[k][j] for j=0,1,2,... is better than
  //							  B[k][j] for k=0,1,2, ... (Occurred address jumping in every iteration, less data locality)

  int pid = * (int *) data;		// pthread ID or index ( 0 ~ num_threads-1)

  int i_slice = M / num_threads;	// separate jobs by num_threads
  int i_start = pid * i_slice;		// divide by row of A[i][] as starting row
  int i_end = (pid == num_threads - 1) ? M : i_start + i_slice;		// ending row will be (slice size or M (max row of A)
				  
  float Aik;
  // int idx_iK;
  // int idx_iN;
  // int idx_kN;
	
  //int bs = 16;	// BLOCKSIZE: 16 --> around 200 GFLOPS, run time: 0.68 sec.(avg.)
  int bs = 32;	// BLOCKSIZE: 32 --> around 297 GFLOPS, run time: 0.46 sec.(avg.)
  //int bs = 64;	// BLOCKSIZE: 64 --> around 256 GFLOPS, run time: 0.54 sec.(avg.)
  //int bs = 96;	// BLOCKSIZE: 96 --> around 208 GFLOPS, run time: 0.66 sec.(avg.)
  //int bs = 128;	// BLOCKSIZE: 128 --> around 203 GFLOPS, run time: 0.67 sec.(avg.)

  int min_kk;

  for (int kk = 0; kk < K; kk += bs) {	// Tiling according to column of A or Row of B
	min_kk = ((kk+bs) <= K) ? (kk+bs) : K;

    for (int i = i_start; i < i_end; i++) {	// Threading by pid and its row location calculation

	    //idx_iK = i * K;			// No significant effect 
	    //idx_iN = i * N;			// No significant effect 
		//for(int k = kk; k < MIN(kk+bs, K); k++) {
		for(int k = kk; k < min_kk; k++) {
			Aik = A[i*K + k];			// Reduce iterative operation

			//idx_kN = k * N;			// No significant effect 

			for(int j=0; j < N; j++) {
				C[ i*N + j] += Aik * B[ k * N + j];			// slightly increase the performance

				//C[ i*N + j] += A[i*K + k] * B[ k * N + j];
				//C[ idx_iN + j] += Aik * B[ idx_kN + j];			// No significant effect 

			}	// for j
		}	// for k
	  }	 // for i
  }	// for kk

  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
  // Divide-and-Conquer: USE Multiple threads

//  pthread_t thread;
//  pthread_create(&thread, NULL, mat_mul_thread, NULL);
//  pthread_join(thread, NULL);


  pthread_t * threads;
  threads = (pthread_t *) malloc(sizeof(pthread_t) * num_threads);

  for (int i = 0; i < num_threads; i++) {
	int * pid = (int *) malloc(sizeof(int));
	*pid = i;
	pthread_create(&threads[i], NULL, mat_mul_thread, pid);		// Send pid index as arguments for each thread
  }

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

}