chundoong-lab-ta/SamsungDS22/submissions/HW2/hongpooh.kim/mat_mul.cpp

#include "mat_mul.h"

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

//#define BLOCKSIZE 128		// 200.00 GFLOPS
//#define BLOCKSIZE 64		// 287.41 GFLOPS
#define BLOCKSIZE 32		// 293.77 GFLOPS (290 ~ 295)
//#define BLOCKSIZE 16		// 199.14 GFLOPS
//#define BLOCKSIZE 4		// 69.86 GFLOPS
#define MIN(blk, total) (blk < total ? blk : total)

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


static void* mat_mul_thread(void *data) {
  // TODO: parallelize & optimize matrix multiplication
  	int i, j, k, kk;
	int pid;
	int div;
	int div_start, div_end;

	pid = *(int *)data;
	div = M / num_threads;
	div_start = pid * div;
	div_end = (pid == num_threads -1) ? M : (pid + 1) * div;

	/* 1) original code
	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];
					}
			}
	}
	 */

	/* 2) pthread only 
	 * ./run.sh -v -n 5 -t 1 1024 1024 1024 => 0.678401 GFLOPS
	 * ./run.sh -v -n 5 -t 2 1024 1024 1024 => 1.356989 GFLOPS
	 * ./run.sh -v -n 5 -t 3 1024 1024 1024 => 2.008897 GFLOPS
	 * ./run.sh -v -n 5 -t 4 1024 1024 1024 => 2.702098 GFLOPS
	 * ./run_performance.sh => Force Terminated (time over: 1 min) 
	 
	for (int i = div_start; i < div_end; ++i) {
			//for (int j = Ndiv_start; j < Ndiv_end ; ++j) {
			for (int j = 0; j < N ; ++j) {
					int k = 0;
					for ( ; k < K; ++k) {
							C[i * N + j] += A[i * K + k] * B[k * N + j];
					}
			}
	}
	 */

	/* 3) prhead + block (final)
	 * ./run_performance.sh
	 * #define BLOCKSIZE 128	=> 200.00 GFLOPS
	 * #define BLOCKSIZE 64		=> 287.41 GFLOPS
	 * #define BLOCKSIZE 32		=> 293.77 GFLOPS (290 ~ 295), 0.479 sec
	 * #define BLOCKSIZE 16		=> 199.14 GFLOPS
	 * #define BLOCKSIZE 4		=> 69.86 GFLOPS
	 *
	 * ./run.sh -v -n 10 -t 40 4096 4096 4096 => 0.411518 sec, 333.980281 GFLOPS
	 */ 
	for (kk =0; kk < K; kk += BLOCKSIZE) {
			for (i = div_start; i < div_end; i++) {
					for(k = kk; k < MIN(kk + BLOCKSIZE, K); k++) {
							for (j = 0; j < N; j++) {
								C[i * N + j] += A[i * K + k] * B[k * N + j];
							}	
					}
			}
	} 

	/* test code - #1
	 * ./run.sh -v -n 5 -t 40 4096 4096 4096	=> 4.808995 sec, 28.579559 GFLOPS
	 * ./run.sh -v -n 5 -t 20 4096 4096 4096	=> time out
	 
	for (kk =0; kk < K; kk += BLOCKSIZE) {
			for (i = div_start; i < div_end; i++) {
					for (j = 0; j < N; j++) {
							for(k = kk; k < MIN(kk + BLOCKSIZE, K); k++) {
									C[i * N + j] += A[i * K + k] * B[k * N + j];
							}	
					}
			}
	}
	 */

	/* test code - #2
	 * ./run_performance.sh	=> 4.621823 sec, 29.736956 GFLOPS

	for (kk =0; kk < K; kk += BLOCKSIZE) {
			for (i = div_start; i < div_end; i++) {
					for(k = kk; k < MIN(kk + BLOCKSIZE, K); k++) {
							float tmpC = 0;
							for (j = 0; j < N; j++) {
								//C[i * N + j] += A[i * K + k] * B[k * N + j];
								tmpC += A[i * K + k] * B[k * N + j];
							}	
							C[i * N + j] = tmpC;
					}
			}
	} 
	 */

	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
  /* single thread ref. code
	 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);
	}

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