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

92 lines
2.3 KiB
C++

#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;
inline int min(int a, int b) { if(a>b) return b; else return a; }
pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
static void* mat_mul_thread(void *data) {
// TODO: parallelize & optimize matrix multiplication
static int tid = 0;
int num;
pthread_mutex_lock(&lock);
tid%=num_threads;
tid++;
num=tid;
pthread_mutex_unlock(&lock);
// printf("start: num = %d\n", num);
int q, r;
r = M % num_threads;
q = M / num_threads;
// if(num==1){ printf("q = %d, r = %d\n", q, r); }
if(r){
q++;
}
int i, j, k;
int kk, off_k;
int jj, off_j;
off_k = 64;
off_j = 1024;
float A_tmp;
if(num == num_threads){
for(jj = 0; jj < N; jj += off_j){
for(kk = 0; kk < K; kk += off_k){
for(i = q * (num-1); i < min(q * num + r, M); ++i){
for(k = kk; k < min(kk+off_k, K); ++k){
A_tmp = A[i * K + k];
// #pragma GCC unroll 2
for(j = jj; j < min(jj+off_j, N); ++j){
C[i * N + j] += A_tmp * B[k * N + j];
}
}
}
}
}
}else{
for(jj = 0; jj < N; jj += off_j){
for(kk = 0; kk < K; kk += off_k){
for(i = q * (num-1); i < min(q * num, M); ++i){
for(k = kk; k < min(kk+off_k, K); ++k){
A_tmp = A[i * K + k];
// #pragma GCC unroll 2
for(j = jj; j < min(jj+off_j, N); ++j){
C[i * N + j] += A_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;
// TODO: create '_num_threads' pthreads
int i;
pthread_t thread[num_threads];
for(i=0; i<num_threads; i++){
pthread_create(&thread[i], NULL, mat_mul_thread, NULL);
}
for(i=0; i<num_threads; i++){
pthread_join(thread[i], NULL);
}
}