69 lines
2.0 KiB
C++
69 lines
2.0 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;
|
|
|
|
#define min(a, b) (((a) < (b)) ? (a) : (b))
|
|
#define BLOCKSIZE 45
|
|
|
|
static void* mat_mul_thread(void *data) {
|
|
int pid = * (int *) data;
|
|
|
|
int slice = M / num_threads; // 4096/20 : 204
|
|
u_int64_t start = pid * slice;
|
|
//eatch thread does 204 entiries Matrix Mul start: pid*204.
|
|
u_int64_t end = pid == num_threads - 1 ? M : (pid + 1) * slice;
|
|
float Aik;
|
|
int bs = BLOCKSIZE;
|
|
for (u_int64_t kk = 0; kk < (u_int64_t)K; kk += bs) { //4ea
|
|
u_int64_t mk = min(kk + bs, (u_int64_t)K);
|
|
for (u_int64_t i = start; i < end; ++i) {
|
|
for (u_int64_t k = kk; k < mk; ++k) {
|
|
Aik = A[i * K + k];
|
|
u_int64_t mmj = (N>>3)<<3;
|
|
u_int64_t j = 0;
|
|
for ( ; j < mmj; j+=8) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
C[i * N + j+1] += Aik * B[k * N + j+1];
|
|
C[i * N + j+2] += Aik * B[k * N + j+2];
|
|
C[i * N + j+3] += Aik * B[k * N + j+3];
|
|
C[i * N + j+4] += Aik * B[k * N + j+4];
|
|
C[i * N + j+5] += Aik * B[k * N + j+5];
|
|
C[i * N + j+6] += Aik * B[k * N + j+6];
|
|
C[i * N + j+7] += Aik * B[k * N + j+7];
|
|
}
|
|
for( ; j<(u_int64_t)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 * 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);
|
|
}
|
|
}
|