63 lines
1.6 KiB
C++
63 lines
1.6 KiB
C++
#include "mat_mul.h"
|
|
|
|
#include <immintrin.h>
|
|
#include <cstdlib>
|
|
#include <cstdio>
|
|
#include <pthread.h>
|
|
|
|
static float *A, *B, *C;
|
|
static int M, N, K;
|
|
static int num_threads;
|
|
|
|
static void* mat_mul_thread(void *data) {
|
|
int tid = *(int *)data;
|
|
|
|
int num = (M + num_threads - 1) / num_threads;
|
|
int start = num * tid;
|
|
int end = (tid == num_threads -1) ? M : num * (tid + 1);
|
|
int blocksizek = 32;
|
|
int blocksizem = 64;
|
|
int blocksizen = 4096;
|
|
|
|
if (start >= M) return NULL;
|
|
|
|
// TODO: parallelize & optimize matrix multiplication
|
|
for (int ii = start ; ii < end ; ii +=blocksizem) {
|
|
for (int jj = 0 ; jj < N ; jj +=blocksizen) {
|
|
for (int kk = 0 ; kk < K ; kk += blocksizek) {
|
|
int endm = ii + blocksizem < end ? (ii + blocksizem) : end;
|
|
int endn = jj + blocksizen < N ? (jj + blocksizen) : N;
|
|
int endk = kk + blocksizek < K ? (kk + blocksizek) : K;
|
|
for (int i = ii ; i < endm ; i++) {
|
|
for (int k = kk ; k < endk ; k++) {
|
|
for (int j = jj ; j < endn ; j++) {
|
|
C[i * N +j] += A[i * K + k] * 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[80];
|
|
int tid[80];
|
|
|
|
for (int i = 0 ; i < num_threads ; i++) {
|
|
tid[i] = i;
|
|
pthread_create(&thread[i], NULL, mat_mul_thread, &tid[i]);
|
|
}
|
|
|
|
for (int i = 0 ; i < num_threads ; i++) {
|
|
pthread_join(thread[i], NULL);
|
|
}
|
|
}
|