55 lines
1.3 KiB
C++
55 lines
1.3 KiB
C++
|
#include "mat_mul.h"
|
||
|
|
#include <pthread.h>
|
||
|
|
#include <immintrin.h>
|
||
|
|
|
||
|
|
static float *A, *B, *C;
|
||
|
|
static int M, N, K;
|
||
|
|
static int num_threads;
|
||
|
|
|
||
|
|
static int min(int x, int y) {
|
||
|
|
return x < y ? x : y;
|
||
|
|
}
|
||
|
|
|
||
|
|
#define ITILESIZE (32)
|
||
|
|
#define JTILESIZE (1024)
|
||
|
|
#define KTILESIZE (1024)
|
||
|
|
|
||
|
|
static void* mat_mul_thread(void *data) {
|
||
|
|
int tid = (long)data;
|
||
|
|
int is = M / num_threads * tid + min(tid, M % num_threads);
|
||
|
|
int ie = M / num_threads * (tid + 1) + min(tid + 1, M % num_threads);
|
||
|
|
|
||
|
|
for (int ii = is; ii < ie; ii += ITILESIZE) {
|
||
|
|
for (int jj = 0; jj < N; jj += JTILESIZE) {
|
||
|
|
for (int kk = 0; kk < K; kk += KTILESIZE) {
|
||
|
|
|
||
|
|
for (int k = kk; k < min(K, kk + KTILESIZE); k++) {
|
||
|
|
for (int i = ii; i < min(ie, ii + ITILESIZE); i++) {
|
||
|
|
float ar = A[i * K + k];
|
||
|
|
for (int j = jj; j < min(N, jj + JTILESIZE); j+=1) {
|
||
|
|
C[i * N + j] += ar * 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;
|
||
|
|
pthread_t threads[num_threads];
|
||
|
|
for (long i = 0; i < num_threads; ++i) {
|
||
|
|
pthread_create(&threads[i], NULL, mat_mul_thread, (void*)i);
|
||
|
|
}
|
||
|
|
for (long i = 0; i < num_threads; ++i) {
|
||
|
|
pthread_join(threads[i], NULL);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|