chundoong-lab-ta/SamsungDS22/submissions/HW2/h0.kim/mat_mul.cpp

63 lines
1.4 KiB
C++

#include "mat_mul.h"
#include <cstdlib>
#include <cstdio>
#include <pthread.h>
#include <immintrin.h>
static float *A, *B, *C;
static int M, N, K;
static int num_threads;
#define MIN(X, Y) (((X) < (Y)) ? (X) : (Y))
#ifdef __AVX512F__
#define IBS 64
#define KBS 40
#else
#define IBS 32
#define KBS 32
#endif
static void* mat_mul_thread(void *data) {
const int m = M / num_threads;
const int tid = *((int *) data);
const int begin = (tid * m);
const int end = (tid == num_threads - 1) ? M : (tid + 1) * m;
float Aik;
for (int ii = begin; ii < end; ii += IBS) {
for (int kk = 0; kk < K; kk += KBS) {
for (int i = ii; i < MIN(ii + IBS, end); ++i) {
for (int k = kk; k < MIN(kk + KBS, K); ++k) {
Aik = A[i * K + k];
for (int j = 0; j < 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;
pthread_t *threads;
threads = (pthread_t *) malloc(sizeof(pthread_t) * num_threads);
for (int i = 0; i < num_threads; ++i) {
int *tid = (int *) malloc(sizeof(int));
*tid = i;
pthread_create(&threads[i], NULL, mat_mul_thread, tid);
}
for (int i = 0; i < num_threads; ++i) {
pthread_join(threads[i], NULL);
}
}