204 lines
4.7 KiB
C++
204 lines
4.7 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;
|
|
|
|
//static void* mat_mul_thread(void *data) {
|
|
// // TODO: parallelize & optimize matrix multiplication
|
|
// for (int i = 0; i < M; ++i) {
|
|
// for (int j = 0; j < N; ++j) {
|
|
// for (int k = 0; k < K; ++k) {
|
|
// C[i * N + j] += A[i * K + k] * B[k * N + j];
|
|
// }
|
|
// }
|
|
// }
|
|
//
|
|
// return NULL;
|
|
//}
|
|
|
|
//static void* mat_mul_thread(void *data) {
|
|
// float acc00, acc01, acc10, acc11;
|
|
// // TODO: parallelize & optimize matrix multiplication
|
|
// for (int i = 0; i < M; i+=2) {
|
|
// for (int j = 0; j < N; j+=2) {
|
|
// acc00 = acc01 = acc10 = acc11 = 0;
|
|
// for (int k = 0; k < K; ++k) {
|
|
// //acc00 = B[k*2+j] * A[i*2+k];
|
|
// acc00 += A[i * K + k] * B[k * N + j];
|
|
// acc01 += A[i * K + k] * B[k * N + j+1];
|
|
// acc10 += A[(i+1) * K + k] * B[k * N + j];
|
|
// acc11 += A[(i+1) * K + k] * B[k * N + j+1];
|
|
// //C[i * N + j] += A[i * K + k] * B[k * N + j];
|
|
// }
|
|
// C[i*N+j] += acc00;
|
|
// C[i*N+j+1] += acc01;
|
|
// C[(i+1)*N+j] += acc10;
|
|
// C[(i+1)*N+j+1] += acc11;
|
|
// }
|
|
// }
|
|
//
|
|
// return NULL;
|
|
//}
|
|
|
|
/*
|
|
static void* mat_mul_thread(void *data) {
|
|
// TODO: parallelize & optimize matrix multiplication
|
|
for (int i = 0; i < M; ++i) {
|
|
for (int j = 0; j < N; ++j) {
|
|
for (int k = 0; k < K; ++k) {
|
|
C[i * N + j] += A[i * K + k] * B[k * N + j];
|
|
}
|
|
}
|
|
}
|
|
|
|
return NULL;
|
|
}
|
|
*/
|
|
/*
|
|
static void* mat_mul_thread(void *data) {
|
|
*/
|
|
static int min (int a, int b) {
|
|
return ( a > b ) ? b : a;
|
|
}
|
|
|
|
static void* worker(void *data) {
|
|
|
|
int pid = * (int *) data;
|
|
|
|
int slice = M / num_threads;
|
|
int start = pid * slice;
|
|
int end = pid == num_threads - 1 ? M : (pid + 1) * slice;
|
|
|
|
float Aik;
|
|
int k_bs = 32;
|
|
int j_bs = 1024;
|
|
//int bs = 1024;
|
|
//int bs = 512;
|
|
//int bs = 256;
|
|
//int bs = 2048;
|
|
// int bs = BLOCKSIZE;
|
|
/*
|
|
for (int kk = 0; kk < K; kk += bs) {
|
|
for (int jj = 0; jj < N; jj += bs) {
|
|
for (int i = start; i < end; ++i) {
|
|
for (int k = kk; k < min(kk + bs, K); ++k) {
|
|
Aik = A[i * K + k];
|
|
for (int j = jj; j < min(jj + bs, N); ++j) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
*/
|
|
|
|
/*
|
|
sum1 = 0, sum2 = 0;
|
|
for(i=0; i<n/2; i++) sum1 += a[2*i], sum2 += a[2*i+1];
|
|
for(i=(n/2)*2; i<n; i++) sum += a[i]; // clean up for n odd
|
|
sum += sum1 + sum2;
|
|
*/
|
|
/*
|
|
__m256 a0;
|
|
__m256 b0;
|
|
__m256 s0;
|
|
s0 = _mm256_set_ps(0.,0.,0.,0.,0.,0.,0.,0.);
|
|
*/
|
|
|
|
for (int kk = 0; kk < K; kk += k_bs) {
|
|
for (int jj = 0; jj < N; jj += j_bs) {
|
|
for (int i = start; i < end; i++) {
|
|
for (int k = kk; k < min(kk + k_bs, K); k++) {
|
|
Aik = A[i * K + k];
|
|
/*
|
|
*/
|
|
for (int j = jj; j < min(jj + j_bs, N); j++) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
/*
|
|
// ORG Start
|
|
// 194 GFLOPs
|
|
for (int j = jj; j < min(jj + bs, N) ; j=j+2) {
|
|
//if (i==0 && j==256) printf("##### %d %d\n",Aik, j);
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
C[i * N + j + 1] += Aik * B[k * N + j + 1];
|
|
}
|
|
for (int j = (min(jj + bs, N)/2)*2; j < min(jj + bs, N); j++) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
// ORG End
|
|
*/
|
|
/*
|
|
// No j
|
|
// 320
|
|
for (int j = 0; j < N ; j++) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
*/
|
|
/*
|
|
for (int j = jj; j < min(jj + bs, N) ; j=j+8) {
|
|
//if (i==0 && j==256) printf("##### %d %d\n",Aik, j);
|
|
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];
|
|
//C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
*/
|
|
/*
|
|
for (int j = jj; j < min(jj + bs, N) ; j=j+8) {
|
|
a0 = _mm256_load_ps(a + j * 8);
|
|
b0 = _mm256_load_ps(b + j * 8);
|
|
s0 = _mm256_fmadd_ps(a0, b0, s0);
|
|
}
|
|
*/
|
|
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
return NULL;
|
|
}
|
|
|
|
|
|
/*
|
|
*/
|
|
|
|
static void* mat_mul_thread(void *data) {
|
|
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, worker, pid);
|
|
}
|
|
|
|
for (int i = 0; i < num_threads; i++) {
|
|
pthread_join(threads[i], NULL);
|
|
}
|
|
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;
|
|
pthread_create(&thread, NULL, mat_mul_thread, NULL);
|
|
pthread_join(thread, NULL);
|
|
}
|