176 lines
6.9 KiB
C++
176 lines
6.9 KiB
C++
#include "mat_mul.h"
|
|
|
|
#include <cstdlib>
|
|
#include <cstdio>
|
|
#include <pthread.h>
|
|
#include <immintrin.h>
|
|
|
|
// __restrict directive means that no memory aliasing on those pointer variables, not much impressive effects
|
|
static float *__restrict A, * __restrict B, * __restrict C;
|
|
static int M, N, K;
|
|
static int num_threads;
|
|
|
|
// Debug
|
|
#define USING_ONE_THREAD (0)
|
|
|
|
#define VECTOR_SIZE (128)
|
|
#define BLOCKSIZE (48)
|
|
#define JBLOCKSIZE (VECTOR_SIZE * 8)
|
|
#define OPTIMAL_BLOCKSIZE (32)
|
|
#define OPTIMAL_MATRIX_SIZE (4096)
|
|
#define OPTIMAL_THREAD_CNT (20)
|
|
#define MIN(__A,__B) ((__A) < (__B) ? (__A) : (__B)) // Can std::min goes to inline function??
|
|
|
|
static void* mat_mul_thread_Opt(void* data) {
|
|
int pid = *(int*)data;
|
|
int slice = OPTIMAL_MATRIX_SIZE / OPTIMAL_THREAD_CNT;
|
|
int start = pid * slice;
|
|
int end = pid == (OPTIMAL_THREAD_CNT - 1) ? OPTIMAL_MATRIX_SIZE : (pid + 1) * slice;
|
|
|
|
for (int kk = 0; kk < OPTIMAL_MATRIX_SIZE; kk += BLOCKSIZE)
|
|
{
|
|
for (int jj = 0; jj < OPTIMAL_MATRIX_SIZE; jj += JBLOCKSIZE)
|
|
{
|
|
// Prefetch, not much impressive effects
|
|
_mm_prefetch((const char*)&A[start * OPTIMAL_MATRIX_SIZE + kk], _MM_HINT_T0);
|
|
for (int i = start; i < end; ++i)
|
|
{
|
|
#if 1
|
|
// Prefetch, not much impressive effects
|
|
_mm_prefetch((const char*)&C[i * OPTIMAL_MATRIX_SIZE + jj], _MM_HINT_T0);
|
|
_mm_prefetch((const char*)&B[kk * OPTIMAL_MATRIX_SIZE + jj], _MM_HINT_T0);
|
|
for (int k = kk; k < MIN(kk + BLOCKSIZE, OPTIMAL_MATRIX_SIZE); ++k)
|
|
{
|
|
// GCC merges aik variables, meaningless but tried
|
|
__m512 aik, aik2, aik3, aik4;
|
|
// Fill A values to aik (32 bit per float * 16EA = 512 bit (__m512))
|
|
aik4 = aik3 = aik2 = aik = _mm512_set1_ps(A[i * OPTIMAL_MATRIX_SIZE + k]);
|
|
|
|
// 4096 % 128 == 0, don't need to use MIN macro
|
|
for (int j = jj; j < jj + JBLOCKSIZE; j += VECTOR_SIZE)
|
|
{
|
|
__m512 c1, c2, c3, c4, c5, c6, c7, c8;
|
|
|
|
// Load C matrix variables
|
|
c1 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j]);
|
|
c2 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x10]);
|
|
c3 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x20]);
|
|
c4 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x30]);
|
|
c5 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x40]);
|
|
c6 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x50]);
|
|
c7 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x60]);
|
|
c8 = _mm512_load_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x70]);
|
|
|
|
// FMA Local C = B * A + C
|
|
c1 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j]), aik, c1);
|
|
c2 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x10]), aik2, c2);
|
|
c3 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x20]), aik3, c3);
|
|
c4 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x30]), aik4, c4);
|
|
c5 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x40]), aik, c5);
|
|
c6 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x50]), aik2, c6);
|
|
c7 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x60]), aik3, c7);
|
|
c8 = _mm512_fmadd_ps(_mm512_load_ps(&B[k * OPTIMAL_MATRIX_SIZE + j + 0x70]), aik4, c8);
|
|
|
|
// Store C matrix values
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j], c1);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x10], c2);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x20], c3);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x30], c4);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x40], c5);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x50], c6);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x60], c7);
|
|
_mm512_store_ps(&C[i * OPTIMAL_MATRIX_SIZE + j + 0x70], c8);
|
|
}
|
|
}
|
|
#else
|
|
for (int k = kk; k < kk + BLOCKSIZE; ++k)
|
|
{
|
|
float Aik, Aik2;
|
|
Aik2 = Aik = A[i * K + k];
|
|
for (int j = 0; j < N; j = j + 4)
|
|
{
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
C[i * N + j + 1] += Aik2 * B[k * N + j + 1];
|
|
C[i * N + j + 2] += Aik * B[k * N + j + 2];
|
|
C[i * N + j + 3] += Aik2 * B[k * N + j + 3];
|
|
}
|
|
}
|
|
#endif
|
|
}
|
|
}
|
|
}
|
|
|
|
return NULL;
|
|
}
|
|
|
|
static void* mat_mul_thread(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 bs = BLOCKSIZE;
|
|
|
|
for (int kk = 0; kk < K; kk += 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 = 0; j < N; ++j) {
|
|
C[i * N + j] += Aik * B[k * N + j];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
return NULL;
|
|
}
|
|
|
|
void mat_mul(float * __restrict _A, float * __restrict _B, float * __restrict _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;
|
|
|
|
if (M == OPTIMAL_MATRIX_SIZE
|
|
&& N == OPTIMAL_MATRIX_SIZE
|
|
&& K == OPTIMAL_MATRIX_SIZE
|
|
&& num_threads == OPTIMAL_THREAD_CNT)
|
|
{
|
|
// Performance checking condition, can reduce unroll cost, constant folding ...
|
|
pthread_t threads[OPTIMAL_THREAD_CNT];
|
|
int pid[OPTIMAL_THREAD_CNT];
|
|
|
|
#if (USING_ONE_THREAD)
|
|
_num_threads = 1;
|
|
#endif // USING_ONE_THREAD
|
|
|
|
for (int i = 0; i < _num_threads; ++i) {
|
|
pid[i] = i;
|
|
pthread_create(&threads[i], NULL, mat_mul_thread_Opt, &pid[i]);
|
|
}
|
|
|
|
for (int i = 0; i < _num_threads; ++i) {
|
|
pthread_join(threads[i], NULL);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
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);
|
|
}
|
|
}
|
|
}
|