1

I wrote a program that tests _mm_prefetch(with gcc -O3): there are two random arrays,

  1. one of which traverses 1/600 of the length of the first random array (hardware prefetch, cache hit)
  2. another traverses the entire the second random array by 600 steps (hardware prefetch, cache miss)

I used the _mm_prefetch prefetch in the second test, but the performance still didn't improve.Why is this？what should I do? please tell me

The Code in C:

#include <stdio.h>
#include <time.h>
#include <stdlib.h>
#include <xmmintrin.h>
#define LENGTH 10000000
#define Max    10000000

//Generate a random array
int * random_arr(int lo,int hi,int num){
    int *arr = malloc(sizeof *arr *num);
    srand((unsigned)time(NULL));
    for (int i=0; i<num; i++) {
        arr[i]=rand()%(hi-lo)+lo;
    }
    return arr;
}

int main(int argc, const char * argv[]) {
    int * random_array1 = random_arr(0, Max, LENGTH); // the first  random array
    int * random_array2 = random_arr(0, Max, LENGTH); // the second random array

    clock_t start1 = clock();
    //-------------------------------------------------------------- first test
    for (int i = 0; i<LENGTH/600; ++i) {
        random_array1[i] = i;
    }

    //--------------------------------------------------------------
    clock_t end1 = clock();
    double duration1 = ((double)(end1 - start1)) / CLOCKS_PER_SEC;
    printf("The first  test running time: %lf s\n", duration1);
    //=========================================

    clock_t start2 = clock();
    //-------------------------------------------------------------- second test
    for (int i = 0; i < LENGTH; i+=600) {
        _mm_prefetch(&random_array2[i+600], _MM_HINT_T0);
        _mm_prefetch(&random_array2[i+1200], _MM_HINT_T0);
        random_array2[i] = i;
    }

    //--------------------------------------------------------------
    clock_t end2 = clock();
    double duration2 = ((double)(end2 - start2)) / CLOCKS_PER_SEC;
    printf("The second test running time: %lf s\n", duration2);

    return 0;
}

Code execution result：

The first  test running time: 0.000076 s
The second test running time: 0.001859 s
Linke
  • 336
  • 1
  • 10
  • You do a lot of unnecessary prefetching (about 16x too much) since each one does a whole cache line. My experience has been that it only is useful to do once before a loop as long as you're doing sequential access. Even then you only get minor improvement that is probably only useful in high speed trading. On older cpus, unrolling the loop to operate on a cache line at a time and prefetching the next one yielded significant improvements but the newest ones seem to do it for you. See: https://stackoverflow.com/q/46521694/1162141 – technosaurus Sep 21 '19 at 16:23
  • @technosaurus What I want to do is jump access to the array instead of sequential access. In this case, the element that access next should be prefetched. – Linke Sep 21 '19 at 18:00
  • As your code is now, you are double prefetching. Move one out of the loop and only prefetch the next.... Of course it would be better to organize your data for sequential access if possible. – technosaurus Sep 21 '19 at 19:45

0 Answers0