Question

Here is a small test I did and the result surprised me: doing the same loop twice was approximately twice as fast as looping once. I am guessing it as because of memory access?

float* A = new float[1000000];
float* B = new float[1000000];
int h,w;
h = w = 1000;
CString txt;
double time1, time2;

time1 = Timer::instance()->getTime();
for(int j = 0; j < h; j++){
    for(int i = 0; i < w; i++){
        A[i+j*w] = 1;
        B[i+j*w] = 1;
    }
}
time2 = Timer::instance()->getTime();
txt.Format(_T("Both in same loop = %f"),time2-time1);
AfxMessageBox(txt);

time1 = Timer::instance()->getTime();
for(int j = 0; j < h; j++){
    for(int i = 0; i < w; i++){
        A[i+j*w] = 1;
    }
}
for(int j = 0; j < h; j++){
    for(int i = 0; i < w; i++){
        B[i+j*w] = 1;
    }
}
time2 = Timer::instance()->getTime();
txt.Format(_T("Different loops = %f"),time2-time1);
AfxMessageBox(txt);

Answer 1

It could be CPU cache, but more likely it’s the concurrent memory access. When you access array1[x], and then immediately after that array2[x], those are two very different locations in memory and it’s difficult to optimize. However array[0], array[1], array[2] etc are all in contiguous memory and much more efficient to access. Intel seems to agree.