Optimising loops for pixel processing - optimization

Question

I'm implementing an algorithm (in OpenCV) that iterates over every pixel in an image, and for each pixel calculates block matches with pixels in the neighbourhood in order to evalaute the similarity of these neighbouring pixels. A "naive" implementation with very deep loops is very slow, so I was wondering how I might try to improve the performance. The following is an extract of my current code:
for(nCh=1;nCh<=channels;nCh++) { // Loop over three channels
for(i=0;i<h;i++) { // "vertical" loop
for(j=0;j<w;j++) { // "horizontal" loop
for (si=-sw_height; si<sw_height; si++){ // vertical search window loop
for (sj=-sw_width; sj<sw_width; sj++){ // horizontal search window loop
dist = 0;
for (blki=0; blki<blk_height; blki++){ // block match loop
for (blkj=0; blkj<blk_width; blkj++){ // block match loop
current_pxl = data[(i+blki)*step+(j+blkj)*channels+nCh];
search_pxl = data[(i+blki+si)*step+(j+blkj+sj)*channels+nCh];
dist += pow((current_pxl - search_pxl),2);
}
}
// ... further processing
}
}
}
}
}

Answer 1

You're calling pow in the innermost loop. Don't.
Also you're doing a lot of index calculation in there.
I bet you can move some out that out of the inner loop.
You should be able to get it so your inner loop looks more like this:
for (blkj = 0; blkj < blk_width; blkj++, pc += channels, ps += channels){
int diff = (*pc - *ps);
dist += (diff * diff);
}
And then, you might even want to unroll it a bit.
BTW, a little more whitespace might help :-)