I finally (sigh) started working on master thesis and that convolution I showed above is just a small prerequisite.

Since I am trying to improve upon an existing method for detecting and recognizing faces, accuracy is more important, not so speed. The convolution is made in frequency domain, so, together with loading the image, calculating the kernel(s) and other (infrastructural) overhead everything runs within ~230ms, which is OK. I am not yet done with multithreading yet and there are other aspects which are not that efficient yet, but that is not yet important

Ask again if I start evaluating the system, when I'm done.

@PadMalcolm: I don't think so, but the thesis itself will be available online, I guess. If my approach is a promising improvement with rock solid results backing it, well, I don't know what happens then