On Mon, Jul 07, 2014 at 03:26:10AM +0200, Krzysztof Kosiński wrote:
2014-07-07 2:32 GMT+02:00 Jon A. Cruz <jon@...18...>:
What about tiling the filtering? At some point to address 'c' we should be looking more into breaking up to sections. However I'm not following that code at the momentenough to know if we're getting close to that point yet or not.
The filtering is already broken into tiles. However, when the Gaussian blur radius is for example 1000 screen pixels, then every pixel of the output depends on a 2000x2000 px area of the input. There is no way around that.
Actually there is, you can do it with short-time fourier transforms, which compute a large scale component just once and share it across all the tiles, then only compute the information around each tile at the local scale. There are similar tricks built around the fact that a large scale binomial distribution is the same as a gaussian.
It might even be worth implementing if this is a common problem. But I would look at approximations for the local area before spending time on that.
njh