Hacker News Reader

Comment by Gravityloss | Hacker News Reader

Why is blur supposed to be computationally intensive? It's very simple mathematically...

It's not the computation itself, it's the GPU memory bandwidth it takes up. (Assuming you're doing it in real-time of course.)

if you blur an MxM area with a NxN kernel you have approximately MxMxNxN multiply adds. That can add up to a lot especially on a retina display. You are right that it is simple mathematically though :)

Shaders.

Most certainly you want to use a shader, but that doesn't fix all your problems. You have to push a (very large) texture across the bus. The GPU will chew on it and spit out the result, which you then have to pull back across the bus into main memory so that it plays nicely with other UI elements that are CPU-bound and not cached in VRAM.

So you've removed the processing bottleneck in exchange for a bus bandwidth bottleneck :(

are you sure it has to be pulled back to main memory?

Yeah, that doesn't make sense to me either. Especially since the entire point of the recent changes to iOS is precisely to get everything on-screen handled by the GPU, in the first place ..

As is factoring large numbers :)

Actually, in a naive implementation it is very computationally intensive, especially with large convolution cores (like on iOS 7 where it appears huge) and high resolutions.

Even optimized implementations on GPU are still computationally intensive compared to alpha blending.

Reply on news.ycombinator.com