Welcome to Fractal Forums

As I mentioned in my other thread where I posted a video tour of the Buddhabrot (http://www.fractalforums.com/movies-showcase-%28rate-my-movie%29/detailed-buddhabrot-exploration/ (http://www.fractalforums.com/movies-showcase-%28rate-my-movie%29/detailed-buddhabrot-exploration/)), I have been working on my own Buddhabrot rendering software for a while.  I was originally focusing on high quality, detailed, stills but took a break from that to work on some animation features.

Here are some stills that I find particularly compelling.  I call his "the Infinity Fields."  I think the first one is more aesthetic but the second one conveys the idea of "infinity" a little better.

(http://www.aluminumstudios.com/photography/photoblog/wmilberry-infinity_fields02.jpg)

(http://www.aluminumstudios.com/photography/photoblog/milberry-infinity_fields_01.jpg)

wouaaaaaaaaaaaaaaaaaaaaaaaaah  ;D

On 2nd tought :

Part me of me want to keep working on gpu stuff (see : http://www.fractalforums.com/programming/mandelbrot-in-xna-plus-hlsl/ )
And the other part of me want to code a buddhabrot renderer ... again. Because your rendering is so much better than mine.

Mmm ....
Well... I'll see if i can manage to render a buddhabrot using OpenCL (or cuda, or Direct Compute, or any GPGPU stuff).

The GPU is not supposed to be efficient with that kind of computation, but let's try ! :)

The speed problem is with the scattered read-modify-writes to global memory. I am thinking mainly about the CUDA architecture now:

Wouldn't it be faster to use the 16kb of shared memory (64 kb on Fermi) as some kind of independent "mini framebuffer" tiles, and accumulating the writes within shared memory only - as if one rendered a lot of independent deep zooms? Techniques to render Buddhabrot zooms exist (with appropriate non-uniform sampling optimizations). And applying these techniques to individual tiles that make up a larger image might just work. Every multiprocessor would get to work on its own tile (shared memory is individual to each multiprocessor). For those tiles that finish rendering, the multiprocessor will immediately get to work on another tile (on Fermi at least). One would not need any writes to global memory - speeding up the process by an order of magnitude maybe -  until the very end when the completed tile is written out.

Christian

Mmmmmmmmm.... interesting idea  ;D

Thanks for the "woooahhh!" :)

I'm hoping to focus on feedback about the images in this post.  Maybe we should start a thread in the programming forum about it?

here : http://www.fractalforums.com/programming/buddhabrot-on-gpu/

those look amazing :D

Does your current setup only alow 2nd or nth power bbrots or does it allow aribitary formulae?

I'm at work right now, so I can't post any examples, but I can use arbitrary formula including Julia style if I wanted.  Some formulas give really interesting images too.  I will post some soon.

try tan(x+y*i)(x+y*i)+a+b*i - it has a lot of amazing details :)