(08-13-2012, 01:00 AM)xoft Wrote: You might not believe it, but I have considered that as soon as I started writing the new generator architecture. I didn't know anything about GPU programming at that time, so I did a little research, and the results are not good: There is no good standard that would be generally compatible with every gfx card. CUDA only works with nVidia cards and ATI has some other API. Now, we want to run on other devices, too - typically, internet servers in ISP server-housing don't have gfx cards at all. Raspberry Pi has a gfx chip, but doesn't provide the programming API. And so forth. So there's just no way to make this solution universal. Hence, it's not worth writing (yet).
I think the overall performance is good enough for now, so that we don't need to go this way immediately.
OpenCL should work on both ATI and nVidia. It's a pain to work with and I think the only change in performance will be from lighting and generating. Managing what area to (re-)light and what stages of generating to do is already confusing (to me at least XD) doing it with GPGPU will be even more confusing haha. Performance could even be less, because you wouldn't only be copying from CPU cache and RAM, but also from VRAM which is much further away from the CPU/RAM