I haven't looked into it, but I know that when I worked at Philips people there have looked into it for image processing. What I'm writing here is what I understood from their presentations.
1. The result of calculations is not exact and can be different on different hardware. Since audio is (at least) 16 bit and images are 8, I expect problems here.
2. Every time when you need to do some processing that's not suitable for GPU processing, you need to go back and forth between CPU and GPU, which - in case of video cards - means that you need to send the data through the PCI(e) connector. Which is very inefficient.
3. Result of tests with one complex algorithm was that there was no gain in performance when using GPU instead of CPU. This might be because the people who did the test didn't understand what they were doing...
Anyway, I think it will cause more problems than it will solve. On modern Intel CPU's, the CPU load of Stereo Tool with everything enabled (full FM preset, composite clipping, Stokkemask filtering) is around 6%. So there's not much reason to do this.
|