I’ve written a filter for ffmpeg that needs to look at over 80mln pixels for each frame. This seems like something that CUDA would help with, specifically on Amazon’s EC2 Linux GPU cluster box. I got 3.8x the bandwith using OpenMP on a 4-core i7, so maybe CUDA would be even better. I ran the profiler, and for this task, 90+% of the processing time is in my filter.
I’ve seen many people asking about CUDA for decoding with ffmpeg and many other people explaining the problems using GPU power for that, but has anyone figured out how to tweak or avoid the Makefile to enable CUDA in the filters library? On Linux?