How feasible is CUDA support for GPU acceleration???

[mjv2]

DSS seems like a software that could benefit massively from parallelized processing, and given that so many people have relatively powerful consumer grade GPUs, it could really cut down processing time for amateurs and pros alike. I'm sure it would be a huge undertaking but I would be happy to help in whatever ways I can.

[perdrix52]

We already make use of parallel processing in the form of OpenMP and also exploit SIMD. There's a lot of I/O as well as processing involved, so offloading the data to a GPU may not make sense.

I agree if would be a pretty huge undertaking. If you are serious and have CUDA skills already feel free to create a fork and try some experiments. I suspect the gain might not be a large as you hope for.

[mjv2]

I see, was just curious as I've been learning it at university. One of the main topics covered is about throughput, bandwith limitation, utilizing shared memory, etc. and so far I've only seen CUDA accelerate somewhat specific bottleneck scenarios, so I see what you mean. If file I/O and similar things are responsible for most of the processing time there's not much parallelization would do, but if there are significant delays from running a lot of DSP like convolutions that are still sequential, I would be interested in looking further into those parts of the code.