my kernel does some complex computation and blocks the gpu for quite some time, thus making the normal GUI almost unusable; i’ve tested that on Linux, MacOS and Windows. I instantly loose lot of performance when I reduce the size of the input buffer in order to reduce the single-call-runtime and free some time for the GUI so this is not really an option.
My question now is if there is a way to schedule the kernel calls into idle gpu time or detect if e.g. no directx or opengl applications are running; I could fall back to cpu-only computation in that case.
Somehow the nVidia System Tools 6.0.2 (free download from nVidia) allow you to display GPU usage in %. So there must be an API to determine this. However I wonder if this API is public.