I have some CUDA code that I want to use inside a plugin into another application, which itself is not CUDA enabled.
Unfortunately, if I link with and use cudart.dll, it can’t be found when my plugin is loaded into the host application.
It sounds like what I want is to use the dynamic linked driver api (as used in the matrix demo). I tried this approach, but I still get unresolved symbols from my compiled .cu code.
Note that I need my plugin to run even if a CUDA-enabled driver is not present, and I can’t get away with sprinkling CUDA DLLs through the file-system. If I distribute any extra files, they need to be along-side my plugin.
Is there an easier way to do this?