I’m getting some funny issues with CUDA 5.5 toolkit used with the drivers that come with CUDA 7.0, I can’t remember the exact version but I’ll get it if it’ll help with any analysis/commentary you guys may have.
Essentially the problem is no matter what I do if I try to use one device for FFT related stuff and the other for custom kernels the FFT’s invariably fail with a message that says the plan wasn’t created on the right card. Each card driver is its own thread in the same process. I ran cudaGetDevice before/after planning and when calling execute and they come out to the correct device number. I did not have this problem with using multiple K40s in the same manner.
Has anyone seen this type of issue before? If so, is there a known workaround?
Thanks for any help!