I am joining a team which is working heavily with CUDA and am currently configuring my computer to run the code. It seems everything is working except for the cufftExecR2C call, which returns all 0 regardless of the input. I know the code I’m running the tests on is correct, as the same tests run correctly on every other persons computer here.
So my question is, what configuration settings could cause the FFT to fail on one computer when the code works for another? I’m passing all of the other tests, so most of CUDA and JCufft appears to be configured correctly.
It’s probably difficult to find the possible reasons for this behavior locally, on the respective machine, but it’s even more difficult to do this online… So far, I can only ask some questions to narrow down the problem space:
Is the driver up to date?
Are all libraries up to date (i.e. are the JCuda libraries and the installed CUDA version matching, and no other, older libraries present in the PATH, which may cause the wrong library to be loaded)?
Is there a noticable pattern concerning the hardware (i.e. the GPU) on which the program is working or not working?
Apart from that, it might be interesting how exactly you run the test. Did you set
JCuda.setExceptionsEnabled(true);
JCufft.setExceptionsEnabled(true);
to be informed about errors that might have occured before the call to cufftExecR2C (or, alternatively, did you check all return codes of all function calls)?