Potential cuFFT Precision Regression in 418.67

Today, the NVIDA 418.67 driver became available to me.

In an FFT of size 8096 using cuComplex type, the precision of the output has been reduced in a significant way. Using numpy’s allclose we were testing a knowing input signal, the rtol parameter absolute(a - b) <= (atol + rtol * absolute(b)) used to function at 1e-05. And now will only function with relative tolerance of 1e-04.

Is this known and expected behavior? Or is there an issue in our usage?

The CUFFT library itself doesn’t change when a new driver comes out. It’s not impossible for a driver to have an effect, but unlikely to be noticeable.

On the other hand CUDA 10.1 Update 1 did also become available (which depends on this newer driver 418.67) so if you actually installed CUDA 10.0 Update 1, it brought a new version of CUFFT library with it. There were definitely changes in CUFFT library from CUDA 10.1 to CUDA 10.1 Update 1.

I doubt this is known and expected behavior. I couldn’t comment whether there is an issue in your usage, since I have no idea what your usage is. If you can provide a complete test case, that would certainly help. In that case, you could file a bug if you wish. The instructions to do so are linked at the top of the CUDA programming forum.

Thank you Robert. I will do further analysis to confirm if an issue in fact exists, and if so will follow the guidelines documented here: https://devtalk.nvidia.com/default/topic/1044668/cuda-programming-and-performance/-how-to-report-a-bug/