I am using CUFFT. I have found that in my application an in place 1d 1024 point C2R (513 complex values generating a 1024 point real output) is giving me numerically imprecise results when I select CUFFT_COMPATIBILITY_NATIVE mode. However, when I switch to CUFFT_COMPATIBILITY_FFTW_ASYMMETRIC mode then the results are reliable. Does this make any sense? Can you describe why CUFFT_COMPATIBILITY_FFTW_ASYMMETRIC mode would have an affect on a real C2R transform in which I would think conjugate symmetry would be guaranteed by definition.