CUFFT accuracy in EMULATIOn

Hi all,

i’m performig a 2D cufft transform in an emulation mode with a CUDA cufft call. Starting from the same input data, I compared the result from cuFFT and matlab fft2 and the difference is quite unacceptable. Indeed, in some points there is a difference of 15% in the value, while maximum values are identical.

I’m searching on internet about the accuracy of cuFFT but is seems that in my case it should be very accurate as input size are small (8x8 matrix) and powers of 2, so I guessed that the problem could be maybe the EMULATION. Unfortunately, I do not have available a GPU in this moment to check it out directly and I’m wondering if this is a known issue.

Does anybody can help me in understanding why there is a so evident difference.