GPU FFT2 Discrepancies ( Different from CPU based FFT2 )

Question

Kevin Blank 2013-10-22

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/91024-gpu-fft2-discrepancies-different-from-cpu-based-fft2

评论： Kevin Blank 2013-10-22

Hello, I have noticed a problem that Matlab FFT2 does not produce the same result on a GPU array as it does on a "normal" array.

The following code was used to ensure that my observation is correct.

totalDiff = 0;
for loop = 0:10000
      a = rand( [2 2] ); % Obtain random numbers
      cpuResult = fft2(a); % Calculate on CPU
      gpuResult = gather( fft2( gpuArray(a) ) ); % Calculate on GPU
      diff = sum( gpuResult(:) - cpuResult(:) ); % Calculate difference
      totalDiff = totalDiff + diff; % Get a running total of differences
end
totalDiff / 10000 %Output average change per matrix

I appreciate any feedback regarding this question: Why is this changing result being observed and can it be minimized or removed? While these values are not extremely different (differing by 1e-17 or so), I would still like to ensure an exact equality between my GPU and CPU code.

P.S. fft also has changing results based on the device used for computation; but, for my system, the matrix has to be larger than 4 x 4 to observe the difference.