Issue/bug with 1D FFT/IFFT along third dimension of 3D matrix when FFTW_MEASURE or FFTW_PATIENT are used (FFTW_ESTIMATE works)

[fftw]

Issue/bug with 1D FFT/IFFT along third dimension of 3D matrix when FFTW_MEASURE or FFTW_PATIENT are used (FFTW_ESTIMATE works)

[rdolbeau]

I'm not sure the way you use fftw_execute_split_dft() is legit. It is quite sensitive to data placement. The comment in the source file is:
"/* guru interface: requires care in alignment, r - i, etcetera. */"

In particular, planning & executing on the same arrays in your code seems to work. Separate allocation with a fixed offset between real_* and imag_* (i.e. single allocation, fixed offsets) works as well (but only for forward or backward, you would need 2 plans).

[stevengj]

Yes, the split interface assumes that the re - im offset does not change between planning and execution. If you want to apply it to a different pair of arrays, the best way is to do a single fftw_malloc for both real and imaginary parts (with the first half being real and the second half being imaginary).

Thanks for looking at this, @rdolbeau

[stevengj]

In particular, the manual states the following requirement:

• For split arrays, the separations between the real and imaginary parts, ii-ri and io-ro, are the same as they were for the input and output arrays when the plan was created. (This condition is automatically satisfied for interleaved arrays.)

[mlgill]

Thank you both for the response. I read the manual thoroughly, but I didn't realize what the statement in question meant. I'll let you know if I have further issues.

[mlgill]

As a quick followup, everything is now working correctly, so thank you again. I did have to create a separate IFFT plan for the scenario I'm using to correctly reflect the relative location in memory of real/imaginary data when their position is swapped in the execute command.