topic fftwf_plan_r2r_2d failing - expects rank of 1? in Intel® oneAPI Math Kernel Library

fftwf_plan_r2r_2d failing - expects rank of 1?

stele — Thu, 07 Apr 2011 16:14:07 GMT

I am evaluating the Intel C++ compiler and MKL and have run into a snag. I have some existing fftw code that uses fftwf_plan_r2r_2d() to do an in-place FFT on non-power-of-two image data.

Using MKL fftwf_plan_r2r_2d is returning null. I brought in the wrapper code so I could step into it and the problem is in fftw_plan_guru64_r2r() where it expects a rank of 1.

If fftwf_plan_r2r_2d() handles rank 2 data and it delegates down to fftw_plan_guru64_r2r(), shouldn't that function handle rank 2d data?

I confess I'm a complete moron when it comes to FFT stuff - I'm just using some code I inherited. But it was working with the real fftw library.

Any tips?

fftwf_plan_r2r_2d failing - expects rank of 1?

Dmitry_B_Intel — Fri, 08 Apr 2011 00:44:23 GMT

Hi Paul,

Thank you for trying MKL. FFTW3 interfaces that MKL provides do not support multidimensional r2r transforms.

Dima

fftwf_plan_r2r_2d failing - expects rank of 1?

stele — Fri, 08 Apr 2011 21:29:06 GMT

Thanks for the confirming that Dmitry. Can you tell me what an alternative way to do a 2D non-power-of-two transform and back - a rough equivalent to fftwf_plan_r2r_2d, that MKL does support? I was using that function to do an in-place cosine only transform.

fftwf_plan_r2r_2d failing - expects rank of 1?

Dmitry_B_Intel — Tue, 19 Apr 2011 12:54:31 GMT

Paul,

To date MKL supports only a limited set ofr2rtransforms, specifically 1D stride-1 one-transform-per-call sine and cosine transforms. An option to do 2D cosine with MKL is to implement row-col algorithm using these 1D transforms and transposition functions.

Thanks
Dima