topic I'm glad I'm not the only one in Intel® oneAPI Math Kernel Library

MKL's cosine transform for strided data

Anonymous — Tue, 12 Jul 2011 19:31:00 GMT

MKL has a cosine transform among its Trigonometric Transforms (Chapter 13 of the manual). The cosine transform is significantly faster than that of FFTW.

However, there is no documented way to apply cosine transform to strided data. In particular, suppose

a b a b a b a b a b a b ....

is a sequence of doubles. I want to apply a cosine transform to the a's and to the b's. Is there a way to do that?

MKL's cosine transform for strided data

mecej4 — Tue, 12 Jul 2011 22:06:25 GMT

If you are calling from Fortran, you can use an array expression for the input argument.

In Fortran as well as C, you can, of course, extract the 'a's into one array, the 'b's into another and use such an array as the input argument.

You could put your data into a 2-D array:

a b
a b
a b

and pass each column to the FFT routine.

MKL's cosine transform for strided data

Anonymous — Tue, 12 Jul 2011 23:31:47 GMT

One can copy the interleaved data into two stride 1 arrays and use MKL, but that incurs a very significant penalty. If n=1024, copying will increase the cost of a single cosine transform (assuming data is not in cache but the copying area is in cache) by nearly 30%.

I'm glad I'm not the only one

Gerald_H_ — Fri, 17 Mar 2017 21:47:58 GMT

I'm glad I'm not the only one who is attempting to vectorize operations with interleaved data -- i.e. complex numbers.