topic MKL performance w/regard to transform length in Intel® oneAPI Math Kernel Library

MKL performance w/regard to transform length

mike_p_1 — Thu, 29 Jan 2015 15:23:52 GMT

on https://software.intel.com/en-us/intel-mkl/details

The description of MKL perforce states:

"Fast Fourier Transforms
Intel MKL FFTs include many optimizations and should provide significant performance gains over other libraries ****** for medium and large transform sizes. ******" (emphasis mine).

What is the expected value of "medium"? I am comparing performance of MKL FFT against FFTPACK 5.1 and am seeing a break-even value that *I* think is quite large; way bigger than I expected.

My benchmark includes setup overhead for both MKL and FFTPACK and does a forward/backward transform. The retrieved data series is virtually the same as input.

Thanks.

Hello Mike,

Devorah_H_Intel — Thu, 29 Jan 2015 21:42:47 GMT

Hello Mike,

Can you please send us your test results in private message. I will investigate your question and get back to you with update.

Thank you,

Devorah

Hi Mike,

Dmitry_B_Intel — Fri, 30 Jan 2015 13:17:06 GMT

Hi Mike,

Performance of MKL FFT is measured without setup overhead.

Medium sizes of a 1D FFT are approximately 100...100k. A size becomes 'large' when the working data set does not fit in last level cache, which may vary depending on precision and placement of transform, parallelization, and the computer used.

Thanks
Dima