topic Re: FTTW3 wrapper question in Intel® oneAPI Math Kernel Library

FTTW3 wrapper question

michFBK — Mon, 30 Jun 2025 15:26:53 GMT

Hello,

I use oneMKL to calculate FFTs. The program originally was written for FFTW and with version 2024.0 and 2024.1 it runs smoothly using the included wrappers.

The plan is created with this instruction

fftw_plan_FFT_Mon = fftw_plan_dft_r2c_1d(Nmonitor, pfftw_in_FFT_Mon, pfftw_out_FFT_Mon, FFTW_ESTIMATE);

Nmonitor is a power of 2 from 2^10 to 2^19. The program compiles and runs with both FFTW and the 2024.0 and 2024.1 releases. I use MingW64 under Windows 10.

From release 2025.1 (and likewise 2025.2) the program compiles but the fftw_plan_dft_r2c_1d instruction works only up to Nmonitor==2^12. It crashes when the program runs with Nmonitor>=2^13.
I have casually observed that if the program starts in a directory with mkl_intel_thread.2.dll version 2024.1 (instead of the correct one from release 2025.1 or 2025.2), the problem does not occur.

What is the correct data type for the size of the FFT (Nmonitor in my case)? Did it change from release 2024.1 to 2025.1?

Re: FTTW3 wrapper question

Fengrui — Wed, 02 Jul 2025 03:03:59 GMT

Hi,

I tried the in-package example dp_plan_dft_r2c_1d.c and changed N to 2^13. With oneMKL 2025.1, the test passed without issues. Could you share a code example that can reproduce the crash?

Thanks,

Fengrui

Re: FTTW3 wrapper question

michFBK — Wed, 02 Jul 2025 14:59:02 GMT

Hi Fengrui,
I compiled the in-package example and find the same error. Here I report the output of the example in the two cases:

//// RELEASE 2024.1
Example dp_plan_dft_r2c_1d
1D real-to-complex out-of-place transform
Configuration parameters:
N = 8192
H = -1
Allocate data arrays: real x(8192), complex y(4097)
Create FFTW plan for 1D r2c out-of-place FFT
Initialize input for r2c transform
Compute r2c FFT
Verify the result of forward FFT
Check if err is below errthr 7.22e-15
Verified, maximum error was 9.2e-17
Destroy FFTW plan
Free data arrays
TEST PASSED

Process returned 0 (0x0) execution time : 0.044 s
Press any key to continue.
////

//// RELEASE 2025.1
Example dp_plan_dft_r2c_1d
1D real-to-complex out-of-place transform
Configuration parameters:
N = 8192
H = -1
Allocate data arrays: real x(8192), complex y(4097)
Create FFTW plan for 1D r2c out-of-place FFT

Process returned -1066598274 (0xC06D007E) execution time : 1.481 s
Press any key to continue.
////

A few observations:
-Release 2025 works with N<=4096
-Release 2025 works for odd values of N, even if greater than 4096
-The example works for all values of N when compiled with FFTW3.

Do you have any suggestions for me to check on my compiler?

Thank you

Michele

Re: FTTW3 wrapper question

Fengrui — Wed, 02 Jul 2025 18:55:20 GMT

Hi Michele,

The test I did was in Linux. Please note that Mingw-w64 is not listed in the oneMKL system requirements. But I think it is worth trying to build the fftw3 wrappers with icx 2025.1 and see if that works.

Thanks,

Fengrui