FS#44886 - FFTW accuracy issues

Attached to Project: Arch Linux
Opened by Ruben Van Boxem (rubenvb) - Thursday, 07 May 2015, 09:20 GMT
Last edited by Doug Newgard (Scimmia) - Thursday, 07 May 2015, 13:27 GMT
Task Type Bug Report
Category Packages: Extra
Status Closed
Assigned To No-one
Architecture x86_64
Severity Medium
Priority Normal
Reported Version
Due in Version Undecided
Due Date Undecided
Percent Complete 100%
Votes 1
Private No

Details

Description:
The FFTW transforms show a much larger error than should be when calculated on Arch Linux.

Additional info:
* fftw-3.3.4
* source code attached comparing to-and-fro transform to the original.

Steps to reproduce:
Compile and run this code with
g++ -O2 main.cpp -lfftw3 -o fftw_test && ./fftw_test

This produces an error of around 1e-8. Running the exact same code on Windows, using MinGW-w64 GCC and the FFTW library installable from within MSYS2, the error is of the order 1e-17, which is more like what I expect, looking at FFTW's own accuracy benchmarks:
http://www.fftw.org/accuracy/
I don't have other Linux distro's installed, but they should show the 1e-17 order of magnitude as well.
   main.cpp (1 KiB)
This task depends upon

Closed by  Doug Newgard (Scimmia)
Thursday, 07 May 2015, 13:27 GMT
Reason for closing:  Duplicate
Additional comments about closing:   FS#44888 
Comment by Ruben Van Boxem (rubenvb) - Thursday, 07 May 2015, 11:42 GMT
Hmm, I may have been bitten by the low RAND_MAX value on Windows.

If I use C++11 random number generators, the errors match regardless of the platform.

As an apology, please accept an improved version of the FFTW PKGBUILD:
https://bugs.archlinux.org/task/44888

Loading...