I used an optimized FFTW3, but I just tried using FFTW (-D_FFTW) and got the same results:
Total grid elements on this process: 16200 Error < 1e-18: 9515 Error < 1e-15: 14191 Error < 1e-12: 14191 Error < 1e-09: 14191 Error < 1e-06: 14191 Error < 1e-03: 14586 Error < 1e+00: 16200 Error >= 1e+0: 0 So most of the elements have small errors (which is as expected), but there are a large number which have very large errors (>1e-06), and I don't understand why this should happen. Here is a typical sample of elements that don't agree after the G->R->G FFT steps: ( 0.0000000000000000 , 0.0000000000000000 ) -> (-6.19446221078556020E-004,-1.74898798599905760E-004) ( 0.0000000000000000 , 0.0000000000000000 ) -> ( 9.03523919712864205E-005, 9.20886154392430039E-004) ( 0.0000000000000000 , 0.0000000000000000 ) -> ( 3.07966800915878281E-004, 1.44523973744239645E-004) ( 0.0000000000000000 , 0.0000000000000000 ) -> (-1.71711886309486506E-004, 3.97860015055829619E-004) ( 0.0000000000000000 , 0.0000000000000000 ) -> (-4.92121624328117836E-004,-5.62768457197758659E-004) ( 0.0000000000000000 , 0.0000000000000000 ) -> ( 6.02880566549388744E-005,-2.14744832649235777E-004) I don't think this can be put down to numerical error in the FFT? - Iain -- Iain Bethune Applications Consultant, EPCC Email: ibethune at epcc.ed.ac.uk Tel/Fax: +44 (0)131 650 5201/6555 Mob: +44 (0)7598317015 Addr: 2404 JCMB, King's Buildings, Mayfield Road, Edinburgh, EH9 3JZ The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336. > -----Original Message----- > From: Lorenzo Paulatto [mailto:paulatto at sissa.it] > Sent: 13 January 2010 16:34 > To: ibethune at epcc.ed.ac.uk; PWSCF Forum > Subject: Re: [Pw_forum] Errors in parallel 3D FFT > > On Wed, 13 Jan 2010 16:51:20 +0100, Iain Bethune > <ibethune at staffmail.ed.ac.uk> wrote: > > Any idea what might be causing this? > > Hi Iain, > are you using "vanilla" FFTW libraries or architecture optimized ones? > (e.g. ACML, ESSL) Some high performance libraries can trade a little of > accuracy for speed, and to be honest. I'm not even sure that perfect > matching (whatever the definition of perfect) is guaranteed even with the > standard open-source FFTW libraries. > > best regards > > -- > Lorenzo Paulatto > > *** Note: my affiliation has changed! please send future > correspondence to: <Lorenzo.Paulatto at impmc.upmc.fr> *** > > post-doc @ IMPMC/UPMC - Universit? Paris 6 > phone: +33 (0)1 44 27 74 89 > www: http://www-int.impmc.upmc.fr/~paulatto/ > > previously: > phd student @ SISSA & DEMOCRITOS (Trieste) > phone: +39 040 3787 511 > www: http://people.sissa.it/~paulatto/
