843 lines
		
	
	
		
			37 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
	
	
			
		
		
	
	
			843 lines
		
	
	
		
			37 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
	
	
|             FFTW FREQUENTLY ASKED QUESTIONS WITH ANSWERS
 | |
|                             14 Sep 2021
 | |
| 			     Matteo Frigo
 | |
| 			   Steven G. Johnson
 | |
|  			    <fftw@fftw.org>
 | |
| 
 | |
| This is the list of Frequently Asked Questions about FFTW, a collection of
 | |
| fast C routines for computing the Discrete Fourier Transform in one or
 | |
| more dimensions.
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Index
 | |
| 
 | |
|  Section 1.  Introduction and General Information
 | |
|  Q1.1        What is FFTW?
 | |
|  Q1.2        How do I obtain FFTW?
 | |
|  Q1.3        Is FFTW free software?
 | |
|  Q1.4        What is this about non-free licenses?
 | |
|  Q1.5        In the West? I thought MIT was in the East?
 | |
| 
 | |
|  Section 2.  Installing FFTW
 | |
|  Q2.1        Which systems does FFTW run on?
 | |
|  Q2.2        Does FFTW run on Windows?
 | |
|  Q2.3        My compiler has trouble with FFTW.
 | |
|  Q2.4        FFTW does not compile on Solaris, complaining about const.
 | |
|  Q2.5        What's the difference between --enable-3dnow and --enable-k7?
 | |
|  Q2.6        What's the difference between the fma and the non-fma versions?
 | |
|  Q2.7        Which language is FFTW written in?
 | |
|  Q2.8        Can I call FFTW from Fortran?
 | |
|  Q2.9        Can I call FFTW from C++?
 | |
|  Q2.10       Why isn't FFTW written in Fortran/C++?
 | |
|  Q2.11       How do I compile FFTW to run in single precision?
 | |
|  Q2.12       --enable-k7 does not work on x86-64
 | |
| 
 | |
|  Section 3.  Using FFTW
 | |
|  Q3.1        Why not support the FFTW 2 interface in FFTW 3?
 | |
|  Q3.2        Why do FFTW 3 plans encapsulate the input/output arrays and not ju
 | |
|  Q3.3        FFTW seems really slow.
 | |
|  Q3.4        FFTW slows down after repeated calls.
 | |
|  Q3.5        An FFTW routine is crashing when I call it.
 | |
|  Q3.6        My Fortran program crashes when calling FFTW.
 | |
|  Q3.7        FFTW gives results different from my old FFT.
 | |
|  Q3.8        FFTW gives different results between runs
 | |
|  Q3.9        Can I save FFTW's plans?
 | |
|  Q3.10       Why does your inverse transform return a scaled result?
 | |
|  Q3.11       How can I make FFTW put the origin (zero frequency) at the center 
 | |
|  Q3.12       How do I FFT an image/audio file in *foobar* format?
 | |
|  Q3.13       My program does not link (on Unix).
 | |
|  Q3.14       I included your header, but linking still fails.
 | |
|  Q3.15       My program crashes, complaining about stack space.
 | |
|  Q3.16       FFTW seems to have a memory leak.
 | |
|  Q3.17       The output of FFTW's transform is all zeros.
 | |
|  Q3.18       How do I call FFTW from the Microsoft language du jour?
 | |
|  Q3.19       Can I compute only a subset of the DFT outputs?
 | |
|  Q3.20       Can I use FFTW's routines for in-place and out-of-place matrix tra
 | |
| 
 | |
|  Section 4.  Internals of FFTW
 | |
|  Q4.1        How does FFTW work?
 | |
|  Q4.2        Why is FFTW so fast?
 | |
| 
 | |
|  Section 5.  Known bugs
 | |
|  Q5.1        FFTW 1.1 crashes in rfftwnd on Linux.
 | |
|  Q5.2        The MPI transforms in FFTW 1.2 give incorrect results/leak memory.
 | |
|  Q5.3        The test programs in FFTW 1.2.1 fail when I change FFTW to use sin
 | |
|  Q5.4        The test program in FFTW 1.2.1 fails for n > 46340.
 | |
|  Q5.5        The threaded code fails on Linux Redhat 5.0
 | |
|  Q5.6        FFTW 2.0's rfftwnd fails for rank > 1 transforms with a final dime
 | |
|  Q5.7        FFTW 2.0's complex transforms give the wrong results with prime fa
 | |
|  Q5.8        FFTW 2.1.1's MPI test programs crash with MPICH.
 | |
|  Q5.9        FFTW 2.1.2's multi-threaded transforms don't work on AIX.
 | |
|  Q5.10       FFTW 2.1.2's complex transforms give incorrect results for large p
 | |
|  Q5.11       FFTW 2.1.3's multi-threaded transforms don't give any speedup on S
 | |
|  Q5.12       FFTW 2.1.3 crashes on AIX.
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Section 1.  Introduction and General Information
 | |
| 
 | |
|  Q1.1        What is FFTW?
 | |
|  Q1.2        How do I obtain FFTW?
 | |
|  Q1.3        Is FFTW free software?
 | |
|  Q1.4        What is this about non-free licenses?
 | |
|  Q1.5        In the West? I thought MIT was in the East?
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 1.1.  What is FFTW?
 | |
| 
 | |
| FFTW is a free collection of fast C routines for computing the Discrete
 | |
| Fourier Transform in one or more dimensions.  It includes complex, real,
 | |
| symmetric, and parallel transforms, and can handle arbitrary array sizes
 | |
| efficiently.  FFTW is typically faster than other publically-available FFT
 | |
| implementations, and is even competitive with vendor-tuned libraries.
 | |
| (See our web page for extensive benchmarks.)  To achieve this performance,
 | |
| FFTW uses novel code-generation and runtime self-optimization techniques
 | |
| (along with many other tricks).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 1.2.  How do I obtain FFTW?
 | |
| 
 | |
| FFTW can be found at the FFTW web page.  You can also retrieve it from
 | |
| ftp.fftw.org in /pub/fftw.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 1.3.  Is FFTW free software?
 | |
| 
 | |
| Starting with version 1.3, FFTW is Free Software in the technical sense
 | |
| defined by the Free Software Foundation (see Categories of Free and
 | |
| Non-Free Software), and is distributed under the terms of the GNU General
 | |
| Public License.  Previous versions of FFTW were distributed without fee
 | |
| for noncommercial use, but were not technically ``free.''
 | |
| 
 | |
| Non-free licenses for FFTW are also available that permit different terms
 | |
| of use than the GPL.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 1.4.  What is this about non-free licenses?
 | |
| 
 | |
| The non-free licenses are for companies that wish to use FFTW in their
 | |
| products but are unwilling to release their software under the GPL (which
 | |
| would require them to release source code and allow free redistribution).
 | |
| Such users can purchase an unlimited-use license from MIT.  Contact us for
 | |
| more details.
 | |
| 
 | |
| We could instead have released FFTW under the LGPL, or even disallowed
 | |
| non-Free usage.  Suffice it to say, however, that MIT owns the copyright
 | |
| to FFTW and they only let us GPL it because we convinced them that it
 | |
| would neither affect their licensing revenue nor irritate existing
 | |
| licensees.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 1.5.  In the West? I thought MIT was in the East?
 | |
| 
 | |
| Not to an Italian.  You could say that we're a Spaghetti Western (with
 | |
| apologies to Sergio Leone).
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Section 2.  Installing FFTW
 | |
| 
 | |
|  Q2.1        Which systems does FFTW run on?
 | |
|  Q2.2        Does FFTW run on Windows?
 | |
|  Q2.3        My compiler has trouble with FFTW.
 | |
|  Q2.4        FFTW does not compile on Solaris, complaining about const.
 | |
|  Q2.5        What's the difference between --enable-3dnow and --enable-k7?
 | |
|  Q2.6        What's the difference between the fma and the non-fma versions?
 | |
|  Q2.7        Which language is FFTW written in?
 | |
|  Q2.8        Can I call FFTW from Fortran?
 | |
|  Q2.9        Can I call FFTW from C++?
 | |
|  Q2.10       Why isn't FFTW written in Fortran/C++?
 | |
|  Q2.11       How do I compile FFTW to run in single precision?
 | |
|  Q2.12       --enable-k7 does not work on x86-64
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.1.  Which systems does FFTW run on?
 | |
| 
 | |
| FFTW is written in ANSI C, and should work on any system with a decent C
 | |
| compiler.  (See also Q2.2 `Does FFTW run on Windows?', Q2.3 `My compiler
 | |
| has trouble with FFTW.'.) FFTW can also take advantage of certain
 | |
| hardware-specific features, such as cycle counters and SIMD instructions,
 | |
| but this is optional.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.2.  Does FFTW run on Windows?
 | |
| 
 | |
| Yes, many people have reported successfully using FFTW on Windows with
 | |
| various compilers.  FFTW was not developed on Windows, but the source code
 | |
| is essentially straight ANSI C.  See also the FFTW Windows installation
 | |
| notes, Q2.3 `My compiler has trouble with FFTW.', and Q3.18 `How do I call
 | |
| FFTW from the Microsoft language du jour?'.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.3.  My compiler has trouble with FFTW.
 | |
| 
 | |
| Complain fiercely to the vendor of the compiler.
 | |
| 
 | |
| We have successfully used gcc 3.2.x on x86 and PPC, a recent Compaq C
 | |
| compiler for Alpha, version 6 of IBM's xlc compiler for AIX, Intel's icc
 | |
| versions 5-7, and Sun WorkShop cc version 6.
 | |
| 
 | |
| FFTW is likely to push compilers to their limits, however, and several
 | |
| compiler bugs have been exposed by FFTW.  A partial list follows.
 | |
| 
 | |
| gcc 2.95.x for Solaris/SPARC produces incorrect code for the test program
 | |
| (workaround: recompile the libbench2 directory with -O2).
 | |
| 
 | |
| NetBSD/macppc 1.6 comes with a gcc version that also miscompiles the test
 | |
| program. (Please report a workaround if you know one.)
 | |
| 
 | |
| gcc 3.2.3 for ARM reportedly crashes during compilation.  This bug is
 | |
| reportedly fixed in later versions of gcc.
 | |
| 
 | |
| Versions 8.0 and 8.1 of Intel's icc falsely claim to be gcc, so you should
 | |
| specify CC="icc -no-gcc"; this is automatic in FFTW 3.1.  icc-8.0.066
 | |
| reportely produces incorrect code for FFTW 2.1.5, but is fixed in version
 | |
| 8.1.  icc-7.1 compiler build 20030402Z appears to produce incorrect
 | |
| dependencies, causing the compilation to fail.  icc-7.1 build 20030307Z
 | |
| appears to work fine.  (Use icc -V to check which build you have.)  As of
 | |
| 2003/04/18, build 20030402Z appears not to be available any longer on
 | |
| Intel's website, whereas the older build 20030307Z is available.
 | |
| 
 | |
| ranlib of GNU binutils 2.9.1 on Irix has been observed to corrupt the FFTW
 | |
| libraries, causing a link failure when FFTW is compiled.  Since ranlib is
 | |
| completely superfluous on Irix, we suggest deleting it from your system
 | |
| and replacing it with a symbolic link to /bin/echo.
 | |
| 
 | |
| If support for SIMD instructions is enabled in FFTW, further compiler
 | |
| problems may appear:
 | |
| 
 | |
| gcc 3.4.[0123] for x86 produces incorrect SSE2 code for FFTW when -O2 (the
 | |
| best choice for FFTW) is used, causing FFTW to crash (make check crashes).
 | |
| This bug is fixed in gcc 3.4.4.  On x86_64 (amd64/em64t), gcc 3.4.4
 | |
| reportedly still has a similar problem, but this is fixed as of gcc 3.4.6.
 | |
| 
 | |
| gcc-3.2 for x86 produces incorrect SIMD code if -O3 is used.  The same
 | |
| compiler produces incorrect SIMD code if no optimization is used, too.
 | |
| When using gcc-3.2, it is a good idea not to change the default CFLAGS
 | |
| selected by the configure script.
 | |
| 
 | |
| Some 3.0.x and 3.1.x versions of gcc on x86 may crash.  gcc so-called 2.96
 | |
| shipping with RedHat 7.3 crashes when compiling SIMD code.  In both cases,
 | |
| please upgrade to gcc-3.2 or later.
 | |
| 
 | |
| Intel's icc 6.0 misaligns SSE constants, but FFTW has a workaround. icc
 | |
| 8.x fails to compile FFTW 3.0.x because it falsely claims to be gcc; we
 | |
| believe this to be a bug in icc, but FFTW 3.1 has a workaround.
 | |
| 
 | |
| Visual C++ 2003 reportedly produces incorrect code for SSE/SSE2 when
 | |
| compiling FFTW.  This bug was reportedly fixed in VC++ 2005;
 | |
| alternatively, you could switch to the Intel compiler. VC++ 6.0 also
 | |
| reportedly produces incorrect code for the file reodft11e-r2hc-odd.c
 | |
| unless optimizations are disabled for that file.
 | |
| 
 | |
| gcc 2.95 on MacOS X miscompiles AltiVec code (fixed in later versions).
 | |
| gcc 3.2.x miscompiles AltiVec permutations, but FFTW has a workaround.
 | |
| gcc 4.0.1 on MacOS for Intel crashes when compiling FFTW; a workaround is
 | |
| to compile one file without optimization: cd kernel; make CFLAGS=" "
 | |
| trig.lo.
 | |
| 
 | |
| gcc 4.1.1 reportedly crashes when compiling FFTW for MIPS; the workaround
 | |
| is to compile the file it crashes on (t2_64.c) with a lower optimization
 | |
| level.
 | |
| 
 | |
| gcc versions 4.1.2 to 4.2.0 for x86 reportedly miscompile FFTW 3.1's test
 | |
| program, causing make check to crash (gcc bug #26528).  The bug was
 | |
| reportedly fixed in gcc version 4.2.1 and later.  A workaround is to
 | |
| compile libbench2/verify-lib.c without optimization.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.4.  FFTW does not compile on Solaris, complaining about const.
 | |
| 
 | |
| We know that at least on Solaris 2.5.x with Sun's compilers 4.2 you might
 | |
| get error messages from make such as
 | |
| 
 | |
| "./fftw.h", line 88: warning: const is a keyword in ANSI C
 | |
| 
 | |
| This is the case when the configure script reports that const does not
 | |
| work:
 | |
| 
 | |
| checking for working const... (cached) no
 | |
| 
 | |
| You should be aware that Solaris comes with two compilers, namely,
 | |
| /opt/SUNWspro/SC4.2/bin/cc and /usr/ucb/cc.  The latter compiler is
 | |
| non-ANSI.  Indeed, it is a perverse shell script that calls the real
 | |
| compiler in non-ANSI mode.  In order to compile FFTW, change your path so
 | |
| that the right cc is used.
 | |
| 
 | |
| To know whether your compiler is the right one,  type cc -V.  If the
 | |
| compiler prints ``ucbcc'', as in
 | |
| 
 | |
| ucbcc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2
 | |
| 
 | |
| then the compiler is wrong.  The right message is something like
 | |
| 
 | |
| cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.5.  What's the difference between --enable-3dnow and --enable-k7?
 | |
| 
 | |
| --enable-k7 enables 3DNow! instructions on K7 processors (AMD Athlon and
 | |
| its variants).  K7 support is provided by assembly routines generated by a
 | |
| special purpose compiler.  As of fftw-3.2, --enable-k7 is no longer
 | |
| supported.
 | |
| 
 | |
| --enable-3dnow enables generic 3DNow! support using gcc builtin functions.
 | |
| This works on earlier AMD processors, but it is not as fast as our special
 | |
| assembly routines.  As of fftw-3.1, --enable-3dnow is no longer supported.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.6.  What's the difference between the fma and the non-fma versions?
 | |
| 
 | |
| The fma version tries to exploit the fused multiply-add instructions
 | |
| implemented in many processors such as PowerPC, ia-64, and MIPS.  The two
 | |
| FFTW packages are otherwise identical.  In FFTW 3.1, the fma and non-fma
 | |
| versions were merged together into a single package, and the configure
 | |
| script attempts to automatically guess which version to use.
 | |
| 
 | |
| The FFTW 3.1 configure script enables fma by default on PowerPC, Itanium,
 | |
| and PA-RISC, and disables it otherwise.  You can force one or the other by
 | |
| using the --enable-fma or --disable-fma flag for configure.
 | |
| 
 | |
| Definitely use fma if you have a PowerPC-based system with gcc (or IBM
 | |
| xlc).  This includes all GNU/Linux systems for PowerPC and the older
 | |
| PowerPC-based MacOS systems.  Also use it on PA-RISC and Itanium with the
 | |
| HP/UX compiler.
 | |
| 
 | |
| Definitely do not use the fma version if you have an ia-32 processor
 | |
| (Intel, AMD, MacOS on Intel, etcetera).
 | |
| 
 | |
| For other architectures/compilers, the situation is not so clear.  For
 | |
| example, ia-64 has the fma instruction, but gcc-3.2 appears not to exploit
 | |
| it correctly.  Other compilers may do the right thing, but we have not
 | |
| tried them.  Please send us your feedback so that we can update this FAQ
 | |
| entry.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.7.  Which language is FFTW written in?
 | |
| 
 | |
| FFTW is written in ANSI C.  Most of the code, however, was automatically
 | |
| generated by a program called genfft, written in the Objective Caml
 | |
| dialect of ML.  You do not need to know ML or to have an Objective Caml
 | |
| compiler in order to use FFTW.
 | |
| 
 | |
| genfft is provided with the FFTW sources, which means that you can play
 | |
| with the code generator if you want.  In this case, you need a working
 | |
| Objective Caml system.  Objective Caml is available from the Caml web
 | |
| page.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.8.  Can I call FFTW from Fortran?
 | |
| 
 | |
| Yes, FFTW (versions 1.3 and higher) contains a Fortran-callable interface,
 | |
| documented in the FFTW manual.
 | |
| 
 | |
| By default, FFTW configures its Fortran interface to work with the first
 | |
| compiler it finds, e.g. g77.  To configure for a different, incompatible
 | |
| Fortran compiler foobar, use ./configure F77=foobar when installing FFTW.
 | |
| (In the case of g77, however, FFTW 3.x also includes an extra set of
 | |
| Fortran-callable routines with one less underscore at the end of
 | |
| identifiers, which should cover most other Fortran compilers on Linux at
 | |
| least.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.9.  Can I call FFTW from C++?
 | |
| 
 | |
| Most definitely.  FFTW should compile and/or link under any C++ compiler.
 | |
| Moreover, it is likely that the C++ <complex> template class is
 | |
| bit-compatible with FFTW's complex-number format (see the FFTW manual for
 | |
| more details).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.10.  Why isn't FFTW written in Fortran/C++?
 | |
| 
 | |
| Because we don't like those languages, and neither approaches the
 | |
| portability of C.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.11.  How do I compile FFTW to run in single precision?
 | |
| 
 | |
| On a Unix system: configure --enable-float.  On a non-Unix system: edit
 | |
| config.h to #define the symbol FFTW_SINGLE (for FFTW 3.x).  In both cases,
 | |
| you must then recompile FFTW.  In FFTW 3, all FFTW identifiers will then
 | |
| begin with fftwf_ instead of fftw_.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 2.12.  --enable-k7 does not work on x86-64
 | |
| 
 | |
| Support for --enable-k7 was discontinued in fftw-3.2.
 | |
| 
 | |
| The fftw-3.1 release supports --enable-k7.  This option only works on
 | |
| 32-bit x86 machines that implement 3DNow!, including the AMD Athlon and
 | |
| the AMD Opteron in 32-bit mode.  --enable-k7 does not work on AMD Opteron
 | |
| in 64-bit mode.  Use --enable-sse for x86-64 machines.
 | |
| 
 | |
| FFTW supports 3DNow! by means of assembly code generated by a
 | |
| special-purpose compiler.  It is hard to produce assembly code that works
 | |
| in both 32-bit and 64-bit mode.
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Section 3.  Using FFTW
 | |
| 
 | |
|  Q3.1        Why not support the FFTW 2 interface in FFTW 3?
 | |
|  Q3.2        Why do FFTW 3 plans encapsulate the input/output arrays and not ju
 | |
|  Q3.3        FFTW seems really slow.
 | |
|  Q3.4        FFTW slows down after repeated calls.
 | |
|  Q3.5        An FFTW routine is crashing when I call it.
 | |
|  Q3.6        My Fortran program crashes when calling FFTW.
 | |
|  Q3.7        FFTW gives results different from my old FFT.
 | |
|  Q3.8        FFTW gives different results between runs
 | |
|  Q3.9        Can I save FFTW's plans?
 | |
|  Q3.10       Why does your inverse transform return a scaled result?
 | |
|  Q3.11       How can I make FFTW put the origin (zero frequency) at the center 
 | |
|  Q3.12       How do I FFT an image/audio file in *foobar* format?
 | |
|  Q3.13       My program does not link (on Unix).
 | |
|  Q3.14       I included your header, but linking still fails.
 | |
|  Q3.15       My program crashes, complaining about stack space.
 | |
|  Q3.16       FFTW seems to have a memory leak.
 | |
|  Q3.17       The output of FFTW's transform is all zeros.
 | |
|  Q3.18       How do I call FFTW from the Microsoft language du jour?
 | |
|  Q3.19       Can I compute only a subset of the DFT outputs?
 | |
|  Q3.20       Can I use FFTW's routines for in-place and out-of-place matrix tra
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.1.  Why not support the FFTW 2 interface in FFTW 3?
 | |
| 
 | |
| FFTW 3 has semantics incompatible with earlier versions: its plans can
 | |
| only be used for a given stride, multiplicity, and other characteristics
 | |
| of the input and output arrays; these stronger semantics are necessary for
 | |
| performance reasons.  Thus, it is impossible to efficiently emulate the
 | |
| older interface (whose plans can be used for any transform of the same
 | |
| size).  We believe that it should be possible to upgrade most programs
 | |
| without any difficulty, however.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.2.  Why do FFTW 3 plans encapsulate the input/output arrays and not just the algorithm?
 | |
| 
 | |
| There are several reasons:
 | |
| 
 | |
| * It was important for performance reasons that the plan be specific to
 | |
|   array characteristics like the stride (and alignment, for SIMD), and
 | |
|   requiring that the user maintain these invariants is error prone.
 | |
| * In most high-performance applications, as far as we can tell, you are
 | |
|   usually transforming the same array over and over, so FFTW's semantics
 | |
|   should not be a burden.
 | |
| * If you need to transform another array of the same size, creating a new
 | |
|   plan once the first exists is a cheap operation.
 | |
| * If you need to transform many arrays of the same size at once, you
 | |
|   should really use the plan_many routines in FFTW's "advanced" interface.
 | |
| * If the abovementioned array characteristics are the same, you are
 | |
|   willing to pay close attention to the documentation, and you really need
 | |
|   to, we provide a "new-array execution" interface to apply a plan to a
 | |
|   new array.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.3.  FFTW seems really slow.
 | |
| 
 | |
| You are probably recreating the plan before every transform, rather than
 | |
| creating it once and reusing it for all transforms of the same size.  FFTW
 | |
| is designed to be used in the following way:
 | |
| 
 | |
| * First, you create a plan.  This will take several seconds.
 | |
| * Then, you reuse the plan many times to perform FFTs.  These are fast.
 | |
| 
 | |
| If you don't need to compute many transforms and the time for the planner
 | |
| is significant, you have two options.  First, you can use the
 | |
| FFTW_ESTIMATE option in the planner, which uses heuristics instead of
 | |
| runtime measurements and produces a good plan in a short time.  Second,
 | |
| you can use the wisdom feature to precompute the plan; see Q3.9 `Can I
 | |
| save FFTW's plans?'
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.4.  FFTW slows down after repeated calls.
 | |
| 
 | |
| Probably, NaNs or similar are creeping into your data, and the slowdown is
 | |
| due to the resulting floating-point exceptions.  For example, be aware
 | |
| that repeatedly FFTing the same array is a diverging process (because FFTW
 | |
| computes the unnormalized transform).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.5.  An FFTW routine is crashing when I call it.
 | |
| 
 | |
| Did the FFTW test programs pass (make check, or cd tests; make bigcheck if
 | |
| you want to be paranoid)?  If so, you almost certainly have a bug in your
 | |
| own code.  For example, you could be passing invalid arguments (such as
 | |
| wrongly-sized arrays) to FFTW, or you could simply have memory corruption
 | |
| elsewhere in your program that causes random crashes later on.  Please
 | |
| don't complain to us unless you can come up with a minimal self-contained
 | |
| program (preferably under 30 lines) that illustrates the problem.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.6.  My Fortran program crashes when calling FFTW.
 | |
| 
 | |
| As described in the manual, on 64-bit machines you must store the plans in
 | |
| variables large enough to hold a pointer, for example integer*8.  We
 | |
| recommend using integer*8 on 32-bit machines as well, to simplify porting.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.7.  FFTW gives results different from my old FFT.
 | |
| 
 | |
| People follow many different conventions for the DFT, and you should be
 | |
| sure to know the ones that we use (described in the FFTW manual).  In
 | |
| particular, you should be aware that the FFTW_FORWARD/FFTW_BACKWARD
 | |
| directions correspond to signs of -1/+1 in the exponent of the DFT
 | |
| definition.  (*Numerical Recipes* uses the opposite convention.)
 | |
| 
 | |
| You should also know that we compute an unnormalized transform.  In
 | |
| contrast, Matlab is an example of program that computes a normalized
 | |
| transform.  See Q3.10 `Why does your inverse transform return a scaled
 | |
| result?'.
 | |
| 
 | |
| Finally, note that floating-point arithmetic is not exact, so different
 | |
| FFT algorithms will give slightly different results (on the order of the
 | |
| numerical accuracy; typically a fractional difference of 1e-15 or so in
 | |
| double precision).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.8.  FFTW gives different results between runs
 | |
| 
 | |
| If you use FFTW_MEASURE or FFTW_PATIENT mode, then the algorithm FFTW
 | |
| employs is not deterministic: it depends on runtime performance
 | |
| measurements.  This will cause the results to vary slightly from run to
 | |
| run.  However, the differences should be slight, on the order of the
 | |
| floating-point precision, and therefore should have no practical impact on
 | |
| most applications.
 | |
| 
 | |
| If you use saved plans (wisdom) or FFTW_ESTIMATE mode, however, then the
 | |
| algorithm is deterministic and the results should be identical between
 | |
| runs.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.9.  Can I save FFTW's plans?
 | |
| 
 | |
| Yes. Starting with version 1.2, FFTW provides the wisdom mechanism for
 | |
| saving plans; see the FFTW manual.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.10.  Why does your inverse transform return a scaled result?
 | |
| 
 | |
| Computing the forward transform followed by the backward transform (or
 | |
| vice versa) yields the original array scaled by the size of the array.
 | |
| (For multi-dimensional transforms, the size of the array is the product of
 | |
| the dimensions.)  We could, instead, have chosen a normalization that
 | |
| would have returned the unscaled array. Or, to accomodate the many
 | |
| conventions in this matter, the transform routines could have accepted a
 | |
| "scale factor" parameter. We did not do this, however, for two reasons.
 | |
| First, we didn't want to sacrifice performance in the common case where
 | |
| the scale factor is 1. Second, in real applications the FFT is followed or
 | |
| preceded by some computation on the data, into which the scale factor can
 | |
| typically be absorbed at little or no cost.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.11.  How can I make FFTW put the origin (zero frequency) at the center of its output?
 | |
| 
 | |
| For human viewing of a spectrum, it is often convenient to put the origin
 | |
| in frequency space at the center of the output array, rather than in the
 | |
| zero-th element (the default in FFTW).  If all of the dimensions of your
 | |
| array are even, you can accomplish this by simply multiplying each element
 | |
| of the input array by (-1)^(i + j + ...), where i, j, etcetera are the
 | |
| indices of the element.  (This trick is a general property of the DFT, and
 | |
| is not specific to FFTW.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.12.  How do I FFT an image/audio file in *foobar* format?
 | |
| 
 | |
| FFTW performs an FFT on an array of floating-point values.  You can
 | |
| certainly use it to compute the transform of an image or audio stream, but
 | |
| you are responsible for figuring out your data format and converting it to
 | |
| the form FFTW requires.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.13.  My program does not link (on Unix).
 | |
| 
 | |
| The libraries must be listed in the correct order (-lfftw3 -lm for FFTW
 | |
| 3.x) and *after* your program sources/objects.  (The general rule is that
 | |
| if *A* uses *B*, then *A* must be listed before *B* in the link command.).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.14.  I included your header, but linking still fails.
 | |
| 
 | |
| You're a C++ programmer, aren't you?  You have to compile the FFTW library
 | |
| and link it into your program, not just #include <fftw3.h>.  (Yes, this is
 | |
| really a FAQ.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.15.  My program crashes, complaining about stack space.
 | |
| 
 | |
| You cannot declare large arrays with automatic storage (e.g. via
 | |
| fftw_complex array[N]); you should use fftw_malloc (or equivalent) to
 | |
| allocate the arrays you want to transform if they are larger than a few
 | |
| hundred elements.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.16.  FFTW seems to have a memory leak.
 | |
| 
 | |
| After you create a plan, FFTW caches the information required to quickly
 | |
| recreate the plan.  (See Q3.9 `Can I save FFTW's plans?') It also
 | |
| maintains a small amount of other persistent memory.  You can deallocate
 | |
| all of FFTW's internally allocated memory, if you wish, by calling
 | |
| fftw_cleanup(), as documented in the manual.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.17.  The output of FFTW's transform is all zeros.
 | |
| 
 | |
| You should initialize your input array *after* creating the plan, unless
 | |
| you use FFTW_ESTIMATE: planning with FFTW_MEASURE or FFTW_PATIENT
 | |
| overwrites the input/output arrays, as described in the manual.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.18.  How do I call FFTW from the Microsoft language du jour?
 | |
| 
 | |
| Please *do not* ask us Windows-specific questions.  We do not use Windows.
 | |
| We know nothing about Visual Basic, Visual C++, or .NET.  Please find the
 | |
| appropriate Usenet discussion group and ask your question there.  See also
 | |
| Q2.2 `Does FFTW run on Windows?'.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.19.  Can I compute only a subset of the DFT outputs?
 | |
| 
 | |
| In general, no, an FFT intrinsically computes all outputs from all inputs.
 | |
| In principle, there is something called a *pruned FFT* that can do what
 | |
| you want, but to compute K outputs out of N the complexity is in general
 | |
| O(N log K) instead of O(N log N), thus saving only a small additive factor
 | |
| in the log.  (The same argument holds if you instead have only K nonzero
 | |
| inputs.)
 | |
| 
 | |
| There are some specific cases in which you can get the O(N log K)
 | |
| performance benefits easily, however, by combining a few ordinary FFTs.
 | |
| In particular, the case where you want the first K outputs, where K
 | |
| divides N, can be handled by performing N/K transforms of size K and then
 | |
| summing the outputs multiplied by appropriate phase factors.  For more
 | |
| details, see pruned FFTs with FFTW.
 | |
| 
 | |
| There are also some algorithms that compute pruned transforms
 | |
| *approximately*, but they are beyond the scope of this FAQ.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 3.20.  Can I use FFTW's routines for in-place and out-of-place matrix transposition?
 | |
| 
 | |
| You can use the FFTW guru interface to create a rank-0 transform of vector
 | |
| rank 2 where the vector strides are transposed.  (A rank-0 transform is
 | |
| equivalent to a 1D transform of size 1, which.  just copies the input into
 | |
| the output.)  Specifying the same location for the input and output makes
 | |
| the transpose in-place.
 | |
| 
 | |
| For double-valued data stored in row-major format, plan creation looks
 | |
| like this:
 | |
| 
 | |
| fftw_plan plan_transpose(int rows, int cols, double *in, double *out)
 | |
| {
 | |
|     const unsigned flags = FFTW_ESTIMATE; /* other flags are possible */
 | |
|     fftw_iodim howmany_dims[2];
 | |
| 
 | |
|     howmany_dims[0].n  = rows;
 | |
|     howmany_dims[0].is = cols;
 | |
|     howmany_dims[0].os = 1;
 | |
| 
 | |
|     howmany_dims[1].n  = cols;
 | |
|     howmany_dims[1].is = 1;
 | |
|     howmany_dims[1].os = rows;
 | |
| 
 | |
|     return fftw_plan_guru_r2r(/*rank=*/ 0, /*dims=*/ NULL,
 | |
|                               /*howmany_rank=*/ 2, howmany_dims,
 | |
|                               in, out, /*kind=*/ NULL, flags);
 | |
| }
 | |
| (This entry was written by Rhys Ulerich.)
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Section 4.  Internals of FFTW
 | |
| 
 | |
|  Q4.1        How does FFTW work?
 | |
|  Q4.2        Why is FFTW so fast?
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 4.1.  How does FFTW work?
 | |
| 
 | |
| The innovation (if it can be so called) in FFTW consists in having a
 | |
| variety of composable *solvers*, representing different FFT algorithms and
 | |
| implementation strategies, whose combination into a particular *plan* for
 | |
| a given size can be determined at runtime according to the characteristics
 | |
| of your machine/compiler.  This peculiar software architecture allows FFTW
 | |
| to adapt itself to almost any machine.
 | |
| 
 | |
| For more details (albeit somewhat outdated), see the paper "FFTW: An
 | |
| Adaptive Software Architecture for the FFT", by M. Frigo and S. G.
 | |
| Johnson, *Proc. ICASSP* 3, 1381 (1998), also available at the FFTW web
 | |
| page.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 4.2.  Why is FFTW so fast?
 | |
| 
 | |
| This is a complex question, and there is no simple answer.  In fact, the
 | |
| authors do not fully know the answer, either.  In addition to many small
 | |
| performance hacks throughout FFTW, there are three general reasons for
 | |
| FFTW's speed.
 | |
| 
 | |
| * 	FFTW uses a variety of FFT algorithms and implementation styles that
 | |
|   can be arbitrarily composed to adapt itself to a machine.  See Q4.1 `How
 | |
|   does FFTW work?'.
 | |
| * 	FFTW uses a code generator to produce highly-optimized routines for
 | |
|   computing small transforms.
 | |
| * 	FFTW uses explicit divide-and-conquer to take advantage of the memory
 | |
|   hierarchy.
 | |
| 
 | |
| For more details (albeit somewhat outdated), see the paper "FFTW: An
 | |
| Adaptive Software Architecture for the FFT", by M. Frigo and S. G.
 | |
| Johnson, *Proc. ICASSP* 3, 1381 (1998), available along with other
 | |
| references at the FFTW web page.
 | |
| 
 | |
| ===============================================================================
 | |
| 
 | |
| Section 5.  Known bugs
 | |
| 
 | |
|  Q5.1        FFTW 1.1 crashes in rfftwnd on Linux.
 | |
|  Q5.2        The MPI transforms in FFTW 1.2 give incorrect results/leak memory.
 | |
|  Q5.3        The test programs in FFTW 1.2.1 fail when I change FFTW to use sin
 | |
|  Q5.4        The test program in FFTW 1.2.1 fails for n > 46340.
 | |
|  Q5.5        The threaded code fails on Linux Redhat 5.0
 | |
|  Q5.6        FFTW 2.0's rfftwnd fails for rank > 1 transforms with a final dime
 | |
|  Q5.7        FFTW 2.0's complex transforms give the wrong results with prime fa
 | |
|  Q5.8        FFTW 2.1.1's MPI test programs crash with MPICH.
 | |
|  Q5.9        FFTW 2.1.2's multi-threaded transforms don't work on AIX.
 | |
|  Q5.10       FFTW 2.1.2's complex transforms give incorrect results for large p
 | |
|  Q5.11       FFTW 2.1.3's multi-threaded transforms don't give any speedup on S
 | |
|  Q5.12       FFTW 2.1.3 crashes on AIX.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.1.  FFTW 1.1 crashes in rfftwnd on Linux.
 | |
| 
 | |
| This bug was fixed in FFTW 1.2.  There was a bug in rfftwnd causing an
 | |
| incorrect amount of memory to be allocated.  The bug showed up in Linux
 | |
| with libc-5.3.12 (and nowhere else that we know of).
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.2.  The MPI transforms in FFTW 1.2 give incorrect results/leak memory.
 | |
| 
 | |
| These bugs were corrected in FFTW 1.2.1.  The MPI transforms (really, just
 | |
| the transpose routines) in FFTW 1.2 had bugs that could cause errors in
 | |
| some situations.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.3.  The test programs in FFTW 1.2.1 fail when I change FFTW to use single precision.
 | |
| 
 | |
| This bug was fixed in FFTW 1.3.  (Older versions of FFTW did work in
 | |
| single precision, but the test programs didn't--the error tolerances in
 | |
| the tests were set for double precision.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.4.  The test program in FFTW 1.2.1 fails for n > 46340.
 | |
| 
 | |
| This bug was fixed in FFTW 1.3.  FFTW 1.2.1 produced the right answer, but
 | |
| the test program was wrong.  For large n, n*n in the naive transform that
 | |
| we used for comparison overflows 32 bit integer precision, breaking the
 | |
| test.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.5.  The threaded code fails on Linux Redhat 5.0
 | |
| 
 | |
| We had problems with glibc-2.0.5.  The code should work with glibc-2.0.7.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.6.  FFTW 2.0's rfftwnd fails for rank > 1 transforms with a final dimension >= 65536.
 | |
| 
 | |
| This bug was fixed in FFTW 2.0.1.  (There was a 32-bit integer overflow
 | |
| due to a poorly-parenthesized expression.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.7.  FFTW 2.0's complex transforms give the wrong results with prime factors 17 to 97.
 | |
| 
 | |
| There was a bug in the complex transforms that could cause incorrect
 | |
| results under (hopefully rare) circumstances for lengths with
 | |
| intermediate-size prime factors (17-97).  This bug was fixed in FFTW
 | |
| 2.1.1.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.8.  FFTW 2.1.1's MPI test programs crash with MPICH.
 | |
| 
 | |
| This bug was fixed in FFTW 2.1.2.  The 2.1/2.1.1 MPI test programs crashed
 | |
| when using the MPICH implementation of MPI with the ch_p4 device (TCP/IP);
 | |
| the transforms themselves worked fine.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.9.  FFTW 2.1.2's multi-threaded transforms don't work on AIX.
 | |
| 
 | |
| This bug was fixed in FFTW 2.1.3.  The multi-threaded transforms in
 | |
| previous versions didn't work with AIX's pthreads implementation, which
 | |
| idiosyncratically creates threads in detached (non-joinable) mode by
 | |
| default.
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.10.  FFTW 2.1.2's complex transforms give incorrect results for large prime sizes.
 | |
| 
 | |
| This bug was fixed in FFTW 2.1.3.  FFTW's complex-transform algorithm for
 | |
| prime sizes (in versions 2.0 to 2.1.2) had an integer overflow problem
 | |
| that caused incorrect results for many primes greater than 32768 (on
 | |
| 32-bit machines).  (Sizes without large prime factors are not affected.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.11.  FFTW 2.1.3's multi-threaded transforms don't give any speedup on Solaris.
 | |
| 
 | |
| This bug was fixed in FFTW 2.1.4.  (By default, Solaris creates threads
 | |
| that do not parallelize over multiple processors, so one has to request
 | |
| the proper behavior specifically.)
 | |
| 
 | |
| -------------------------------------------------------------------------------
 | |
| 
 | |
| Question 5.12.  FFTW 2.1.3 crashes on AIX.
 | |
| 
 | |
| The FFTW 2.1.3 configure script picked incorrect compiler flags for the
 | |
| xlc compiler on newer IBM processors.  This is fixed in FFTW 2.1.4.
 | |
| 
 | 
