[SciPy-dev] fftpack2 for review
eric at enthought.com
Tue Oct 1 23:50:14 CDT 2002
> On Tue, 1 Oct 2002, eric jones wrote:
> > Hey Pearu,
> > Overall, first tests went fairly smoothly. Even without djbfft and
> > fftw, things appeared to install correctly on win32. *very* nice
> > system checking and handling the lack of libraries.
> > All tests passed. Sure am glad to finally have some unit tests for
> > stuff!
> > I do not have fftw installed. I tested once without djbfft and once
> > with it. I didn't notice any material difference in timings (the
> > non-djbfft set given below). My tests are on a 850 MHz PIII and
> > are slower than yours. This leads me to believe, even when trying
> > use djbfft, I'm not getting it linked correctly. Setup said it
> > my djbfft libraries though (I defined the environment variable
> > to my libs to get this to work).
> Note that measuring time is currently reliable only on Linux platform
> where real jiffies [*] are used. Under windows and other unices either
> real jiffies are not available or I don't know how to use them, and
> time.time is used instead.
> [*] jiffies are used to measure how much the given process uses CPU
> unit is 1/100th of a second. Under Linux one can get user jiffies and
> system jiffies, in scipy bench tests the last ones are ignored when
> measuring performance of algorithms.
> > Also freqshift puts the Nyquist component at the end of the result
> > while fftshift puts it at the beginning of the result. This is for
> > the consistency among varios functions in the fftpack module.
> Let me add here that it would be good for performance not to fix the
> storage order and provide helper functions such as dftfreq function
> return this information, also provide functions for manipulating fft
> results. For example, djbfft uses a rather strange ordering (that may
> the key for its good performance) and to transform the results from
> routines to "standard" convention (and back), two nontrivial copying
> arrays is needed. And this may be the performance killer when using
> the djbfft library.
> If no one will object this, I'll put it in to my todo list.
> But it is fine with me to stick to one and only one convention if
> really want that.
Hmmm. I think it is fine to stick with the current behavior for now.
Having multiple output formats is fairly tricky to deal with.
Well... I guess it worth asking now. Does anyone need absolute speed
with fft so much that they are willing to call some obscure function
name and then deal with the ordering themselves? Speak now or forever
hold your peace. (Or add it on your own later...)
> > And now some installation notes about what I did to get djbfft
> > Install notes:
> > 1.
> > win32 (cygwin and mingw) required #include "errno.h" in any djbfft
> > that used errno.
> So, we need to patch djbfft. Probably D.J.B. does not like this.
Yes. I haven't ever gotten a response from him, so I think the chances
> > 2.
> > I installed the latest f2py but still got the warning error:
> > fftpack2 requires F2PY version 2.23.190-1367 or higher but got
> > 2.23.190-1354
> > I'm hitting the hay right now, so I'll check out why this happened
> > tomorrow.
> That is strange. I would expect that Python will crash (with
> segfault) when running convolve tests if older F2PY version than
> 2.23.190-1367 was used for building. Any change that you have several
> f2py2e packages lying around in your system?
I don't think so, but I'll look into it again tonight.
> > 3.
> > How hard would it be to get setup.py to build the djbfft library if
> > wasn't detected. The makefile builds a lot of the files on the fly
> > using stuff like:
> > 4c0.c: \
> > idea/c0.c pentium/c0.c ppro/c0.c sparc/c0.c auto_opt
> > sed 1s/PRE/pre4.c/ `cat auto_opt`/c0.c > 4c0.c
> > 4c0.o: \
> > compile 4c0.c pre4.c fftc4.h complex4.h real4.h fftr4.h real4.h 4i.c
> > ./compile 4c0.c
> > I'm not sure what all is going on here, but it seems like we could
> > either do the same in python or auto-generate a generic version of
> > stuff and include it in SciPy. Of course, since djbfft isn't
> > this isn't necessary.
> In principle, this could be done. However, D.J.B. certainly has made
> more difficult for us to accomplish that because djbfft uses a rather
> unique building process, and also the conditions of altering this
> software are rather restrictive.
If we do it within our own CVS tree, I think it should be Ok with him,
but I'm also not sure it is worth it.
> I think that having setup.py to build djbfft library does not have a
> high priority right now, especially if the performance boost is not so
> great on all platforms and arhitectures. So, I would put it into the
> list and deal with it in future.
More information about the Scipy-dev