[FFmpeg-cvslog] r14698 - in trunk/libavcodec: Makefile dsputil.h fft.c i386/fft_3dn.c i386/fft_3dn2.c i386/fft_mmx.asm i386/fft_sse.c

[Michael Niedermayer]

On Wed, Aug 13, 2008 at 07:35:49PM -0700, Mike Melanson wrote:
> Michael Niedermayer wrote:
> > On Thu, Aug 14, 2008 at 12:40:57AM +0200, Michael Niedermayer wrote:
> >> On Wed, Aug 13, 2008 at 01:15:27AM -0600, Loren Merritt wrote:
> >>> On Tue, 12 Aug 2008, Mike Melanson wrote:
> >>>>> Author: lorenm
> >>>>> Date: Tue Aug 12 02:26:58 2008
> >>>>> New Revision: 14698
> >>>>>
> >>>>> Log:
> >>>>> split-radix FFT
> >>>>> c is 1.9x faster than previous c (on various x86 cpus), sse is 1.6x faster than previous sse.
> >>>> This change broke the WMA v1 encoding test on the following configurations:
> >>>>
> >>>> x86_32 / gcc 4.3.1
> >>>> x86_32 / gcc 4.4.0-SVN
> >>>>
> >>>> The failure looks like this:
> >>>>
> >>>> diff -u -w "/home/melanson/fate/source"/tests/ffmpeg.regression.ref
> >>>> tests/data/vsynth.regression
> >>>> --- /home/melanson/fate/source/tests/ffmpeg.regression.ref
> >>>> 2008-08-08 16:34:02.000000000 -0700
> >>>> +++ tests/data/vsynth.regression        2008-08-12 18:48:02.000000000
> >>>> -0700
> >>>> @@ -205,7 +205,7 @@
> >>>>  353368 ./tests/data/a-flac.flac
> >>>>  95e54b261530a1bcf6de6fe3b21dc5f6 *./tests/data/flac.vsynth.out.wav
> >>>>  stddev:    0.00 PSNR:999.99 bytes:  1058444/  1058444
> >>>> -4435d87463cd6c5407bd88cca241ca56 *./tests/data/a-wmav1.asf
> >>>> +6f7f3116b801ea641ce14f827de05cce *./tests/data/a-wmav1.asf
> >>>>  106004 ./tests/data/a-wmav1.asf
> >>>>  stddev:12251.38 PSNR: 14.56 bytes:  1056812/  1058444
> >>>>  stddev: 2105.96 PSNR: 29.85 bytes:  1048620/  1058444
> >>>> make: *** [codectest] Error 1
> >>> I have no idea what's special about those configurations. For that matter, 
> >>> why didn't all the other mdct codecs and all the other configurations 
> >>> change? 
> >>> And how could anyone expect md5 matches for floating-point codecs?
> >> wasnt expected, it just happened to work out by luck well enough to be
> >> useable as a regression test
> > 
> > replacing float functions like the fft/mdct by a identity function if
> > CODEC_FLAG_BITEXACT is set would be welcome.
> 
> In the meantime, perhaps we should remove the test from the regression 
> suite?

have you tried simply disabling the SIMD cases when CODEC_FLAG_BITEXACT is
set? maybe that would be enough ...

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Opposition brings concord. Out of discord comes the fairest harmony.
-- Heraclitus
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-cvslog/attachments/20080814/5e9da2ad/attachment.pgp>