[Ffmpeg-devel] [BUG] regression tests failure in r7867

[Guillaume Poirier]

Hi,

Diego Biurrun wrote:
> On Fri, Feb 09, 2007 at 11:20:27PM +0100, Guillaume POIRIER wrote:
> 
>>On 2/7/07, Baptiste Coudurier <baptiste.coudurier at smartjog.com> wrote:
>>
>>>M?ns Rullg?rd wrote:
>>>
>>>>Benoit Fouet said:
>>>>All the tests (except ffserver) passed last night on my x86-64.  That 
>>>
>>>was before
>>>
>>>>the pixfmt checkin though, which is why I suggested it might be the 
>>>
>>>culprit.
>>>
>>>No problem here, I just fixed some libav problems.
>>
>>Just tonight, on PPC, with r7903, the regression tests are as follow:
> 
> 
> The failure starts with r7965 where the WMA regression test was added.

That version number can't be right. Current version of ffmpeg is 7950.
Did you use IPOT (IP over Time) to fetch your svn copy of ffmpeg? ;o)
Ok, I guess you meant 7865? ;o)
Anyway, I think you're right about the fact that it's the WMA regtests
that "broke" regression tests on PPC.
I agree with Michael's analysis, it's gotta be a rounding problem.
For the record, the regtests also fail on x86-64:
--- ./data/ffmpeg.regression    2007-02-12 17:02:24.000000000 +0100
+++
/passauvegarde/dori/gpoirier/snapshot/ffmpeg/tests/ffmpeg.regression.ref
   2007-02-12 16:56:23.000000000 +0100
@@ -190,15 +190,15 @@
 stddev: 33.31 PSNR:65.87 bytes:1040384
 768305cec342e01b7a9d38e971ef0174 *./data/a-wmav1.asf
 99596 ./data/a-wmav1.asf
-2a36f36c883da068630be6c8fbfb8243 *./data/out.wav
+2ab193df22bb473e61015e56ac11faf3 *./data/out.wav
 stddev:12251.50 PSNR:14.56 bytes:1056768
 stddev:2106.00 PSNR:29.85 bytes:1048576
 c706c7cf9db10df7a7353a40feddcfdd *./data/a-wmav2.asf
 99602 ./data/a-wmav2.asf
-6b6283520caf395f55f2822e2bbae70f *./data/out.wav
+8d05d345bc14153f3d0da95fdf34e084 *./data/out.wav
 stddev:12255.92 PSNR:14.55 bytes:1056768
 stddev:2099.31 PSNR:29.88 bytes:1048576
-3e12dbc3672b3e17af98f18eb39e9da2 *./data/a-vorbis.asf
+ef725cb9c13b9437a41428eabadeba52 *./data/a-vorbis.asf
 35686 ./data/a-vorbis.asf
-07777d0f623aa346fc4383c8bd014e7f *./data/out.wav
+e46368ef818f2ec3abcc3f8390d14df5 *./data/out.wav
 stddev:3882.46 PSNR:24.54 bytes:1056768

Regression test: Error.
make[1]: *** [codectest] Error 1

Which makes sense since on x86-32, the x87 floating point stack is
used for FP computations, whereas on x86-64 all FP computation are
done with scalar SSE.
I bet both units have a different way to round things internally
(IFAIR, x87 keeps more "significant" bits internally).

Guillaume