[Ffmpeg-devel] [patch] another experiment w/ ffserver regression test

[Aurelien Jacobs]

On Fri, 3 Mar 2006 01:18:21 +0100
Michael Niedermayer <michaelni at gmx.at> wrote:

> Hi
> 
> On Fri, Mar 03, 2006 at 12:57:41AM +0100, Aurelien Jacobs wrote:
> > On Thu, 02 Mar 2006 21:48:03 +0100
> > Jindrich Makovicka <makovick at kmlinux.fjfi.cvut.cz> wrote:
> > 
> > > Aurelien Jacobs wrote:
> > > > On Wed, 01 Mar 2006 20:32:58 +0100
> > > > Jindrich Makovicka <makovick at kmlinux.fjfi.cvut.cz> wrote:
> > > > 
> > > > 
> > > >>Hi,
> > > >>
> > > >>could you please try the server-regression test with this patch? It adds
> > > >>the BitExact option to ffserver and fixes the test.conf file. My results
> > > >>of the test are also attached.
> > > > 
> > > > 
> > > > Well, this patch smells good.
> > > > The good news is that I tested it on amd64 and several x86 and got the
> > > > same md5 as you.
> > > > Now the bad news is that I tested it on ppc and it gives totally
> > > > different results :
> > > > 9b6ccb5cdd4f6677b79d0e06557fdb7e  ff-test_h.avi
> > > > c8bd4680156fa412635f011909eef081  ff-test_l.avi
> > > > b892f296c47d1b9a2b64edc502f6819b  ff-test.swf
> > > > b987e804ca459245f8fd572717bb7714  ff-test_h.asf
> > > > 25d55bb8e9d43ea2fb57d94af215f56e  ff-test_l.asf
> > > > fbe0a4ed463f5a589d4ba9f68fd4fae4  ff-test_h.rm
> > > > 45792412d2120b610824ce7e2ef9365a  ff-test_l.rm
> > > > e18ff0e88892c88549bab394fe28ef8d  ff-test.jpg
> > > > 228057fa69aededfd078e657ae12eb88  ff-test_small.jpg
> > > > 7bcccb92c3c9e38b5a9d54e44360bbb1  ff-test.mjpg
> > > > 
> > > > 
> > > >>I am just curious if it's possible to make a regression test which
> > > >>actually tests something :)
> > > > 
> > > > 
> > > > Indeed, with your patch, the regression tests seems much more useful ;-)
> > > > It really produces playable files, and results are consistents.
> > > > So I think it should be applied.
> > > > And then someone will have to dig into the ppc problem (which is maybe
> > > > an endianness issue somewhere).
> > > 
> > > Couldn't it be that PPC doesn't honor the bitexact flag? Can you run the
> > > test with Altivec optimizations disabled?
> > 
> > Hum... I did the test on a G3, so yes, altivec was disabled ;-)
> 
> you didnt forget:
> AVCodecContext.idct_algo= FF_IDCT_SIMPLE;
> AVCodecContext. dct_algo= FF_DCT_INT;

Well, forcing those flags in ffserver.c improves the situation. Now
there's only the .asf and .rm tests which gives different results on ppc.

BTW: the .asf produced by the ffserver regression tests don't seems to
be playable...

Aurel