[FFmpeg-devel] [PATCH 8/9] x86: dct-test: add 10 bits versions.

Michael Niedermayer michael at niedermayer.cc
Sat Oct 10 19:10:27 CEST 2015


On Sat, Oct 10, 2015 at 06:14:13PM +0200, Christophe Gisquet wrote:
> Hi,
> 
> 2015-10-10 13:20 GMT+02:00 Michael Niedermayer <michael at niedermayer.cc>:
> > On Fri, Oct 09, 2015 at 11:53:45PM +0200, Christophe Gisquet wrote:
> >> ---
> >>  libavcodec/x86/dct-test.c | 13 ++++++++++++-
> >>  1 file changed, 12 insertions(+), 1 deletion(-)
> >
> > with this patch
> > libavcodec/dct-test -i 0 10
> > end in "Error: 1." after testing SIMPLE10-C
> 
> Well, the source of the error is omse > 0.02. What a pity that
> simple10-c was implemented but fails that (omse=0.03060703). The x86
> implementation simply tries to be bit-exact to it, so if there's
> something to fix, that's the C version first.

indeed, i must have mixed the 9bit (which doesnt produce an error)
with 10bit


> 
> We can probably get one more bit of precision from it by changing the
> scaling (dct-test passing with omse=0.01663750 and no change to PSNR
> in dnxhd fate tests although the hashes change), but I would assume
> the persons that chose those coefficients had reasons... And that
> would break prores and its extra shift.

i would be in favor of inproving the precission but this is indeed
unrelated to the patchset and there indeed may be reasons against it
(i dont know)
maybe the threshold should be adjusted for 10bit or the specific idct
though


> 
> > libavcodec/dct-test  0 10
> > tests IJG-LLM-INT10:which looks like it fails the test
> 
> Huh? Even on master, I don't see a IJG-LLM-INT10 being tested when I
> run this command. IJG-LLM-INT is indeed ran, but fails utterly.

Iam just confused that the tested idcts are changed by the patch
from a set of mostly failing and some passing ones to a single new
one that fails


> 
> > the FAAN dct would pass and was tested previously but is not tested
> > anymore (this was possibly caused by the previous patch not this one)
> 
> Yes it passes, and the patch should be so that this doesn't change.
> 
> -- 
> Christophe
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

I do not agree with what you have to say, but I'll defend to the death your
right to say it. -- Voltaire
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20151010/697e0aec/attachment.sig>


More information about the ffmpeg-devel mailing list