[FFmpeg-devel] [PATCH] Port x264 SSE2 deblocking code to H.264 decoder

[Jason Garrett-Glaser]

On Fri, Dec 19, 2008 at 7:18 PM, Mike Melanson <mike at multimedia.cx> wrote:
> Jason Garrett-Glaser wrote:
>> On Thu, Dec 18, 2008 at 8:28 PM, Jason Garrett-Glaser
>> <darkshikari at gmail.com> wrote:
>>> On Thu, Dec 18, 2008 at 5:34 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
>>>> On Thu, Dec 18, 2008 at 05:22:51PM -0800, Jason Garrett-Glaser wrote:
>>>>> On Thu, Dec 18, 2008 at 5:09 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
>>>>>> On Thu, Dec 18, 2008 at 04:47:24PM -0800, Jason Garrett-Glaser wrote:
>>>>>>> OK, now we have luma_intra in DSPutil, so this should be easier.
>>>>>>>
>>>>>>> Michael: how should we rename the x264 functions?  My thought was just
>>>>>>> to s/x264/ff_h264/ or something of the sort, which would modify the
>>>>>>> code from x264's version but make it trivial to modify the code before
>>>>>>> committing any updates from x264.  It wouldn't need ugly #defines
>>>>>>> either.
>>>>>> didint loren post some patch that changed cglobal to add a prefix
>>>>>> automagically ...
>>>>> Shouldn't that be a separate patch?  If it's fine, can I commit his
>>>>> patch now then?
>>>> ive approved his patch, so yes you can apply it of course
>>> applied (sorry for forgetting credit, forgot to add Loren's name to
>>> commit message).
>>>
>>> Updated x264 deblock patch attached.
>>>
>>> Dark Shikari
>>>
>>
>> Small error in patch fixed.
>
> In case you did not see my response on ffmpeg-cvslog, this code causes
> the H.264 decoder to segfault (on about 90 of the conformance samples)
> when compiled with Intel's C Compiler, which we are sort of trying to
> support. I had time to look at it a little more but did not find much
> more useful information. I know that a build from a straight
> './configure --cc="icc"' config is fine, but not if --enable-gpl is also
> specified.
>
> Here comes the info dump, using the sample from this test spec:
> http://fate.multimedia.cx/index.php?test_spec=8

libavcodec requires an aligned stack on x86_32--however, as far as I
know, this is the first asm to be inserted into a very widely-used
location that explicitly requires that alignment.  If a compiler
cannot provide this, it cannot correctly compile libavcodec.

If this is a problem, perhaps we should revisit the issue of stack
alignment (and accordingly either implement x264's method of stack
alignment, or simply remove all SSE code from ffmpeg).

Dark Shikari