[FFmpeg-devel] [PATCH] VC-1: Faster C loopfilter

Jason Garrett-Glaser darkshikari
Sun Jan 18 08:15:04 CET 2009


On Sun, Jan 18, 2009 at 2:12 AM, Jason Garrett-Glaser
<darkshikari at gmail.com> wrote:
> On Sun, Jan 18, 2009 at 3:00 AM, Kostya <kostya.shishkov at gmail.com> wrote:
>> On Sun, Jan 18, 2009 at 12:48:41AM -0500, David Conrad wrote:
>>> On Jan 18, 2009, at 12:39 AM, Jason Garrett-Glaser wrote:
>>>
>>> >On Sat, Jan 17, 2009 at 11:54 PM, Jason Garrett-Glaser
>>> ><darkshikari at gmail.com> wrote:
>>> >>Yuvi requested I optimize this one, so I did.  This patch can be
>>> >>summed up as follows:
>>> >>
>>> >>13:20 <@mru> I see a lot of xor
>>> >>13:20 <@mru> that's usually a sign of magic
>>> >>
>>> >>338 -> 315 cycles for vc1_loop_filter on Core 2 Duo.
>>> >>Overall decoding 12.047s -> 11.781s
>>> >>
>>> >>Dark Shikari
>>> >>
>>> >
>>> >Also, it seems forcing inlining of vc1_filter_line improves
>>> >performance greatly (315 -> 275 cycles from the above); is it fine to
>>> >av_always_inline it?
>>>
>>> I am working on mmx/sse versions you know :P
>>> Anyways you should most of the inlining benefits with separate h/v and
>>> size=4/8/16 functions (h/v much more so than 4/8/16), e.g. with the
>>> attached (I'll split/update it once your optimizations hit svn)
>>
>> Disabling loop filter at all does not hurt anything and gives more speedup
>> than Dark Shikari's patch :P.
>>
>> Anyway, I have no objections against this patch as well.
>
> Fixed version attached, since I forgot about ffmpeg bracket rules on
> one line, apparently.
>
> Dark Shikari
>

Eh, since it was approved by Kostya, I just went ahead and applied it.

Dark Shikari




More information about the ffmpeg-devel mailing list