Hi,

>-----Original Message-----
>From: ext lisa yue [mailto:lisayue...@gmail.com]
>Sent: 29 November, 2010 10:10
>To: MeeGo-dev@meego.com
>Cc: Kost Stefan (Nokia-MS/Helsinki)
>Subject: SSE optimization for gstreamer audio resample element
>
>Hi, group
>
>I am looking at Meego bug #5603:
>
>Bug 5603 - SSE misalignment removal in gstreamer base plugin
>"audioresample"
>
>(http://bugs.meego.com/show_bug.cgi?id=5603)
>
>It's about removing the SSE misalignment in gstreamer audio resample
>element.
>
>Since the input array can't ensure 16-byte aligned, when _USE_SSE is
>configured,
>
>_mm_loadu_ps() will be used instead of _mm_load_ps() to load from array.
>
>And _mm_loadu_ps should be the main cause for high CPU usage according
>to BMC#5603.

I totally agree with the analysis. I would suggest to open an upstream bug 
(bugzilla.gnome.org) and file the patches there. The code in question is used 
in gstreamer and in speex. So we should also ensure that the same fix is done 
in both places.

>
>The prototyping codes suggest removing misalignment for oversample=1 and
>2, this will reduce the hot function overhead by about 60%.
>
>Of course the disadvantage of removing the mis-alignment is to increase
>the code complexity a lot.
>

The code complexity increase in expected and jin in this case.

>
>I plan to optimize gstreamer audio resample code now. And since this is
>a common issue for speex, ffmpeg, pulseaudio etc, I wonder what's the
>suggested method to increase SSE optimization utilization for all of
>them~

We are using ORC in several gstreamer plugins already. Using orc in more places 
and having a good orc backend for SSE3 would imho be the way to go. ORC can 
produce code to handle with misaligned data. As a starting point, e.g. volume, 
audioconvert are two audio elements that use orc already (just search for *.orc 
files in gstreamer sources).

Stefan

>
>I googled and found someone asked about SSE optimization for moblin
>audio before but I can't see the cached content.
>
>Welcome your comments!
>
>
>
>Regards,
>
>Lisa
>
>

_______________________________________________
MeeGo-dev mailing list
MeeGo-dev@meego.com
http://lists.meego.com/listinfo/meego-dev

Reply via email to