[gst-devel] TI OMAP DSP support

[Kalle Vahlman]

2008/1/17, g <the_ether2 at yahoo.co.uk>:
> As previously mentioned, a TI article specifically
> states that there is a GStreamer plug-in that uses the
> DSP and yet I cannot find any mention of its existance
> on the GStreamer site.

Anyone can write a GStreamer plugin to do whatever task they need to,
it doesn't mean it will be available from the GStreamer site.

> --- Josep Torra Valles <j.torra at telefonica.net> wrote:
> > I also wrote some assembly to accelerate video
> > decoding on ARM5e and
> > ARM6 cores and I think that probably on ARM6(400
> > MHz) you can achieve
> > more video decoding speed than using the DSP(133
> > MHz).
>
> The clock frequency isn't the only part. What matters
> is what work is done per cycle. The DSP obviously has
> an important part to play. That's why it's there in
> the first place. Given the choice, doing all the
> processing purely on the ARM core and not using the
> DSP for any task is silly - especially for video
> decoding.

What is perhaps not obvious is that whatever TI promises for it's SOC
isn't necessarilary true for any given product that uses the SOC.

For example, if talking about the N8x0 platforms, they have
limitations to their implementation (both hw and sw) that in short
mean:

 - As mentioned, the external LCD controller has lousy memory
bandwidth which limits the framerate even if you could decode faster
 - If a "dynamic" DSP task is loaded, the ARM core is locked to 333
MHz (max is 400MHz)

This in effect means that the usage of DSP for MP3 decoding "wastes"
77MHz of the ARM core which, according to a mplayer dev, is more than
enough to decode MP3 audio. Thus, you can achieve better video
performance by using a software codec to decode both the video and the
audio than what you get by using the DSP codec for the MP3 decoding.
Search the maemo-developers mailing list archives for more details.

I don't know if a DSP video codec would change that scenario though,
but for N8x0 that point is pretty moot since the actual bottleneck is
the slow memory bus to the LCD controller, not decoding power (as far
as I know at least).

Oh, and for the record, despite all the shortcomings I'm happy enough
watching 640x240 at 23.976fps videos (scaled to whatever the aspect ratio
really is) on my N800 :)

-- 
Kalle Vahlman, zuh at iki.fi
Powered by http://movial.fi
Interesting stuff at http://syslog.movial.fi