VP8 performance on ARM Cortex A9 - Neon

[rupesh_kumbhare]

Dear VP8-dev-team,

I am looking into the feasibility of getting 1080p/1080i VP8 decoding
on ARM Cortex A9 (Neon). Can you give the current performance in terms
of MHz/MCPS and memory consumption on Cortex A9?

Any performance comparison with H.264 codec will also help. Appreciate
your help!!!

Thanks,
Rupesh

[Johann Koenig]

On Thu, Aug 26, 2010 at 07:36, rupesh_kumbhare
<rupesh....@gmail.com> wrote:
> I am looking into the feasibility of getting 1080p/1080i VP8 decoding
> on ARM Cortex A9 (Neon). Can you give the current performance in terms
> of MHz/MCPS and memory consumption on Cortex A9?

My dev board is a Beagle Board with an A8 at the moment, so I can't
say how it would run on the A9. Best bet would be to grab the library
and run a test on sample hardware.

> Any performance comparison with H.264 codec will also help.

Sorry, I don't have any.
--
- johann koenig
google

[Vladimir Pantelic]

rupesh_kumbhare wrote:
> Dear VP8-dev-team,
>
> I am looking into the feasibility of getting 1080p/1080i VP8 decoding
> on ARM Cortex A9 (Neon). Can you give the current performance in terms
> of MHz/MCPS and memory consumption on Cortex A9?

http://bloggingthemonkey.blogspot.com/2010/08/ffpv8-neon-720p24.html

720p on an A9 so far using ffvp8 and neon. libvpx on an A8 is far from
reaching 720p.

> Any performance comparison with H.264 codec will also help. Appreciate
> your help!!!

Since h264 and vp8 are somewhat *cough* similar, I'd expect similar
performance running on an ARM CPU. For SOCs with dedicated HW units
(e.g. OMAP3/OMAP4) the differences between VP8 and H264 might make not
all HW units be usable and therefore H264 would perform better than VP8.

[Rob Clark]

On Tue, Aug 31, 2010 at 2:02 AM, Vladimir Pantelic <vlad...@gmail.com> wrote:
> rupesh_kumbhare wrote:
>>
>> Dear VP8-dev-team,
>>
>> I am looking into the feasibility of getting 1080p/1080i VP8 decoding
>> on ARM Cortex A9 (Neon). Can you give the current performance in terms
>> of MHz/MCPS and memory consumption on Cortex A9?
>
> http://bloggingthemonkey.blogspot.com/2010/08/ffpv8-neon-720p24.html
>
> 720p on an A9 so far using ffvp8 and neon. libvpx on an A8 is far from
> reaching 720p.

Hi! libvpx suffers a bit, esp on ARM SoC sort of devices, because of
some extra memcpy's (whenever one of the reference frames is updated),
and more passes over the data, so a bit less efficiency when it comes
to cache utilization. At this point, there are a couple more
functions that can be converted to neon for ffvp8. (And lack of
multi-thread support in ffvp8.) And the neon code could be tuned to
save a few cycles here and there. And a few other misc things. So
still some room for improvement. But as it is it is just a bit over
720p24 avg at the display and dips below that on complex scenes at
1ghz on an A9. But 1080p is a whole lot more than 720p!! I think for
1080p you need to hw accelerate it or have a lot of ghz and memory
bandwidth to keep up ;-)

Really, from a power and performance standpoint, you can't beat asic!

BR,
-R

>
>> Any performance comparison with H.264 codec will also help. Appreciate
>> your help!!!
>
> Since h264 and vp8 are somewhat *cough* similar, I'd expect similar
> performance running on an ARM CPU. For SOCs with dedicated HW units
> (e.g. OMAP3/OMAP4) the differences between VP8 and H264 might make not
> all HW units be usable and therefore H264 would perform better than VP8.
>

> --
> You received this message because you are subscribed to the Google Groups
> "Codec Developers" group.
> To post to this group, send email to codec...@webmproject.org.
> To unsubscribe from this group, send email to
> codec-devel...@webmproject.org.
> For more options, visit this group at
> http://groups.google.com/a/webmproject.org/group/codec-devel/?hl=en.
>
>

[rupesh_kumbhare]

Hi Rob,

Thank you for responding to my query! Please see my inline comments.

Regards,
Rupesh

On Aug 31, 12:02 pm, Vladimir Pantelic <vlado...@gmail.com> wrote:
> rupesh_kumbhare wrote:
> > Dear VP8-dev-team,
>
> > I am looking into the feasibility of getting 1080p/1080i VP8 decoding
> > on ARM Cortex A9 (Neon). Can you give the current performance in terms
> > of MHz/MCPS and memory consumption on Cortex A9?
>
> http://bloggingthemonkey.blogspot.com/2010/08/ffpv8-neon-720p24.html

[Rupesh] Thanks for the link. I too got it after the post, had gone
through it.

>
> 720p on an A9 so far using ffvp8 and neon. libvpx on an A8 is far from
> reaching 720p.
>
> > Any performance comparison with H.264 codec will also help. Appreciate
> > your help!!!
>
> Since h264 and vp8 are somewhat *cough* similar, I'd expect similar
> performance running on an ARM CPU. For SOCs with dedicated HW units
> (e.g. OMAP3/OMAP4) the differences between VP8 and H264 might make not
> all HW units be usable and therefore H264 would perform better than VP8.

[Rupesh] So far everyone has achieved any HD decode using HW
accelerators, and I believe thats the way you can conserve power on
the portable devices. No point in going for SW solution when the
standard is almost carved in stones.

[Vladimir Pantelic]

rupesh_kumbhare wrote:

>> > Any performance comparison with H.264 codec will also help. Appreciate
>> > your help!!!
>>
>> Since h264 and vp8 are somewhat *cough* similar, I'd expect similar
>> performance running on an ARM CPU. For SOCs with dedicated HW units
>> (e.g. OMAP3/OMAP4) the differences between VP8 and H264 might make not
>> all HW units be usable and therefore H264 would perform better than VP8.
>
> [Rupesh] So far everyone has achieved any HD decode using HW
> accelerators, and I believe thats the way you can conserve power on
> the portable devices. No point in going for SW solution when the
> standard is almost carved in stones.

"almost" is not good enough. HW accelerator design in e.g. omap3 and omap4
predates the public release of VP8 by years...

Only chips that are in the design stage currently will have (full) VP8
hw acceleration...