[Libva] How to detect the type of memory returned...

[Gwenole Beauchesne]

Hi,

2014-06-17 10:53 GMT+02:00 Xiang, Haihao <haihao.xiang at intel.com>:
> On Tue, 2014-06-17 at 17:54 +1000, Jean-Yves Avenard wrote:
>> On 17 June 2014 17:27, Peter Frühberger <peter.fruehberger at gmail.com> wrote:
>> > Hi,
>> >
>> > we implemented the SSE4 copy code arround that was referenced here
>> > into xbmc arround christmas, cause we wanted to use state of the art
>> > deinterlacing with ffmpeg's yadif.
>> > Here are some benchmarks with 8K vs 4K buffer and 1080p vs 4K surface sizes:
>>
>> Except that getting the picture via vaDeriveImage returns a NV12 image
>> (at least with Intel, AMD VAAPI doesn't implement vaDeriveImage), in
>> which case that routine serves no purpose until of course you can
>> natively display NV12 image.
>>
>> And vaGetImage returns a YV12 image, which isn't a USMC map memory,
>> and the SSE4 routine is in this case three times as slow as the plain
>> C version using memcpy.
>
> Hi,
>
> Can you give a try with the attached patch ? It uses
> drm_intel_gem_bo_map_gtt() to map a buffer through the aperture rather
> than directly to its CPU cacheable memory

That's what I had in mind except that the plan was also to allocate
tiled surfaces for VA images. That should also benefit to the CSC
process too. Can you implement that?

Thanks,
Gwenole.

>> We always get back to the problem I mentioned in my first email.
>> Unfortunately, there's not a generic solution that can be adapted.
>> If memory used is USWC, you must use SSE4, if not, you certainly don't
>> want to use SSE4 and a buffer
>>
>> > Fun fact:
>> > Slow Baytrail Chips are faster and more stable doing VAAPI Decoder +
>> > sse4 transfert + software deinterlacing than using VPP with Bobbing.
>>
>> You don't have to use VPP for Bob.
>> But yeah, I see your point for other type of VPP's deinterlacers
>> _______________________________________________
>> Libva mailing list
>> Libva at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/libva
>
>
> _______________________________________________
> Libva mailing list
> Libva at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/libva
>



-- 
Gwenole Beauchesne
Intel Corporation SAS / 2 rue de Paris, 92196 Meudon Cedex, France
Registration Number (RCS): Nanterre B 302 456 199