[FFmpeg-devel] Re: [FFFjo] [FFmpeg/FFmpeg] aarch64/vvc: Optimisations of put_luma_v() functions for 10/12-bit (PR #21148)

[Michael Niedermayer via ffmpeg-devel <ffmpeg-devel@ffmpeg.org>]

[-- Attachment #1.1: Type: text/plain, Size: 3023 bytes --]

Hi Zhao Zhili

On Wed, Jan 07, 2026 at 11:34:35PM +0800, Zhao Zhili via ffmpeg-devel wrote:
> 
> 
> > On Jan 7, 2026, at 23:10, Martin Storsjö <code@ffmpeg.org> wrote:
> > 
> > 
> > I don't see the update here in this branch (nor in https://code.ffmpeg.org/george.zaguri/FFmpeg/src/branch/test/put_luma_v) for changing the interleaving for in-order cores?
> > 
> > For what it's worth; at https://code.ffmpeg.org/mstorsjo/FFmpeg/commits/branch/vvc-put-luma-v I pushed the WIP commit I made for testing this, which gives numbers on A53/A55 as shown above. I don't have newer in-order cores than that to test on. (On macOS, it's not possible to pin threads/processes to a specific core, so I'm not aware of a good way of benchmarking on the Apple Silicon efficiency cores.)
> > 
> > @sean_mcg does have a machine with A520 and A720 cores though - I'd appreciate if he could benchmark this and give corresponding numbers.
> > 
> > @sean_mcg : If you have time to help out with this; check out https://code.ffmpeg.org/mstorsjo/FFmpeg/commits/branch/vvc-put-luma-v and the preceding commit, build checkasm, and run it (tests/checkasm/checkasm --bench --test=vvc_mc --bench="put_luma_v_1*"). If you don't have a working setup for allowing user mode access to the cycle counters (i.e. can build and load the kernel module from https://code.videolan.org/janne/arm64-cycle-cnt), you might be able to edit libavutil/aarch64/timer.h and change #if defined(__ANDROID__) || defined(__APPLE__) into #if 1 to choose the slightly worse but more accessible timer register.)
> > 
> > With that setup in place, we'd want numbers for A520 (and ideally A720 too). You can restrict it to run on a specific core by running taskset -c 4 ./tests/checkasm/checkasm ...where 4 is the number of the core you want to run it on. It might require some testing to figure out which cores are the big ones (A720) and which ones are the small ones (A520). The numbers on A520 should be bigger, as it's a slower core. We're mostly interested in the before/after numbers on A520, but if you can time on A720 as well, that'd be appreciated!
> > 
> 
> Can we purchase a ARM V9 board for testing, e.g., radxa o6, orange pi 6 ?

It sounds like we need or would benefit from such a board, thus
Iam in favor of that, please avoid any board that has issues though (i dont know if any of them has)

thx

>
> [1] https://radxa.com/products/orion/o6/
> [2] http://www.orangepi.org/html/hardWare/computerAndMicrocontrollers/details/Orange-Pi-6-Plus.html
> _______________________________________________
> ffmpeg-devel mailing list -- ffmpeg-devel@ffmpeg.org
> To unsubscribe send an email to ffmpeg-devel-leave@ffmpeg.org

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Frequently ignored answer#1 FFmpeg bugs should be sent to our bugtracker. User
questions about the command line tools should be sent to the ffmpeg-user ML.
And questions about how to use libav* should be sent to the libav-user ML.

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 195 bytes --]

[-- Attachment #2: Type: text/plain, Size: 163 bytes --]

_______________________________________________
ffmpeg-devel mailing list -- ffmpeg-devel@ffmpeg.org
To unsubscribe send an email to ffmpeg-devel-leave@ffmpeg.org