Hi there.

Let me add some background here.

The assembly optimization of inv is too complicated to be a GSOC task.
You can refer to the code of dav1d to see how it can be. 
https://code.videolan.org/videolan/dav1d/-/blob/master/src/x86/itx_avx2.asm?ref_type=heads.
We do have an internal version of it. That's why we haven't posted and merged 
it yet.

About the intra prediction, some function of the current C code with the
compilation optimization enabled will run fast enough, so the handwritten
assembly will not definitely be faster than the C code. You may need to run
a performance test to evaluate it.

Thanks,
Jianhua

________________________________
发件人: Frank Plowman via ffmpeg-devel <[email protected]>
发送时间: 2026年3月29日 3:20
收件人: [email protected] <[email protected]>
抄送: Frank Plowman <[email protected]>
主题: [FFmpeg-devel] Re: [GSoC 2026] Proposal/Mentorship Request: VVC x86 SIMD 
Optimization

On 28/03/2026 16:01, Hank Li via ffmpeg-devel wrote:
> Hello,
>
> I'm a bit late to the party, but I'm interested in proposing a GSoC
> project on VVC x86 SIMD optimization. This was on the 2025 ideas list
> but there are currently no SIMD overrides for itx or intra_pred in
> avcodec/x86/vvc/dsp_init.c, only ALF and inter prediction have x86
> assembly.
>
> I would like to propose implementing:
> -AVX2 inverse transforms for DCT-II, DST-VII, and DCT-VIII
> -Intra prediction for DC, planar, and angular modes
> -Corresponding checkasm tests and benchmarks
>
> For qualification, I'm currently working on a checkasm test for VVC
> inverse transforms (tests/checkasm/vvc_itx.c).
>
> I've emailed Nuo Mi about mentoring since he mentored the VVC SIMD
> projects in 2023 and 2024. If anyone else is interested in mentoring
> or has thoughts, I'd appreciate it.
>
> Best,
> Hankang Li
> _______________________________________________
> ffmpeg-devel mailing list -- [email protected]
> To unsubscribe send an email to [email protected]

Hi,

Thanks for your interest in the VVC decoder.

Writing assembly optimizations for the inverse transforms will be
significantly more difficult than for the intra prediction I believe.
There isn't much precedent for rectangular transforms available, whereas
the intra prediction functions are largely similar to HEVC's only with
an increased number of taps.  Perhaps a more suitable plan would be to
begin with the intra prediction functions and then to have the inverse
transform functions as a "stretch goal".  I don't know how familiar you
are with AVX2 and video coding, but that would represent a more natural
introduction and progression of difficulty.

--
Frank
_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to