[FFmpeg-devel] [PATCH 2/2] swscale/x86/input: add AVX2 optimized uyvytoyuv422

[Rémi Denis-Courmont]

Le 6 juin 2024 10:01:24 GMT+03:00, Christophe Gisquet <christophe.gisquet at gmail.com> a écrit :
>Le jeu. 6 juin 2024 à 08:11, Rémi Denis-Courmont <remi at remlab.net> a écrit :
>> >James Almer:
>> >> uyvytoyuv422_c: 23991.8
>> >> uyvytoyuv422_sse2: 2817.8
>> >> uyvytoyuv422_avx: 2819.3
>> >
>> >Why don't you nuke the avx version in a follow-up patch?
>>
>> Same problem with the RGBA stuff as well. Are the AVX functions expected to be faster than SSE2 on processors *without* AVX2?
>
>Something frequent in this type of questions is that people are using
>numbers from a CPU that has had 10 years of arch improvements (and
>probably a doubling in throughput for any instruction set) over one
>that supported at most AVX. The presence of an AVX function (whose
>benefit is only 3-operand instructions, so admittedly small) would
>ideally only be benchmarked on that kind of CPUs.

It feels a bit dense for someone not intimate with x86 innards such as I. Intuitively, 3 operands instructions are certainly helpful in avoiding vector copies around destructive operations. But it should be clear if, for any given function, this does or does not help.

That being said, I have no objections as such to run-time optimisations for middle-aged processors.

If anything, I think we should auto-trim the useless C code at least via DCE if we want to save space. We can assume at least SSE2 on x86-64, no? Of course this will break checkasm and some command line flags.