Re: [FFmpeg-devel] [PATCH v2 1/2] avfilter/vf_blackdetect: add AVX2 SIMD version

[Zhao Zhili <quinkblack@foxmail.com>]

> On Jul 18, 2025, at 19:36, Kacper Michajlow <kasper93@gmail.com> wrote:
> 
> On Thu, 17 Jul 2025 at 12:45, Niklas Haas <ffmpeg@haasn.xyz> wrote:
>> 
>> From: Niklas Haas <git@haasn.dev>
>> 
>> Requested by a user. Even with autovectorization enabled, the compiler
>> performs a quite poor job of optimizing this function, due to not being
>> able to take advantage of the pmaxub + pcmpeqb trick for counting the number
>> of pixels less than or equal-to a threshold.
>> 
>> blackdetect8_c:                                       4625.0 ( 1.00x)
>> blackdetect8_avx2:                                     155.1 (29.83x)
>> blackdetect16_c:                                      2529.4 ( 1.00x)
>> blackdetect16_avx2:                                    163.6 (15.46x)
> 
> I think we should try to have better standards for reporting
> performance metrics. Those numbers without context mean not so much.
> What compiler, flags, cpu were used? Sure, we can omit some
> information if we want to show only the scaling, but if it highly
> depends on those things, then we should at least try to be more
> specific.
> 
> Sorry for being pedantic about those things, but I think it's
> important, especially if we put those values in a commit message which
> will live forever in the repository as a vague reference.

The basic benchmark context can be specified in doc and leave out in commits.
And it’s easy to get the implicit conditions in the mailing list. I’m more worried about
only give the data on social media platforms, e.g., X.

The checkasm benchmark serve our purpose very well, but it’s not a fair benchmark to
compare hand written assembly to compiler optimizations. Sure we can do better than
compiler, but compiler isn’t that sucks. More importantly, compiler isn’t our enemy,
no point in embarrassing the compiler or compiler developer.

"If I have seen further it is by standing on the shoulders of Giants.”
 — By Isaac Newton.

Runtime cpu detection is a thing. There are methods to use compiler auto-vectorize and
runtime cpu detection at the same time. Our approach isn’t the only working approach.

> 
>> Even with autovectorization enabled
> 
> You mention the auto vectorization enabled, yet the reported numbers
> are without it. In my mind this description implies that shown
> performance comparison is with auto vectorization enabled.
> 
> When we compare apples to apples, with avx2 we get a more expectable
> 3.74x (gcc) / 2.38x (clang) depending on the compiler. It's still a
> good improvement, no reason to oversell it.
> 
> For reference some metrics on me end:
> 
> clang 20.1.7
> 
> march=generic (default config):
> 
> blackdetect8_c:                                       1591.1 ( 1.00x)
> blackdetect8_avx2:                                     225.1 ( 7.07x)
> blackdetect16_c:                                       643.5 ( 1.00x)
> blackdetect16_avx2:                                    220.6 ( 2.92x)
> 
> march=core-avx2:
> 
> blackdetect8_c:                                        526.0 ( 1.00x)
> blackdetect8_avx2:                                     220.9 ( 2.38x)
> blackdetect16_c:                                       318.8 ( 1.00x)
> blackdetect16_avx2:                                    225.9 ( 1.41x)
> 
> gcc 14.2.0
> 
> -fno-tree-vectorize (default config):
> 
> blackdetect8_c:                                       5126.6 ( 1.00x)
> blackdetect8_avx2:                                     198.0 (25.89x)
> blackdetect16_c:                                      2151.9 ( 1.00x)
> blackdetect16_avx2:                                    196.8 (10.93x)
> 
> march=generic -ftree-vectorize:
> 
> blackdetect8_c:                                       1354.4 ( 1.00x)
> blackdetect8_avx2:                                     196.9 ( 6.88x)
> blackdetect16_c:                                       644.2 ( 1.00x)
> blackdetect16_avx2:                                    249.8 ( 2.58x)
> 
> march=core-avx2 -ftree-vectorize:
> 
> blackdetect8_c:                                        820.8 ( 1.00x)
> blackdetect8_avx2:                                     219.2 ( 3.74x)
> blackdetect16_c:                                       372.8 ( 1.00x)
> blackdetect16_avx2:                                    201.4 ( 1.85x)
> 
> Again, sorry for being pedantic here, but it gives the wrong
> impression especially if you look at this from outside.
> 
> - Kacper
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".


_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".