Re: [FFmpeg-devel] SIMD Motion Compensation Optimizations
On Monday, November 11th, 2024 at 3:12 AM, Rémi Denis-Courmont wrote: > > I recently read a news article about impressive AVX-512 optimizations > > of Motion Compensation functions and I'd like to learn more. > > Unfortunately, the article merely linked the post on X, which > > included a screenshot and little else. > > > > https://x.com/FFmpeg/status/1852542388851601913 > > That's a screenshot of dav1d's checkasm, so you'll find the code in dav1d, > not FFmpeg. Thanks for the pointer. I'd like to alert folks that the only way I could reproduce performance numbers similar to theirs is when I built it with '-Dbuildtype=debug'. Otherwise, the SIMD versions are all within the typical order of magnitude vs. the generic C implementations you'd normally expect. Again, I'm lacking any context concerning the original screen grab. However, the actual text of the tweet appears to be highly misleading. "A 94x speed improvement demonstrated using handwritten assembly" ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-devel] SIMD Motion Compensation Optimizations
Dear ffmpeg developers, I recently read a news article about impressive AVX-512 optimizations of Motion Compensation functions and I'd like to learn more. Unfortunately, the article merely linked the post on X, which included a screenshot and little else. https://x.com/FFmpeg/status/1852542388851601913 Unfortunately, I do not have an account on X/twitter, nor am I interested in supporting surveillance capitalism by creating one. I think it's inconsistent with open source values to use such a site, when free alternatives exist that do not require an account to browse. Anyway, please advise where I can learn more about the referenced AVX-512 optimizations. I tried searching in the git repo, but was unable to find them. Thanks, in advance. ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] SIMD Motion Compensation Optimizations
On Monday, November 11th, 2024 at 12:09 PM, compn wrote: > are you trying to say that ffmpeg should take down the tweet? I only provided my observations (with the caveat that I'm still potentially lacking context). What anyone might want to do about it isn't for me to say. If you'd like to check for yourself, try running meson with/without the '-Dbuildtype=debug' option. Note that you'll also need to pass meson the '-Dtrim_dsp=false' option, in order to be able to run the test. Once you've built the tree, which only takes about 30 seconds, you can run the exact commandline shown in the image and see how the relative performance ratios look on your machine. If your CPU lacks AVX-512, you can just compare the performance of C vs SSSE3 and AVX2, which should be more than convincing enough. > well you can come on #ffmpeg-devel irc too, if you dont like twitter. Thank you. I will probably do that, at some point. ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
