4 ms·
The AVX2 version was still 64x faster than the C one, so AVX-512 is just 50% improvement over that. Hand vectorized assembly is very much the key to the gains.
by saati 1y ago
The AVX2 version was still 64x faster than the C one, so AVX-512 is just 50% improvement over that. Hand vectorized assembly is very much the key to the gains.
- brigade 1y agoThe only material difference AVX2 makes is that it can't saturate L1 bandwidth. Which would imply that 100x for AVX-512 is only for frames that fit within L1. And... yep, the benchmark on 256x16 frames. [1] [1] https://ffmpeg.org/pipermail/ffmpeg-devel/2025-July/346729.html https://ffmpeg.org/pipermail/ffmpeg-devel/2025-July/346729.h...