Prominent FFmpeg developer Niklas Haas has extended the FFmpeg libswscale to make use of AVX-512's Vector Permute Byte "VPERMB" instructions to help with more efficient video scaling and pixel format conversion. With VPERMB the code now benefits from full cross-lane byte permutation to provide better performance for non-lane-aligned scenarios.
The exciting takeaway for end-users of FFmpeg is that Haas found RGBA24 1920x1080 to RGBA 1920x1080 conversion was 1.372x faster with this reworked code on AVX-512 processors. This is great news for AMD Zen 4 and newer CPUs or Intel Xeon and other select Intel CPUs supporting AVX-512.
More details for those interested in this commit .