With some effort, we can make the current implementation more generic and avoid code duplication by sharing it for all feasible architectures. Maybe even for HalfPackets and Packets (relevant especially for AVX).
I guess overall this is easy enough to be implemented for 3.3
Created attachment 501 [details]
AVX version of the pexp/plog/psqrt math functions
Thank you Benoit for sharing them.
Created attachment 542 [details]
AVX version of the sine function
Benoit, feel free to push them upstream.
Done in commit af4780a, closing this bug.