It is easy, but I'm not sure the coarse-grained levels accurately capture the craziness of which AVX512 feature bits are supported on which models once you get beyond amd64-v3. Maybe it doesn't matter.
If you want to branch around having or not having something like AVX-VNNI you still need to use IFUNC or just patch your own function pointers (this is the better choice, since IFUNC is both slow and a security risk).
This looks really nice. No pain aside from 3x the build time.
Does anyone know if Windows or macOS provide equivalent dynamic dispatch functionality that's resolved automatically? (Less of a big deal with macOS, but if you wanted to use SVE or something.)
It is easy, but I'm not sure the coarse-grained levels accurately capture the craziness of which AVX512 feature bits are supported on which models once you get beyond amd64-v3. Maybe it doesn't matter.
https://en.wikipedia.org/wiki/X86-64#Microarchitecture_level...
If you want to branch around having or not having something like AVX-VNNI you still need to use IFUNC or just patch your own function pointers (this is the better choice, since IFUNC is both slow and a security risk).
This looks really nice. No pain aside from 3x the build time.
Does anyone know if Windows or macOS provide equivalent dynamic dispatch functionality that's resolved automatically? (Less of a big deal with macOS, but if you wanted to use SVE or something.)