Hey Opening an issue instead of a PR for this one because it's super dirty work atm:
Basically on neon aarch64 (M1 Mac) we can add pure f16 intrinsics and get pretty sizeable speedup:
Something like ~2x on most matmuls
LaurentMazare#4
However this requires hacking new intrinsics using arm! macro which seem to confuse the compiler (most likely because I didn't write them properly)
Hey Opening an issue instead of a PR for this one because it's super dirty work atm:
Basically on neon aarch64 (M1 Mac) we can add pure f16 intrinsics and get pretty sizeable speedup:
Something like ~2x on most matmuls
LaurentMazare#4
However this requires hacking new intrinsics using
arm!macro which seem to confuse the compiler (most likely because I didn't write them properly)