In https://github.com/numpy/numpy/pull/25934, we see up to 3x slow down in gathering loads on AArch64 when using Highway. I believe this is due to https://github.com/google/highway/pull/2116, which seems to have subtly changed the code path.
In numpy/numpy#25934, we see up to 3x slow down in gathering loads on AArch64 when using Highway.
I believe this is due to #2116, which seems to have subtly changed the code path.