tree 8d9f1f9cc781a548e0c771be4234f29ce4f6feae
parent c7d034617588bd2e72e1a4b1cebf881372e58e15
author Lei Zhang <antiagainst@google.com> 1676929092 -0800
committer Lei Zhang <antiagainst@google.com> 1686716251 -0700

[metal] Add builtin executables for polyfilling buffer fills

On macOS, the Metal API does not support buffer fills with
non-4-multiple buffer offsets and lengths for fills. It also
does not support 2/4-byte fill patterns. For such cases, we
need to use a kernel dispatch to polyfill it.

Three kernels are provided for different cases, where we can
optimize for 16-byte or 4-byte aligned cases a bit better.