Closed Nielsbishere closed 1 year ago
Halfs are really nice on the GPU. We might want to have a type that:
F16 can be supported on CPU too; but needs manual implementation which is SLOW.
https://en.wikipedia.org/wiki/F16C
Implemented. F16C is apparently slow. So defaulting to software fallback.
Halfs are really nice on the GPU. We might want to have a type that:
F16 can be supported on CPU too; but needs manual implementation which is SLOW.