You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I think there are plenty of reasons to support f16 as a native arithmetic type in Rust, but my primary use-case is ML inference for hardware that supports fp16 arithmetic (e.g. the Cortex-a55).
I've resorted to writing simple functions (multiply - add, dot products, etc...) to operate on _Float16 values in C and calling them because the half crate's conversion cost is really painful for anything low-latency/high frequency (audio processing). It is... far from efficient.
The text was updated successfully, but these errors were encountered:
VariantXYZ
changed the title
f16 and bf16 native support in Rust
f16 native support in Rust
Dec 24, 2022
Sorry in advance if I missed a tracking issue somewhere, the ones I've found seemed to briefly mention but not propose/drive native f16 specifically.
Specifically, I've seen:
#2629
#521
rust-lang/stdarch#344
I think there are plenty of reasons to support f16 as a native arithmetic type in Rust, but my primary use-case is ML inference for hardware that supports fp16 arithmetic (e.g. the Cortex-a55).
I've resorted to writing simple functions (multiply - add, dot products, etc...) to operate on _Float16 values in C and calling them because the half crate's conversion cost is really painful for anything low-latency/high frequency (audio processing). It is... far from efficient.
The text was updated successfully, but these errors were encountered: