Floating point conversion functions.
Software and hardware implementations
The [native] module provides only those conversions which are natively
available on the target hardware.
The [soft] module provides a software implementation of all
conversion functions, for targets which do not provide them natively.
The root of the crate exposes the best version of all available functions.
Rounding mode
- Functions named
_roundround the integer to the closest possible floating point number, and breaks ties to even. - Functions named
_truncatetruncate the result, which means they round towards zero. - Functions named
_anyare an alias of whichever variant is fastest on the target. - Functions without a rounding mode in their name do not round. These conversions are always lossless.
Speed
For conversions that aren't available natively, the software
implementations in this crate seem to be both faster and smaller in almost
all cases compared to the ones currently used by x as f64 (from the
compiler builtins runtime support library).
Work in progress
This crate is usable, but still incomplete:
- There's no support for
f32yet. - There's no support for converting to integers yet.
- Native conversions are only available on ARM (32- and 64-bit) and x86 (32- and 64-bit).
- The software implementations can probably be optimized further.
- More benchmarking still needs to happen.