9 releases

0.2.8 Aug 12, 2020
0.2.7 Aug 12, 2020
0.1.0 Aug 7, 2020

#14 in #floating

27 downloads per month

BSD-2-Clause

39KB
914 lines

Floating point conversion functions.

Software and hardware implementations

The soft module provides a software implementation of all conversion functions, for targets which do not provide them natively. These are implemented without any floating point operations, so are also useful for software that needs to avoid using floating point hardware.

The fast module provides a fast implementation of all conversion functions by making use of native floating point instructions where possible.

Conversion of integers to floating point values

  • Functions named _round round the integer to the closest possible floating point number, and breaks ties to even.
  • Functions named _truncate truncate the result, which means they round towards zero.
  • Functions without a rounding mode in their name do not round. These conversions are always lossless.
  • The only conversion that can overflow is u128_to_f32_round, in which case it returns f32::INFINITY.

Conversion of floating point values to integers

  • These conversions truncate, which means they round towards zero.
  • Values higher than what the integer can represent (including +∞) result in the maximum integer value.
  • Values lower than what the integer can represent (including −∞) result in the minimum integer value.
  • NaN is converted to zero.

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 <type> (from the compiler builtins runtime support library).

Work in progress

This crate is usable, but still incomplete:

  • Native conversions are only available on ARM (32- and 64-bit) and x86 (32- and 64-bit).
  • The truncating functions do not (yet) use any native floating point instructions.

No runtime deps