9 unstable releases (3 breaking)
0.4.1 | Sep 7, 2023 |
---|---|
0.4.0 | Sep 1, 2023 |
0.3.0 | Aug 30, 2023 |
0.2.2 | Aug 8, 2023 |
0.1.1 | Jan 17, 2023 |
#248 in Math
65KB
1K
SLoC
Fixed Math
This library implements analytic/trigonometric functions for fixed point numbers.
Implemented functions:
sqrt
from traitSqrt
Sqrt
is not implemented for fixed numbers with less than 2 integer bits, but there are functions that work on those types:sqrt_i1
, ...
sin_cos
,sin
,cos
,tan
from traitSinCos
SinCos
is not implemented for fixed numbers with less than 8 (TODO 7) integer bits because the there are table values that would overflow on those types.- All calculations are made in degrees
- except that there is a
sin_cos_rad
function which is very imprecise; check source code for why, feel free to fix it (its not a priority for me)
- except that there is a
Optional Features
- std : std feature of fixed, there is no proper no-std support yet...
Examples
There are traits and standalone functions, see examples on how to use them.
Errors
Check the examples to see about how much error this implementation produces.
Usually sqrt
has an error of around 1-2 Delta.
(Delta = the distance to the next representable number)
sin_cos
may produce bigger errors, around 1-2 decimal places.
Benchmarks
You can check or run the benchmarks in benches
.
Here are some conclusions I've got to:
SinCos
2022-09-28
Calculation time for sin_cos varies with the fixed number's byte size.
- I10F6: ~ 8ns
- I16F16: ~ 9ns
- I32F32: ~ 18ns
- I32F96: ~ 210ns
Notes:
- there are many different int bit / frac bit combinations; I did not test them
(int bits must be >= 10 (but maybe I can do something to relax that further)) - these are all calculations in degrees
- code was compiled with native cpu features
- go for FixedI32 instead of FixedI16 unless you are limited by memory much
- I did a benchmark in the same style on
cordic
'ssin_cos
on FixedI64- keep in mind that
cordic
works with radians, I used the same angle values - so they can take
sin_cos
of a lot bigger angle on the same number representation size - this crate was about 1.5-2 times faster on same angle sizes
- keep in mind that
2023-07-30
System info:
-`
.o+` --------
`ooo/ OS: Arch Linux x86_64
`+oooo: Host: X570 AORUS ELITE -CF
`+oooooo: Kernel: 6.4.7-arch1-1
-+oooooo+:
`/:-:++oooo+:
`/++++/+++++++: Shell: fish 3.6.1
`/++++++++++++++: Resolution: 3840x2160
`/+++ooooooooooooo/` DE: Hyprland
./ooosssso++osssssso+`
.oossssso-````/ossssss+`
-osssssso. :ssssssso. Terminal: WezTerm
:osssssss/ osssso+++. CPU: AMD Ryzen 7 5800X (16) @ 3.800GHz
/ossssssss/ +ssssooo/- GPU: AMD ATI Radeon RX 7900 XT/7900 XTX
`/ossssso+/:- -:/+osssso+- Memory: 32014MiB
`+sso+:-` `.-/+oso:
`++:. `-/+/
.` `/
Performance is different now and without native cpu features:
- I10F6: ~ 430ps
- I16F16: ~ 15ns
- I32F32: ~ 25ns
- I32F96: ~ 255ns
With native cpu features : RUSTFLAGS="-C target-cpu=native" cargo bench
:
- I10F6: ~ 420ps
- I16F16: ~ 8ns
- I32F32: ~ 22ns
- I32F96: ~ 222ns
Notes:
- performance regressed at I32F32 and I32F96
- performance improved a lot at I10F6, so we might consider using 16bit fixed point values where we do not need much precision
- cordic benchmark not checked...
License
All code in this repository is dual-licensed under either:
- MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)
- Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
at your option.
Contributing
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.
Attribution
We use code modified from cordic, licensed as BSD-3-Clause:
Dependencies
~2.5MB
~45K SLoC