bsteinb / accurate Goto Github PK

I searched for Kahan, floating point precision, sum, ... and it did not pop up.

It might be worth it to expend some time doing some crates.io SEO, e.g., expanding the readme with more information, maybe mentioning which algorithms are implemented in this crate, etc.

Some unsafe code relies on RawExponent::raw_exponent() < FloatFormat::base_pow_exponent_digits() to be sound. Currently a user could write a custom float type in safe rust and trigger UB by messing up the trait implementation.

Example:
https://github.com/bsteinb/accurate/blob/master/src/sum/onlineexactsum.rs#L63

Hi @bsteinb

thank you for this nice repository!

Could you maybe point me towards a resource on which algorithm to choose? I read the original papers, but of course every paper tries to sell their contribution and thus it is hard to gauge which algorithm is actually a good choice. If you would have any thoughts on this, or resources to point to, I would very much appreciate it.

Thank you
Lukas

Consider exactly representable f32 numbers 4194304.0 and 4194304.5. Their sum is 8388608.5, which is not exactly representable as an f32, so it needs to be rounded to the nearest even mantissa, which is 8388608. However accurate doesn't do this correctly:

let a: f32 = 4194304.0;
let b: f32 = 4194304.5;
println!("a = {:.1}", a);
println!("b = {:.1}", b);
println!("a + b = {:.1}", a + b);

use accurate::traits::SumAccumulator;
let s1 = accurate::sum::OnlineExactSum::zero().absorb([a, b]).sum();
let s2 = accurate::sum::i_fast_sum_in_place(&mut [a, b]);
println!("OnlineExactSum: {:.1}", s1);
println!("i_fast_sum: {:.1}", s2);

The above prints:

a = 4194304.0
b = 4194304.5
a + b = 8388608.0
OnlineExactSum: 8388609.0
i_fast_sum: 8388609.0

error[E0034]: multiple applicable items in scope
--> /home/parrot/.cargo/registry/src/github.com-1ecc6299db9ec823/accurate-0.3.0/src/sum/ifastsum.rs:78:32
|
78 | sm = sm.max(st.abs());
| ^^^ multiple abs found
|
note: candidate #1 is defined in the trait ieee754::Ieee754
--> /home/parrot/.cargo/registry/src/github.com-1ecc6299db9ec823/ieee754-0.2.6/src/traits.rs:472:5
|
472 | fn abs(self) -> Self;
| ^^^^^^^^^^^^^^^^^^^^^
note: candidate #2 is defined in the trait num::Float
--> /home/parrot/.cargo/registry/src/github.com-1ecc6299db9ec823/num-traits-0.2.12/src/float.rs:1191:5
|
1191 | fn abs(self) -> Self;
| ^^^^^^^^^^^^^^^^^^^^^
help: disambiguate the associated function for candidate #1
|
78 | sm = sm.max(ieee754::Ieee754::abs(st));
| ^^^^^^^^^^^^^^^^^^^^^^^^^
help: disambiguate the associated function for candidate #2
|
78 | sm = sm.max(num::Float::abs(st));
| ^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error

For more information about this error, try rustc --explain E0034.
error: could not compile accurate.

To learn more, run the command again with --verbose.
warning: build failed, waiting for other jobs to finish...
error: build failed

Ubuntu 20.04
stable-x86_64-unknown-linux-gnu (default)
rustc 1.47.0 (18bf6b4f0 2020-10-07)