common_traits

A collection of traits and types that can be used to write code that is generic over numerical types. It also provides atomic floats implemented using the integer atomic type with the same number of bits, and support for half-precision floats via the crate half.

Additionally, there are a few traits missing from the standard library, such as Sequence, variants of existing library traits such as Rng and Hash, and macros like invariant.

Finally, we provide traits for casting between types, such as UpcastableInto, and fast implementations of a few primitives such as FastRange and SelectInWord.

Everything is experimental and I'll change them to my needs, respecting semantic versioning. :)

Examples

Mixed precision generic dot products!

use common_traits::*;

#[inline]
pub fn dot_product<MT: Number, RT: Number, A, B>(a: A, b: B) -> RT
where
    A: Sequence,
    B: Sequence,
    A::Item: To<MT>,
    B::Item: To<MT>,
    MT: To<RT>,
    RT: To<MT>,
{
    // Ensure compatibility of the vectors
    assert_eq!(a.len(), b.len());

    // Compute the dot product
    let mut accum = RT::ZERO;
    for (a, b) in a.iter().zip(b.iter()) {
        accum = (a.to()).mul_add(b.to(), accum.to()).to();
    }

    accum
}

let x: Vec<f32> = vec![1.0, 2.0, 3.0];
let w: Vec<u8> = vec![3, 2, 1];
// compute the dot product between f32 and u8, casting to f64 and
// accumulating as u16
let res: u16 = dot_product::<f64, _, _, _>(&x, &w);
println!("{:?}", res);

Numerical traits at a glance

The numerical traits dependency chain is the following. Black arcs represent the trait dependencies, the blue arcs represent the possibility to access an associated type implementing that trait.

Why?

The point of making this crate public is to be able to discuss this as it covers many core missing features from Rust.

The traits in this crate are similar to the ones from num-traits but they are more interconnected (the blue arcs in the previous graph), which makes it possible to write generic code (e.g., code mixing a type and its associated atomic type) more easily and with less trait bounds.

Summary

A highlight of common_traits' most noteworthy features.

Macros

This crate adds the following macros: invariant, invariant_eq, invariant_ne, which are similar to the debug_assert macros, which get checked during debug runs and get replaced with core::hint::unreachable_unchecked on release builds.

Structs

This crate adds emulated atomic floats implemented via fetch_update for the following types:

• f64 as AtomicF64
• f32 as AtomicF32
• half::f16 as AtomicF16
• half::bf16 as AtomicBF16

Numerical Traits

This crate provides the following traits for numerical types:

• Number: something that can be added, subtracted, multiplied, divided and has a zero and a one.
• FiniteRangeNumber: a Number which has a minimum and a maximum.
• Float: float numbers.
• Integer: an integer number represented as a sequence of bits.
• SignedInt: a signed integer represented in two's complement.
• UnsignedInt: an unsigned integer.

Atomic Numerical Traits

There are two main traits for working with atomic values:

• Atomic: for values that can be read and written atomically.
• IntoAtomic: for values that can be converted into atomic types.

Each numerical trait has an atomic equivalent:

• AtomicNumber
• AtomicFiniteRangeNumber
• AtomicFloat
• AtomicInteger
• AtomicSignedInt
• AtomicUnsignedInt

Miscellaneous Traits

The crate also contains a couple of extra traits:

• Sequence, SequenceMut, and SequenceGrowable to abstract over slices and other sequence-like types.
• AsBytes, ToBytes, and FromBytes are traits used to convert types to and from byte arrays.
• NonZero, a version of Self that cannot be zero; UnsignedInt and SignedInt have an associated type implementing this.
• FastRange for faster div, mod, and range operations.
• SelectInWord to find the position of the i-th 1 or 0 in words of memory.
• Splat to broadcast a smaller type on a larger type, mainly used for SWAR.
• Rng for a generic random number generator.
• SameAs, an unsafe marker trait guaranteeing that a type and its atomic variant have the same memory layout.
• Hasher, that is like std::hash::Hasher but allows returning a generic type instead of a u64.
• SeedableHasher, a standard way to initialize hashers.

Conversion Traits

Traits for conversion between types are also provided:

• To, to cast primitive values using as.
• DoubleType and HalfType can be used to access bigger or smaller types in a generic way.
• UpcastableInto and UpcastableFrom to cast primitive values which cannot lose precision.

• DowncastableInto and DowncastableFrom to cast primitive values which can lose precision.

• CastableInto and CastableFrom to cast primitive values which may or may not lose precision. This is the union of DowncastableInto and UpcastableInto.

The difference between CastableInto and To is that CastableInto does not allow casting from f32 to u32 for example, because CastableInto is implemented only between integers and between floats, while To is implemented for all primitive types.

Features

This crate has the following features:

• simd: to enable portable_simd and be able to do generic SIMD code
• std: to enable standard library support
• alloc: to enable allocator support for Vec/Box without full std
• half: to enable support for half::f16 (experimental)
• nightly_f16: to enable support for the native f16 type. Requires a nightly compiler and is mutually exclusive with half.