slas

Static Linear Algebra System

A linear algebra system with a focus on performance, static allocation, statically shaped data and copy-on-write (aka cow) behavior. Safe and fast bindings for blas/blis are also provided out of the box.

The mission

The goal of slas is to provide the best perfomance given the most amount of information that can be known at compile time. This mainly includes shapes and sizes of algebraic objects, target architecture and available hardware features/devices.

Please keep in mind that slas specializes in cases where binaries are compiled and executed on the same system and thus is primarily intended for native compilation.

NOTE: Slas might still be very broken when not using native compilation.

Specialization in hardware and usecases is attempted to be done with the modular backend system, which will support custom allocators in the future.

What is BLAS?

Example

A StaticCowVec can be created both with the moo and the cow_vec macro. They have the exact same syntax, but the cow_vec macro is a great choice for the serious programmer. Better documentation for the moo macro can be found here.

use slas::prelude::*;
use slas::cow_vec;

let a = moo![f32: 1, 2, 3.2];
let b = cow_vec![f32: 3, 0.4, 5];

println!("Dot product of {a:?} and {b:?} is {:?}", a.dot(&b));
println!("{a:?} + {b:?} = {:?}", a.add(&b));

By default slas tries to select a backend for you, you can also choose a static backend yourself. (More about what exactly a backend is and how to configure it.)

use slas::prelude::*;
let a = moo![on slas_backend::Rust:f32: 1, 2, 3.2];
// This will only use rust code for all operations on a

use slas::prelude::*;
let a = moo![on slas_backend::Blas:f32: 1, 2, 3.2];
// This will always use blas for all operations on a

By default slas will choose the backend that is assumed to be the fastest, given options set in environment during build (more about that here).

The StaticCowVec dereferences to StaticVecUnion, which in turn dereferences to [T; LEN], so any method implemented for [T;LEN] can also be called on StaticCowVec and StaticVecUnion.

More example code here.

What is a cow and when is it useful?

The copy-on-write functionality is inspired by std::borrow::cow. The idea is simply that allocations (and time) can be saved, by figuring out when to copy at runtime instead of at compiletime. This can be memory inefficient at times (as an enum takes the size of its largest field + tag size), which is why you can optionally use StaticVecUnions and StaticVecs instead. You can call moo, moo_ref and mut_moo_ref on any type that implements StaticVec to cast it to a appropriate type for it's use-case, with zero overhead.

moo_ref returns a StaticVecRef, which is just a type alias for a reference to a StaticVecUnion. This is most efficient when you know you don't need mutable access or ownership of a vector.

mut_moo_ref returns a MutStaticVecRef. This is a lot like moo_ref, but is useful when you want to mutate your data in place (fx if you wan't to normalize a vector). You should only use this if you want mutable access to a vector WITH side effects.

moo returns a StaticCowVec that references self. This is useful if you don't know if you need mutable access to you vector and you don't want side effects. If you want to copy data into a StaticCowVec then StaticCowVec::from is what you need.

moo_owned will just return a StaticVecUnion. This is useful when you really just want a [T; LEN], but you need methods only implemented for a StaticVecUnion.

Example of cow behavior

use slas::prelude::*;

let source: Vec<f32> = vec![1., 2., 3.];
let mut v = source.moo();

// Here we mutate v,
// so the content of source will be copied into v before the mutation occours.
v[0] = 0.;

assert_eq!(**v, [0., 2., 3.]);
assert_eq!(source, vec![1., 2., 3.]);

The borrow checker won't allow mutating source after v is created, because assignment to borrowed values is not allowed. This can be a problem in some situations.

use slas::prelude::*;

let mut source: Vec<f32> = vec![1., 2., 3.];
let mut v = unsafe { StaticCowVec::<f32, 3>::from_ptr(source.as_ptr()) };

// Here we can mutate source, because v was created from a raw pointer.
source[1] = 3.;
v[0] = 0.;
source[2] = 4.;

assert_eq!(**v, [0., 3., 3.]);
assert_eq!(source, vec![1., 3., 4.]);

In the example above, you can see v changed value the first time source was mutated, but not the second time. This is because v was copied when it was mutated.

Matrix example

use slas::prelude::*;
use slas_backend::*;

let a = moo![f32: 1..=6].matrix::<Blas, 2, 3>();
let b = moo![f32: 1..=6].matrix::<Blas, 3, 2>();
let c = a.matrix_mul(&b);
let d = b.vector_mul(&[1., 2.]);

assert_eq!(c, [22., 28., 49., 64.]);
assert_eq!(d, [5., 11., 17.]);

println!("{a:.0?} * {b:.0?} = {:.0?}", c.matrix::<Blas, 2, 2>());

In slas there is a Matrix type and a Tensor type. A 2D tensor can be used instead of a Matrix for most operations. A matrix dereferences into a 2D tensor and 2D tensors implement Into<Matrix>, both operations have no overhead, as only changes type information. The matrix type has some additional optional generic arguments, including IS_TRANS, which will be true if the matrix has been lazily transposed at compiletime. If this information is not needed for a matrix operation, it should be implemented for a 2D tensor. When indexing into a 2D tensor [usize; 2] will be used, which takes columns first, where as using (usize, usize) for the matricies, will be rows first.

use slas::prelude::*;
use slas_backend::*;

let a = moo![f32: 1..=6].matrix::<Blas, 2, 3>();

assert_eq!((*a)[[0, 1]], a[(1, 0)]);

Tensor example

At the moment tensors can't do much.

NOTE: that tensors (and therefore also matricies) always need to have a associated backend.

use slas::prelude::*;
let t = moo![f32: 0..27].reshape([3, 3, 3], slas_backend::Rust);
assert_eq!(t[[0, 0, 1]], 9.);

let mut s = t.index_slice(1).matrix();

assert_eq!(s[(0, 0)], 9.);
assert_eq!(s.transpose()[(1, 0)], 10.);

That's pretty much it for now...

Why not just use ndarray (or alike)?

Slas can be faster than ndarray in some specific use cases, like when having to do a lot of allocations, or when using referenced data in vector operations. Besides slas should always be atleast as fast as ndarray, so it can't hurt.

Ndarray will always use the backend you choose in your Cargo.toml. With slas you can choose a backend in code and even create your own backend that fits your needs.

Static allocation and the way slas cow behavior works with the borrow checker, also means that you might catch a lot of bugs at compiletime, where ndarray most of the time will let you get away with pretty much anything. For example taking the dot product of two vectors with different sizes, will cause a panic in ndarray and a compiletime error in slas.

Installation

By default slas will assume you have blis installed on your system. You can pretty easily statically link and compile blis, by disabeling default-features and enabelig the blis-static feature. If you want tos choose your own blas provider please set dependencies.slas.default-features = false in your Cargo.toml, and refer to blas-src for further instructions. Remember to add extern crate blas_src; if you use blas-src as a blas provider.

On the crates.io version of slas (v0.1.0 and 0.1.1) blis is compiled automatically.

For now, if you want to use the newest version of slas, you need to install blis/blas on your system.

• On Arch linux blis-cblas v0.7.0 from the AUR has been tested and works fine.
• On Debian you can simply run apt install libblis-dev.
• On Windows openblas-src has been tested. This mean you will need to disable slas default features, follow the installation instructions in the openblas readme and add extern crate openblas_src to your main file.

Enviroment variables

The backend being chosen to use when none is specified, depends on environment variables.

For example SLAS_BLAS_IN_DOT_IF_LEN_GE=50, will use blas by default, for any dot product operation performned on vectors with greater than or equal to 50 elements. SLAS_BLAS_IN_DOT_IF_LEN_GE can be found as a constant in slas::config::BLAS_IN_DOT_IF_LEN_GE.

Again, this is only applicable when no backend is not specified for a vector (fx moo![f32: 1, 2].dot(moo![2, 1])).

Variables and default values

SLAS_BLAS_IN_DOT_IF_LEN_GE = 750

Misc

• Slas is still in very early days, and is subject to a lot of breaking changes.
• Benchmarks, tests and related

TODO

Progress and todos are on trello!

License: Apache-2.0