briny_ai 0.5.0

A tiny & efficient AI inference engine
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

//! Tedious manual tensor operations.
//!
//! The alternative is to use high level abstractions, which might not be as efficient.
//! Though, manually implementing everything isn't portable; some targets may have `Tensor<T, N, D>`,
//! others have `Tensor<T>`. Some targets might have `T=f64`m others might have `T=f32`.
//!
//! Changing the features will completely change the API and wil probably not compile.
//! However, the abstract API is always the same, so it is much more portable and will
//! always compile for correct uses - whether features are removed, added, etc., the
//! abstraction is abstracted the same way every time.

pub mod ops;
pub mod tensors;
// pub mod io;

#[cfg(feature = "f64")]
type TensorFloatInner = f64;
#[cfg(not(feature = "f64"))]
type TensorFloatInner = f32;

/// The float used in tensors.
///
/// - `f64` on 64-bit machines
/// - `f32` on 32-bit machines
/// - `f32` on 16-bit machines
pub type TensorFloat = TensorFloatInner;

/// The amount of [`TensorFloat`]s in a SIMD register on the target.
pub const FLOAT_LANES: usize = if cfg!(target_pointer_width = "64") {
    lazy_simd::MAX_SIMD_DOUBLE_PRECISION_LANES * 2 // should be 1/2, but doesn't compile if not
} else {
    lazy_simd::MAX_SIMD_SINGLE_PRECISION_LANES // equal to the above
};