rten-vecmath

This crate provides portable SIMD types that abstract over SIMD intrinsics on different architectures. Unlike std::simd this works on stable Rust. There is also functionality to detect the available instructions at runtime and dispatch to the optimal implementation.

This crate also contains SIMD-vectorized versions of math functions such as exp, erf, tanh, softmax etc. that are performance-critical in machine-learning models.

lib.rs:

rten-vecmath provides portable SIMD types for implementing vectorized functions that work across different architectures.

Portable SIMD types

The simd_vec module contains types and traits to support writing portable SIMD code, that works on stable Rust.

Supported architectures

SIMD wrappers are provided for the following architectures:

• Arm Neon
• AVX 2 / FMA
• AVX-512 (requires avx512 feature and nightly Rust)
• WebAssembly SIMD

There is also a scalar fallback that works on all platforms, but provides no performance benefit over non-SIMD code.

Vectorized math functions

This crate contains SIMD-vectorized implementations of various math functions that are commonly used in neural networks.

For each function in this library there are multiple variants, which typically include:

• A version that operates on scalars
• A version that reads values from an input slice and writes to the corresponding position in an equal-length output slice. These have a vec_ prefix.
• A version that reads values from a mutable input slice and writes the computed values back in-place. These have a vec_ prefix and _in_place suffix.

All variants use the same underlying implementation and should have the same accuracy.

See the source code for comments on accuracy.