tract-linalg

Crates.io	tract-linalg
lib.rs	tract-linalg
version
source	src
created_at	2019-01-16 15:41:28.476283
updated_at	2024-12-05 10:35:14.91332
description	Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
homepage
repository	https://github.com/snipsco/tract
max_upload_size
id	108938
Cargo.toml error:	TOML parse error at line 18, column 1 \| 18 \| autolib = false \| ^^^^^^^ unknown field `autolib`, expected one of `name`, `version`, `edition`, `authors`, `description`, `readme`, `license`, `repository`, `homepage`, `documentation`, `build`, `resolver`, `links`, `default-run`, `default_dash_run`, `rust-version`, `rust_dash_version`, `rust_version`, `license-file`, `license_dash_file`, `license_file`, `licenseFile`, `license_capital_file`, `forced-target`, `forced_dash_target`, `autobins`, `autotests`, `autoexamples`, `autobenches`, `publish`, `metadata`, `keywords`, `categories`, `exclude`, `include`
size	0

Mathieu Poumeyrol (kali)

documentation

linalg stands for "linear algebra". This is a misnamer. This crates contains low-level, architecture dependant optimisations used by tract-core.

MatMatMul: Extended matrix*matrix product:
- inspired by Gotoblass and BLIS micro kernel approach
- extended for convolution friendly addressing (fused img2col)
- fused output pipeline (min, max, and a few more simple, fast ops)
- f32*f32 -> f32 (à la sgemm)
- i8*i8 -> i32 accumulator -> i32 storage
- i8*i8 -> i32 accumulator -> i8 (with channel zeropoint and scale, and re-quantization pipeline)
f32 sigmoid and f32 tanh: at f32 precision, by a rationale function (no exponentiation)
byte-to-byte lookup table

Commit count: 6847