Crate embedded_nn 

embedded-nn

embedded-nn is a pure Rust, #![no_std] neural network inference library for microcontrollers and embedded targets, inspired by ARM’s CMSIS-NN and TensorFlow Lite Micro.

Modules

• types: Dimensional shapes, parameter structures, error types.
• support: Fixed-point quantization math, rounding division, bit operations.
• activations: ReLU, ReLU6, LeakyReLU, Sigmoid, Tanh.
• basic_math: Elementwise addition, subtraction, and multiplication.
• convolution: 2D Convolution, 1x1 Convolution, Depthwise Convolution, Transposed Conv, 1D Temporal Conv.
• fully_connected: Fully Connected (Linear / Dense) layers and Batch Matrix Multiplication (BatchMatMul).
• pooling: Max Pooling, Average Pooling.
• softmax: Softmax activation.
• [concat]: Depthwise concatenation.
• pad: Tensor padding.
• transpose: Matrix and spatial transposition.
• reshape: Reshaping operations.
• simd: Target SIMD vectorization abstractions, ARM DSP SMLAD assembly, and dot-product acceleration.
• subbyte: Sub-byte 4-bit (s4) quantization layers and packing helpers.
• recurrent: Recurrent neural network layers (LSTM cell, SVDF filter with 8-bit & 16-bit state).
• float_ops: Floating-point (f32 & IEEE-754 f16) fallback layers.
• [dsp_bridge] (feature embedded-dsp): Quantize embedded-dsp FFT output into this crate’s tensors.

Re-exports

pub use support::clamp;

pub use support::divide_by_power_of_two;

pub use support::doubling_high_mult_no_sat;

pub use support::pack_q15x2_32x1;

pub use support::pack_s8x4_32x1;

pub use support::requantize;

pub use support::requantize_s64;

pub use types::Activation;

pub use types::Context;

pub use types::ConvParams;

pub use types::Dims;

pub use types::DwConvParams;

pub use types::Error;

pub use types::FcParams;

pub use types::PerChannelQuantParams;

pub use types::PerTensorQuantParams;

pub use types::PoolParams;

pub use types::QuantParams;

pub use types::Result;

pub use types::SoftmaxParams;

pub use types::Tile;

pub use convolution::convolve_1_x_n_s8;

pub use convolution::transpose_conv_s8;

pub use float_ops::f16_to_f32;

pub use float_ops::f32_to_f16;

pub use fully_connected::batch_matmul_s16;

pub use fully_connected::batch_matmul_s8;

pub use recurrent::lstm_step_s16;

pub use recurrent::lstm_step_s8_s16;

pub use recurrent::svdf_s8;

pub use recurrent::svdf_state_s16_s8;

pub use recurrent::LstmGateParams;

pub use simd::vec_dot_s16;

pub use simd::vec_dot_s8;

pub use subbyte::convolve_s4;

pub use subbyte::fully_connected_s4;

pub use subbyte::pack_s4_pair;

pub use subbyte::unpack_s4_pair;

Modules

activations

Activation functions for quantized tensors.

basic_math

Basic elementwise mathematical operations on quantized tensors.

concat

Concatenation operations along channel/depth dimension for quantized tensors.

convolution

Convolution layer operations for quantized neural networks.

float_ops

Floating-point (f32 and f16) fallback operations and layers.

fully_connected

Fully Connected (Linear / Dense) and Batch Matrix Multiplication operations.

pad

Tensor padding operations for quantized tensors.

pooling

Pooling layer operations (Max Pooling, Average Pooling) for quantized neural networks.

recurrent

Advanced recurrent neural network layers (LSTM, SVDF).

reshape

Tensor reshape operations.

simd

Target SIMD vectorization abstractions and acceleration hooks.

softmax

Softmax activation operations for quantized neural networks.

subbyte

Sub-byte 4-bit (s4) quantization operations and layers.

support

Fixed-point math and support helper functions for embedded-nn.

transpose

Tensor transpose operations.

types

Core types and parameter structures for embedded-nn.