📖 Overview

Welcome to HeliosRT, a specialized fork of TensorFlow Lite for Microcontrollers (TFLM), tailored for Ambiq's Apollo family of ultra-low-power System-on-Chips (SoCs). This project enhances TFLM by optimizing it to leverage the advanced hardware intrinsics of the Apollo series, such as Matrix Vector Extensions (MVE) and Digital Signal Processing (DSP) instructions.

Why HeliosRT?

HeliosRT addresses the unique demands of ultra-low-power devices, offering developers the tools to build AI applications that require minimal energy consumption without sacrificing performance. By harnessing the specific capabilities of Ambiq's SoCs, HeliosRT enables more efficient neural network inference on devices like smart watches, fitness trackers, and other smart IoT devices.

🚀 Key Features

• Optimized Performance: Utilizes MVE and DSP hardware capabilities to enhance computational efficiency and speed.
• Energy Efficiency: Designed to minimize power usage, extending the battery life of edge devices.
• Broad Compatibility: Supports a wide range of Ambiq's Apollo SoCs, ensuring versatile applications across different hardware.

Explore our Getting Started guide to dive into the development with HeliosRT, or check out the Benchmarks section to see how HeliosRT performs under various conditions. Whether you are developing for wearable technology or other smart devices, HeliosRT provides a robust framework for embedding AI into your projects.

📚 Quick Links

• Install HeliosRT and getting up and running in minutes.   Install HeliosRT

• Usage Examples showcasing real-world applications and best practices.   Usage Examples

• Performance Benchmarks comparing HeliosRT to other frameworks.   Performance Benchmarks

Getting Started

Jumpstart your development with neuralSPOT, a robust AI SDK optimized for Ambiq's ultra-low-power Apollo SoCs. This toolkit provides comprehensive resources, including the latest stable releases of HeliosRT, fully configured for both GCC and Arm Clang across debug and release builds. For a detailed step-by-step guide, refer to our Getting Started documentation.

Operator Support Matrix

Below is a detailed matrix that outlines the support for various operators across different backends and data types within HeliosRT. Each entry indicates whether an operator is supported (Yes), not supported (No), or not applicable (N/A) for the specified data type and backend. "N/A" is used where the operation does not logically apply to the data type, whereas "No" indicates that the operation could be supported but currently isn't.

The implementations are categorized under three main technologies: - C: Standard C implementation. - DSP: Utilizes Digital Signal Processing instructions. - MVE: Uses Matrix Vector Extensions.

Operator	C int8	C int16	C int4*	DSP int8	DSP int16	DSP int4*	MVE int8	MVE int16	MVE int4*
add	Yes	Yes	N/A	Yes	Yes	N/A	Yes	Yes	N/A
batch_matmul	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
batch_to_space_nd	Yes	Yes	No	No	No	No	No	No	No
cast	Yes	Yes	No	No	No	No	No	No	No
comparisons	Yes	Yes	No	No	No	No	No	No	No
concatenation	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
conv	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
cumsum	Yes	Yes	No	No	No	No	No	No	No
depth_to_space	Yes	Yes	No	No	No	No	No	No	No
depthwise_conv	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
dequantize	Yes	Yes	No	No	No	No	No	No	No
elementwise	Yes	Yes	No	No	No	No	No	No	No
elu	Yes	Yes	No	No	No	No	No	No	No
embedding_lookup	Yes	Yes	No	No	No	No	No	No	No
expand_dims	Yes	Yes	No	No	No	No	No	No	No
fill	Yes	Yes	No	No	No	No	No	No	No
fully_connected	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
gather_nd	Yes	Yes	No	No	No	No	No	No	No
gather	Yes	Yes	No	No	No	No	No	No	No
hard_swish	Yes	No	No	No	No	No	No	No	No
l2norm	Yes	Yes	No	No	No	No	No	No	No
leaky_relu	Yes	Yes	No	Yes	No	No	Yes	Yes	No
log_softmax	Yes	Yes	No	No	No	No	No	No	No
logistic	Yes	Yes	No	Yes	No	No	Yes	Yes	No
lstm	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
minimum	Yes	Yes	N/A	No	No	N/A	Yes	Yes	N/A
maximum	Yes	Yes	N/A	No	No	N/A	Yes	Yes	N/A
mirror_pad	Yes	Yes	No	No	No	No	No	No	No
mul	Yes	Yes	N/A	Yes	Yes	N/A	Yes	Yes	N/A
neg	Yes	Yes	No	No	No	No	No	No	No
pack	Yes	Yes	No	No	No	No	No	No	No
pad	Yes	Yes	N/A	No	No	N/A	Yes	Yes	N/A
max_pooling	Yes	Yes	N/A	Yes	Yes	N/A	Yes	Yes	N/A
avg_pooling	Yes	Yes	N/A	Yes	Yes	N/A	Yes	Yes	N/A
prelu	Yes	Yes	No	No	No	No	No	No	No
resize_bilinear	Yes	Yes	No	No	No	No	No	No	No
select	Yes	Yes	No	No	No	No	No	No	No
slice	Yes	Yes	No	No	No	No	No	No	No
softmax	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
space_to_batch_nd	Yes	Yes	No	No	No	No	No	No	No
space_to_depth	Yes	Yes	No	No	No	No	No	No	No
split	Yes	Yes	No	No	No	No	No	No	No
square_difference	Yes	Yes	No	No	No	No	No	No	No
strided_slice	Yes	Yes	No	No	No	No	No	No	No
sub	Yes	Yes	No	Yes	No	N/A	No	No	No
svdf	Yes	Yes	No	Yes	Yes	No	Yes	Yes	No
tanh	Yes	Yes	No	Yes	No	No	Yes	No	No
transpose_conv	Yes	No	No	Yes	No	No	Yes	No	No
tranpose	Yes	Yes	No	No	No	No	No	No	No
unpack	Yes	Yes	No	No	No	No	No	No	No
zeros_like	Yes	Yes	No	No	No	No	No	No	No

---

Ready to dive in? Head over to the Getting Started guide and generate your first module in minutes.

📜 License