X11xnn

X11xnn is an open-source neural network inference engine and compiler designed to deliver efficient execution of feedforward models on a range of CPU and GPU backends. The project aims to provide a compact, portable runtime with a focus on small footprint and ease of integration for desktop, server, and embedded applications. It accepts models from common formats such as ONNX and compiles them into optimized kernels that run on available hardware.

Key components include a graph compiler that performs operator fusion and constant folding, a runtime for executing the optimized graph, and a backend layer that provides implementations for x86-64, ARMv8, and optional accelerator backends. The backend layer uses architecture-specific micro-kernels and a scheduling system that maps high-level operators to the best available kernels.

Supported features typically include common neural network operators (convolutions, depthwise convolutions, fully connected layers, activations, pooling), quantization support (int8 and float16), and optional fusion of operations to reduce memory bandwidth. The project emphasizes portability and reproducibility across platforms, with a simple C API for integration and Python bindings for experimentation.

Licensing and development: x11xnn is released under an open-source license (project pages indicate MIT- or Apache-style licensing). It is maintained on public repositories with ongoing contributions from independent developers and users. Documentation covers installation, model import, API usage, and guidelines for extending backends.

See also: XNNPACK, ONNX, neural network inference engines.