oneDNN

oneDNN is an open-source library that provides optimized implementations of deep learning operations such as matrix multiplication, convolution, and attention mechanisms. It targets CPUs across multiple architectures including x64, ARM AArch64, and RISC-V, as well as Intel GPUs. The library supports both inference and training workloads, with options for low-precision computation including integer and floating-point formats. oneDNN is developed by uxlfoundation and is released under the Apache 2.0 license.

The library is written in C++ and integrates with major deep learning frameworks and runtimes including PyTorch, TensorFlow, OpenVINO, and ONNX Runtime. It offers a range of primitives for neural network operations, performance profiling capabilities, and support for advanced features like quantization and graph-based computation. Developers use oneDNN to accelerate deep learning workloads on supported hardware platforms.