oneDNN

Why we included this project

oneDNN sits underneath a lot of the deep learning stack without most users ever seeing it. It is a performance library that supplies the low-level primitive operations, like convolutions and matrix multiplications, that frameworks such as PyTorch, TensorFlow, ONNX Runtime, and llama.cpp rely on for fast training and inference on CPUs and GPUs. Instead of shipping generic kernels, it generates and selects code tuned for the actual hardware at runtime, which is often why workloads run noticeably faster on Intel or Arm machines than you might expect. It is well optimized for Intel and AMD x86 processors, AArch64, and Intel graphics, with experimental support for NVIDIA, AMD, and other architectures. Framework developers and engineers profiling odd performance behavior will find the developer guide and API reference a practical way to understand, and sometimes override, the primitives under their stack.