oneDNN

oneDNN, originally released as MKL-DNN, is an open-source performance library of deep learning building blocks designed to accelerate computation on CPUs and, increasingly, other hardware through the oneAPI ecosystem. It provides optimized implementations of common deep learning primitives such as convolutions, pooling, normalization, activation functions, softmax, and recurrent layers, with a focus on high throughput for both inference and training workloads. The library is intended to be embedded by machine learning frameworks and high-performance applications that require portable, highly optimized kernels rather than bespoke implementations.

The API is exposed as a low-level C interface (with a C++ wrapper) that organizes work into primitives and primitive descriptors. This design lets frameworks select the most efficient kernel for a given operation and memory layout at runtime. oneDNN supports multiple data types, including f32, f16, and bf16 in recent releases, and exposes a variety of memory formats to optimize data locality. Core optimizations target CPU instruction sets such as AVX2, AVX-512, and NEON, enabling efficient execution on x86 and ARM platforms. The library is generally used as a backend for CPU computation in major frameworks and can be compiled into builds to accelerate workloads.

oneDNN is released under the Apache License 2.0 and is hosted on GitHub. It is part of the broader oneAPI ecosystem and benefits from Intel and community contributions. The project emphasizes performance portability across hardware and software stacks, with ongoing additions of primitives, data types, and platform support.