OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Onednn

oneDNN is an open-source C++ library of optimized deep-learning operations, including matrix multiplication, convolution, attention and normalization, for use in machine-learning applications and inference software.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Onednn — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use oneDNN when developing deep-learning software and you need a library of optimized operations that can specialize computation for supported hardware.

Who it is for

  • Deep-learning application and framework developers seeking optimized computation primitives.
  • Inference software developers targeting supported CPU architectures or Intel GPUs.

Documented uses

  • Add optimized matrix multiplication, convolution or attention operations to a deep-learning application.
  • Use oneDNN as a computation backend in inference software.
  • Build CPU-focused inference software for supported architectures.
  • Integrate optimized operations into software using frameworks such as PyTorch, TensorFlow, OpenVINO or ONNX Runtime.
Before you choose

Requirements & limits

The repository describes oneDNN as open source under the Apache-2.0 license. It documents installation through conda-forge and Intel oneAPI packages, as well as building from source.

  • The documented standard CPU targets are Intel 64/AMD64 and AArch64; Intel Graphics is also supported.
  • NVIDIA and AMD GPUs, Power ISA, IBM Z and RISC-V are listed as experimental in the README.
  • The library offers CPU threading interoperability with OpenMP, TBB and custom thread pools, and GPU interoperability with SYCL and OpenCL.
  • The README documents low-precision data types and fusion options for supported operations.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install oneDNN and build the matrix-multiplication example

Install the conda-forge package and build and run the repository’s documented CPU performance example.

Before you start

  • Conda and a C++ compiler
  • The oneDNN package installed from conda-forge

Steps

  1. Install the package with conda install -c conda-forge onednn.

  2. Download examples/matmul_perf.cpp and examples/example_utils.hpp from the oneDNN repository.

  3. Compile the example using the package include and library directories and link against dnnl.

  4. Set LD_LIBRARY_PATH to the package library directory and run the example with the cpu argument.

Expected outcome

The documented example measures matrix-multiplication throughput for f32, f16, bf16 and s8/u8 data types.

Guide limitations

  • The documented run is for CPU execution.
  • The README’s commands use LD_LIBRARY_PATH, so the shown run instructions are specific to that environment.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
7 October 2026
Original website title
uxlfoundation/oneDNN · Official homepage title · retained in its source language
How we review sourcesSuggest a correction