oneAPI Deep Neural Network Library (oneDNN)

oneDNN is an open-source, cross-platform library of optimized deep learning operations for CPUs and GPUs. It provides highly optimized implementations of matrix multiplication, convolution, attention, normalization, and other operations, and is the acceleration layer behind PyTorch, TensorFlow, OpenVINO, and many other AI frameworks.

The library is optimized for Intel 64/AMD64 architecture based processors, Arm(R) 64-bit Architecture (AArch64)-based processors, and Intel Graphics. oneDNN has experimental support for the following architectures: NVIDIA* GPU, AMD* GPU, OpenPOWER* Power ISA (PPC64), IBMz* (s390x), and RISC-V.

oneDNN project is part of the UXL Foundation.

Highlights

• Single API, many devices: The same code targets Intel64/AMD64, and AArch64, RISC-V CPUs, and Intel GPUs.
• Hardware specialization: oneDNN detects the hardware features at runtime and JIT-generates kernels specialized for instruction set, number of cores, and cache sizes.
• Low-precision inference: Supports f32, bf16, f16, fp8, fp4, u8/s8, and u4/s4 data types with various quantization schemes including OCP microscaling formats.
• Operation fusion: Post-ops fuse activations, residual connections, and quantization into matmul and convolution. The Graph API fuses larger patterns such as scaled dot-product attention (SDPA) and grouped-query attention (GQA).
• Interoperability: Supports OpenMP, TBB, and custom threadpools on CPUs. SYCL, and OpenCL on GPUs.

Used By

oneDNN is intended for deep learning applications and framework developers interested in improving application performance on CPUs and GPUs. Deep learning practitioners should use one of the applications enabled with oneDNN:

• Apache SINGA
• DeepLearning4J*
• Flashlight*
• llama.cpp
• ONNX Runtime
• OpenNMT CTranslate2
• OpenVINO(TM) toolkit
• PaddlePaddle*
• PyTorch*
• Tensorflow*

Quick Start

Install pre-built oneDNN from conda-forge:

conda install -c conda-forge onednn

Build and run the matrix multiplication performance example, which measures matmul throughput for f32, f16, bf16, and s8/u8 data types:

wget https://raw.githubusercontent.com/uxlfoundation/oneDNN/main/examples/matmul_perf.cpp \
     https://raw.githubusercontent.com/uxlfoundation/oneDNN/main/examples/example_utils.hpp
g++ matmul_perf.cpp -I$CONDA_PREFIX/include -L$CONDA_PREFIX/lib \
    -ldnnl -o matmul_perf
LD_LIBRARY_PATH=$CONDA_PREFIX/lib ./matmul_perf cpu

Next steps:

• Examples and Tutorials cover quantization, fusion, Graph API, GPU execution, and more.
• Build from Source to customize the build.

If oneDNN is useful in your work, consider giving the project a star on GitHub. It helps others discover it.

Installation

You can download and install the oneDNN library using one of the following options:

• Binary Distribution: You can download pre-built binary packages from the following sources:
  • conda-forge: If the configuration you need is not available on the conda-forge channel, you can build the library using the Source Distribution.
  • Intel oneAPI:
    • Intel® oneAPI Toolkit
    • Intel® oneDNN package
• Source Distribution: You can build the library from source by following the instructions on the Build from Source page.

Documentation

• oneDNN Developer Guide and Reference explains the programming model, supported functionality, implementation details, and includes annotated examples.
• API Reference provides a comprehensive reference of the library API.
• Release Notes explain the new features, performance optimizations, and improvements implemented in each version of oneDNN.

System Requirements

oneDNN supports platforms based on the following architectures:

• Intel 64 or AMD64,
• Arm 64-bit Architecture (AArch64).
• OpenPOWER / IBM Power ISA.
• IBMz z/Architecture (s390x).
• RISC-V 64-bit (RV64).

WARNING

Power ISA (PPC64), IBMz (s390x), and RISC-V (RV64) support is experimental with limited testing validation.

The library is optimized for the following CPUs:

• Intel 64/AMD64 architecture
  • Intel Xeon(R) processor E3, E5, and E7 family v3+ lineups (formerly Haswell and Broadwell)
  • Intel Xeon Scalable processor (formerly Skylake, Cascade Lake, Cooper Lake, Ice Lake, Sapphire Rapids, and Emerald Rapids)
  • Intel Xeon CPU Max Series (formerly Sapphire Rapids HBM)
  • Intel Core Ultra processors (formerly Meteor Lake, Arrow Lake, Lunar Lake, and Panther Lake)
  • Intel Xeon 6 processors (formerly Sierra Forest and Granite Rapids)
  • future Intel Core processor with Intel AVX10.2 instruction set support (code name Nova Lake)
  • future Intel Xeon processor with Intel AVX10.2 instruction set support (code name Diamond Rapids)
• AArch64 architecture
  • Arm Neoverse(TM) N-Series and V-Series

On a CPU based on Intel 64 or on AMD64 architecture, oneDNN detects the instruction set architecture (ISA) at runtime and uses just-in-time (JIT) code generation to deploy the code optimized for the latest supported ISA. Future ISAs may have initial support in the library disabled by default and require the use of run-time controls to enable them. See CPU dispatcher control for more details.

WARNING

On macOS, applications that use oneDNN may need to request special entitlements if they use the hardened runtime. See the Linking Guide for more details.

The library is optimized for the following GPUs:

• Intel discrete GPUs:
  • Intel Arc(TM) A-Series Graphics (formerly Alchemist)
  • Intel Data Center GPU Flex Series (formerly Arctic Sound)
  • Intel Data Center GPU Max Series (formerly Ponte Vecchio)
  • Intel Arc B-Series Graphics and Intel Arc Pro B-Series Graphics (formerly Battlemage)
  • future discrete GPUs based on Xe3p-XPC architecture (code name Crescent Island)
• Intel Graphics integrated with:
  • Intel Graphics for Intel Core Ultra Series 1 processors (formerly Meteor Lake)
  • Intel Graphics for Intel Core Ultra Series 2 processors (formerly Arrow Lake and Lunar Lake)
  • Intel Graphics for Intel Core Ultra Series 3 processors (formerly Panther Lake)
  • Intel Graphics for Intel Core Series 3 processors (formerly Wildcat Lake)
  • Intel Graphics for future Intel Core Ultra processors (code name Nova Lake)

Requirements for Building from Source

oneDNN supports systems meeting the following requirements:

• Operating system with Intel 64/AMD64, AArch 64, PPC64, or s390x architecture support
• C++ compiler with C++11 standard support
• CMake 3.13 or later

The following tools are required to build oneDNN documentation:

• Doxygen 1.8.5 or later
• Doxyrest 2.1.2 or later
• Sphinx 7.4.7 or later
• sphinx-book-theme 1.2.0 or later
• sphinx-copybutton 0.5.2 or later
• graphviz 2.40.1

Configurations of CPU and GPU engines may introduce additional build time dependencies.

CPU Engine

oneDNN CPU engine is used to execute primitives on Intel 64/AMD64 based processors, 64-bit Arm Architecture (AArch64) processors, 64-bit Power ISA (PPC64) processors, IBMz (s390x), and compatible devices.

The CPU engine is built by default but can be disabled at build time by setting ONEDNN_CPU_RUNTIME to NONE. In this case, GPU engine must be enabled. The CPU engine can be configured to use the OpenMP, TBB or SYCL runtime. The following additional requirements apply:

• OpenMP runtime requires C++ compiler with OpenMP 2.0 or later standard support
• TBB runtime requires Threading Building Blocks (TBB) 2017 or later.
• SYCL runtime requires
  • Intel oneAPI DPC++/C++ Compiler
  • Threading Building Blocks (TBB)

Some implementations rely on OpenMP 4.0 SIMD extensions. For the best performance results on Intel Architecture Processors we recommend using the Intel C++ Compiler.

On a CPU based on Arm AArch64 architecture, oneDNN CPU engine can be built with Arm Compute Library (ACL) integration. ACL is an open-source library for machine learning applications and provides AArch64 optimized implementations of core functions. This functionality currently requires that ACL is downloaded and built separately. See Build from Source section of the Developer Guide for details. The minimum supported version of ACL is 53.1.0.

GPU Engine

oneDNN GPU engine is used to execute primitives on various accelerators including Intel integrated and discrete GPUs, NVIDIA GPUs, AMD GPUs, and other devices supporting SYCL programming language. The GPU engine is disabled in the default build configuration and can be enabled by setting ONEDNN_GPU_RUNTIME build option to value other than NONE. Target accelerator vendor must be selected at build time using ONEDNN_GPU_VENDOR build option.

WARNING

Linux will reset GPU when kernel runtime exceeds several seconds. The user can prevent this behavior by disabling hangcheck for Intel GPU driver. Windows has built-in timeout detection and recovery mechanism that results in similar behavior. The user can prevent this behavior by increasing the TdrDelay value.

The following additional requirements apply for Intel integrated and discrete GPUs:

• With OpenCL(TM) runtime:
  • OpenCL SDK (with OpenCL 1.2 support)
  • Intel Graphics Driver with support for OpenCL C 2.0, Intel subgroups support, and USM extensions support
• With SYCL runtime:
  • Intel oneAPI DPC++/C++ Compiler
  • OpenCL SDK (with OpenCL 3.0 support)
  • oneAPI Level Zero with support for API v1.11 or later
  • Intel Graphics Driver with support for OpenCL C 2.0, Intel subgroups support, and USM extensions support

The following additional requirements apply for NVIDIA GPUs:

• oneAPI DPC++ Compiler with support for CUDA or oneAPI for NVIDIA GPUs
• NVIDIA CUDA* driver
• cuBLAS 10.1 or later
• cuDNN 7.6 or later

WARNING

NVIDIA GPU support is experimental. General information, build instructions, and implementation limitations are available in the NVIDIA backend readme.

The following additional requirements apply for AMD GPUs:

• oneAPI DPC++ Compiler with support for HIP AMD or oneAPI for AMD GPUs
• AMD ROCm version 5.3 or later
• MIOpen version 2.18 or later (optional if AMD ROCm includes the required version of MIOpen)
• rocBLAS version 2.45.0 or later (optional if AMD ROCm includes the required version of rocBLAS)

WARNING

AMD GPU support is experimental. General information, build instructions, and implementation limitations are available in the AMD backend readme.

Other devices supporting SYCL programming model require oneAPI DPC++/C++ Compiler that supports the target GPU. Refer to generic GPU vendor documentation for additional details.

Runtime Dependencies

When oneDNN is built from source, the library runtime dependencies and specific versions are defined by the build environment.

Linux

Common dependencies:

• GNU C Library (libc.so)
• GNU Standard C++ Library v3 (libstdc++.so)
• Dynamic Linking Library (libdl.so)
• C Math Library (libm.so)
• POSIX Threads Library (libpthread.so)

Runtime-specific dependencies:

Windows

Common dependencies:

• Microsoft Visual C++ Redistributable (msvcrt.dll)

Runtime-specific dependencies:

macOS

Common dependencies:

• System C/C++ runtime (libc++.dylib, libSystem.dylib)

Runtime-specific dependencies:

Support

Ask questions and share your projects in GitHub Discussions. Submit feature requests and bug reports on the GitHub issues page.

You can also contact oneDNN developers via UXL Foundation Slack using #onednn channel.

Governance

oneDNN project is governed by the UXL Foundation and you can get involved in this project in multiple ways. It is possible to join the AI Special Interest Group (SIG) meetings where the groups discuss and demonstrate work using this project. Members can also join the Open Source and Specification Working Group meetings.

You can also join the mailing lists for the UXL Foundation to be informed of when meetings are happening and receive the latest information and discussions.

Contributing

We welcome community contributions to oneDNN. You can find the oneDNN release schedule and work already in progress towards future milestones in Github's Milestones section. If you are looking for a specific task to start, consider selecting from issues that are marked with the help wanted label.

See contribution guidelines to start contributing to oneDNN. You can also contact oneDNN developers and maintainers via UXL Foundation Slack using #onednn channel.

This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the Contributor Covenant code of conduct.

License

oneDNN is licensed under Apache License Version 2.0. Refer to the "LICENSE" file for the full license text and copyright notice.

This distribution includes third party software governed by separate license terms.

3-clause BSD license:

• Xbyak
• gtest
• Instrumentation and Tracing Technology API (ITT API)
• CMake

2-clause BSD license:

• Sphinx

Apache License Version 2.0:

• Xbyak_aarch64
• OpenCL^TM API Headers
• Arm(R) KleidiAI(TM)

Boost Software License, Version 1.0:

• Boost C++ Libraries

MIT License:

• Intel Graphics Compute Runtime for oneAPI Level Zero and OpenCL Driver
• Intel Graphics Compiler
• oneAPI Level Zero
• Doxyrest
• Intel Metrics Discovery Application Programming Interface
• spdlog
• sphinx-copybutton

This third-party software, even if included with the distribution of other software, may be governed by separate license terms, including without limitation, third party license terms, other software license terms, and open source software license terms. These separate license terms govern your use of the third party programs as set forth in the "THIRD-PARTY-PROGRAMS" file.

Security

Security Policy outlines our guidelines and procedures for ensuring the highest level of security and trust for our users who consume oneDNN.

---

Legal Information