ZenDNN (Zen Deep Neural Network) Library accelerates deep learning inference applications on AMD CPUs. This library, which includes APIs for basic neural network building blocks optimized for AMD CPUs, targets deep learning application and framework developers with the goal of improving inference performance on AMD CPUs across a variety of workloads, including computer vision, natural language processing (NLP), and recommender systems. ZenDNN leverages oneDNN/DNNL v2.6.3's basic infrastructure and APIs. ZenDNN optimizes several APIs and adds new APIs, which are currently integrated into TensorFlow, and PyTorch.
ZenDNN depends on:
- AOCL-BLAS is a high-performant implementation of the Basic Linear Algebra Subprograms (BLAS).
- FBGEMM (Facebook GEneral Matrix Multiplication) is a low-precision, high-performance matrix-matrix multiplications and convolution library for server-side inference.
AOCL-BLAS is a required dependency for ZenDNN, whereas the FBGEMM is optional dependency.
- Zen Deep Neural Network Library (ZenDNN)
- Table of Contents
- Scope
- Release Highlights
- Supported OS and Compilers
- Prerequisites
- AOCL-BLAS Library Installation
- Runtime Dependencies
- Build from Source
- Logs
- License
- Technical Support
The scope of ZenDNN is to support AMD EPYC™ CPUs on the Linux® platform. ZenDNN v5.1 offers optimized primitives, such as Convolution, MatMul, Elementwise, and Pool (Max and Average), Gelu, LayerNorm that improve performance of many convolutional neural networks, recurrent neural networks, transformer-based models, and recommender system models. For the primitives not supported by ZenDNN, execution will fall back to the native path of the framework.
Following are the highlights of this release:
- ZenDNN library is integrated with TensorFlow v2.19 (Plugin), and PyTorch v2.8 (Plugin).
- Python v3.9-v3.12 has been used to generate the TensorFlow v2.19 (Plugin) wheel files (*.whl).
- Python v3.9-v3.13 has been used to generate the PyTorch v2.8 (Plugin) wheel files (*.whl).
ZenDNN library is intended to be used in conjunction with the frameworks mentioned above and cannot be used independently.
The latest information on the ZenDNN release and installers is available on AMD.com portal (https://www.amd.com/en/developer/zendnn.html).
This release of ZenDNN supports the following Operating Systems (OS) and compilers:
Binaires will be supported on
- Ubuntu® 20.04, 22.04, 24.04
- Red Hat® Enterprise Linux® (RHEL) 8.6, 9.2, 9.5
- Anolis 8.8, 8.9
- SLES15 SP6, SP3
- RockyLinux 8.10, 9.4
Build from source will be supported on
- Ubuntu® 22.04, 24.04
- Red Hat® Enterprise Linux® (RHEL) 9.2, 9.5
- Anolis 8.8,8.9
- SLES15 SP6, SP3
- RockyLinux 8.10, 9.4
- GCC 12.2 and later
Theoretically, for wheel files any Linux based OS with GLIBC version later than 2.28 could be supported.
For C++ interface binaries, any Linux based OS with GLIBC version later than 2.28 could be supported.
For TF-JAVA interface binaries, any Linux based OS with GLIBC version 2.33 and later could be supported.
The following prerequisites must be met for this release of ZenDNN:
- AOCL-BLAS v5.1 must be installed.
AOCL-BLAS AOCL-BLAS is a high-performant implementation of the Basic Linear Algebra Subprograms (BLAS). The BLAS was designed to provide the essential kernels of matrix and vector computation and are the most commonly used computationally intensive operations in dense numerical linear algebra. This can be downloaded from https://github.com/amd/blis/releases/tag/AOCL-Sep2025-b1
Note: ZenDNN depends only on AOCL-BLAS and has no dependency on any other AOCL library.
The following points must be considered while installing AOCL-BLAS:
- Change to the preferred directory where AOCL-BLAS will be downloaded.
- This parent folder is referred to as folder
<compdir>in the steps below. - Assume that the parent folder for user setup follows this convention:
/home/<user-id>/my_work.
Complete the following steps to setup the GCC compiled AOCL-BLAS library:
- Execute the command
cd <compdir> - Download BLIS wget https://github.com/amd/blis/archive/refs/tags/AOCL-Sep2025-b1.tar.gz
- Execute the following commands:
tar -xvf AOCL-Sep2025-b1.tar.gz cd blis-AOCL-Sep2025-b1 make clean; make distclean; CC=gcc ./configure -a aocl_gemm --prefix=../amd-blis --enable-threading=openmp --enable-cblas amdzen; make -j install cd ../amd-blis/include mkdir LP64 cp blis/* LP64/ cd ../lib/ mkdir LP64 cp libblis-mt.* LP64/ cd ../ export ZENDNN_BLIS_PATH=$(pwd)
For example:
export ZENDNN_BLIS_PATH=/home/<user-id>/my_work/amd-blisThe bashrc file can be edited to setup ZENDNN_BLIS_PATH environment path. For example, in the case of GCC compiled AOCL-BLAS:
export ZENDNN_BLIS_PATH=/home/<user-id>/my_work/amd-blisZenDNN has the following runtime dependencies:
- GNU C library (glibc.so)
- GNU Standard C++ library (libstdc++.so)
- Dynamic linking library (libdl.so)
- POSIX Thread library (libpthread.so)
- C Math Library (libm.so)
- OpenMP (libomp.so)
- Python v3.9-v3.12 for TensorFlow v2.19(Plugin)
- Python v3.9-v3.13 for PyTorch v2.8 (Plugin)
Since ZenDNN is configured to use OpenMP, a C++ compiler with OpenMP 2.0 or later is required for runtime execution.
Clone ZenDNN git:
git clone https://github.com/amd/ZenDNN.git
cd ZenDNNZENDNN_BLIS_PATH should be defined. example:
export ZENDNN_BLIS_PATH=/home/<user-id>/my_work/amd-blis
make clean
source scripts/zendnn_build.sh gcc lpgemm_v5_0When new terminal is opened, user need to set up environment variables:
source scripts/zendnn_gcc_env_setup.shPlease note above scripts must be sourced only from ZenDNN Folder.
After the library is built on Linux host, user can run unit tests using:
source scripts/runApiTest.shCorresponding tests are located in the tests/api_tests directory. Library logs can be enabled with:
ZENDNN_LOG_OPTS=ALL:2 source scripts/runApiTest.shLogging is disabled in the ZenDNN library by default. It can be enabled using the environment variable ZENDNN_LOG_OPTS before running any tests. Logging behavior can be specified by setting the environment variable ZENDNN_LOG_OPTS to a comma-delimited list of ACTOR:DBGLVL pairs.
The different ACTORS are as follows:
| ACTORS | Usage |
|---|---|
| ALGO | Logs all algorithms executed |
| CORE | Logs all the core ZenDNN library operations |
| API | Logs all the ZenDNN API calls |
| TEST | Logs used in API tests, functionality tests and regression tests |
| PROF | Logs metadata for the op |
| FWK | Logs all the framework (TensorFlow, and PyTorch) specific calls |
For example:
- To turn on info logging, use ZENDNN_LOG_OPTS=ALL:2
- To turn off all logging, use ZENDNN_LOG_OPTS=ALL:-1
- To only log errors, use ZENDNN_LOG_OPTS=ALL:0
- To only log info for ALGO, use ZENDNN_LOG_OPTS=ALL:-1,ALGO:2
- To only log info for CORE, use ZENDNN_LOG_OPTS=ALL:-1,CORE:2
- To only log info for API, use ZENDNN_LOG_OPTS=ALL:-1,API:2
- To only log info for PROF (profile), use ZENDNN_LOG_OPTS=ALL:-1,PROF:3
- To only log info for FWK, use ZENDNN_LOG_OPTS=ALL:-1,FWK:2
Enable Log Profiling
To enable the log profiling of zendnn_primitive_create and zendnn_primitive_execute, set
ZENDNN_PRIMITIVE_LOG_ENABLE=1 ZENDNN_LOG_OPTS=ALL:-1,PROF:3
The Different Debug Levels (DBGLVL) are as follows:
enum LogLevel {
LOG_LEVEL_DISABLED = -1,
LOG_LEVEL_ERROR = 0,
LOG_LEVEL_WARNING = 1,
LOG_LEVEL_INFO = 2,
LOG_LEVEL_VERBOSE0 = 3,
LOG_LEVEL_VERBOSE1 = 4,
LOG_LEVEL_VERBOSE2 = 5
};Refer to the "LICENSE" file for the full license text and copyright notice.
This distribution includes third party software governed by separate license terms.
This third party software, even if included with the distribution of the Advanced Micro Devices software, may be governed by separate license terms, including without limitation, third party 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.
Please email Zendnn.Maintainers@amd.com for questions, issues, and feedback on ZenDNN.
Please submit your questions, feature requests, and bug reports on the GitHub issues page.
