Skip to content

BUILD_FROM_SOURCE.md

Latest commit

 

History

History
144 lines (102 loc) · 4.66 KB

BUILD_FROM_SOURCE.md

File metadata and controls

144 lines (102 loc) · 4.66 KB

Building the ZenTF C++ Package from Source

This guide walks through building the TensorFlow-ZenDNN plug-in C++ package (ZENTF_v5.2.1_C++_SOURCE_BUILD) from source. The package enables ZenDNN-accelerated inference via the TensorFlow C++ API on AMD CPUs.

Prerequisites

Requirement Details
OS Any Linux distribution (tested on Ubuntu, RHEL)
Conda Active conda environment
TensorFlow v2.16.0 -- v2.21.0 installed in the environment
Bazel Version matching the .bazelversion in the repo
GCC >= 13.3
Python 3 Available as python3

Directory Layout

ZenDNN_TensorFlow_Plugin repository must be cloned:

<workspace>/
└── ZenDNN_TensorFlow_Plugin/    # TensorFlow plug-in repo

Step 1 -- Install TensorFlow

pip install tensorflow==2.21.0   # or any version in [2.16.0, 2.21.0]

Verify the installation:

python3 -c "import tensorflow as tf; print(tf.__version__)"

Step 2 -- Configure the Build

cd <workspace>/ZenDNN_TensorFlow_Plugin
bazel clean --expunge --async
./configure

The configure script auto-detects your Python environment and TensorFlow installation. It writes TF_HEADER_DIR and TF_SHARED_LIBRARY_DIR into the Bazel configuration, pointing at the pip-installed TensorFlow headers and shared libraries.

Step 3 -- Build the Plug-in Shared Library

bazel build -c opt //tensorflow_plugin:libamdcpu_plugin_cc.so \
    --verbose_failures \
    --copt=-Wall \
    --copt=-Werror \
    --spawn_strategy=standalone

This produces the plug-in shared library at:

bazel-bin/tensorflow_plugin/libamdcpu_plugin_cc.so

Step 4 -- Assemble the C++ Package

Create the output directory and copy the build artifacts:

cd <workspace>
mkdir -p ZENTF_v5.2.1_C++_SOURCE_BUILD/lib-tensorflow-plugins

# Plug-in library
cp ZenDNN_TensorFlow_Plugin/bazel-bin/tensorflow_plugin/libamdcpu_plugin_cc.so \
   ZENTF_v5.2.1_C++_SOURCE_BUILD/lib-tensorflow-plugins/

# OpenMP runtime
cp ZenDNN_TensorFlow_Plugin/bazel-bin/tensorflow_plugin/libamdcpu_plugin_cc.so.runfiles/llvm_openmp/libiomp5.so \
   ZENTF_v5.2.1_C++_SOURCE_BUILD/lib-tensorflow-plugins/

# Helper scripts and sample code
cp ZenDNN_TensorFlow_Plugin/scripts/c++/zentf_cc_api_setup.sh  ZENTF_v5.2.1_C++_SOURCE_BUILD/
cp ZenDNN_TensorFlow_Plugin/scripts/zentf_env_setup.sh         ZENTF_v5.2.1_C++_SOURCE_BUILD/
cp ZenDNN_TensorFlow_Plugin/examples/c++/sample_inference.cpp   ZENTF_v5.2.1_C++_SOURCE_BUILD/

Step 5 -- Download and Extract TensorFlow Headers & Libraries

The C++ API requires TensorFlow headers and shared libraries. Download the matching wheel and extract them:

TF_VERSION="2.21.0"    # must match the installed version

mkdir -p tf_wheel_tmp
pip download "tensorflow==${TF_VERSION}" --no-deps --only-binary=:all: -d tf_wheel_tmp
unzip tf_wheel_tmp/tensorflow*.whl -d "ZENTF_v5.2.1_C++_SOURCE_BUILD/tensorflow_${TF_VERSION}"
rm -rf tf_wheel_tmp

Create the required symlinks:

cd "ZENTF_v5.2.1_C++_SOURCE_BUILD/tensorflow_${TF_VERSION}/tensorflow"
ln -s libtensorflow_cc.so.2        libtensorflow_cc.so
ln -s libtensorflow_framework.so.2 libtensorflow_framework.so
cd <workspace>

Step 6 -- Set Environment Variables

cd <workspace>/ZENTF_v5.2.1_C++_SOURCE_BUILD

# Sets TF_CC_API_ZENDNN_ROOT, LIBRARY_PATH, LD_LIBRARY_PATH,
# and creates the libomp.so.5 symlink.
source zentf_cc_api_setup.sh

# Enables ZenDNN plug-in and sets recommended performance tuning variables
# (KMP, OMP, ZENDNNL_MATMUL_ALGO, etc.).
source zentf_env_setup.sh

Note: The zentf_env_setup.sh ships with OMP defaults tuned for AMD Turin. Adjust OMP_NUM_THREADS and related variables to match your hardware.

Step 7 -- Compile and Run the Sample

Please follow steps in Build C++ inference application to build and run the sample inference application.

Troubleshooting

Issue Resolution
./configure fails to find TensorFlow Ensure TensorFlow is installed in the active conda env, or set TF_HEADER_DIR / TF_SHARED_LIBRARY_DIR manually.
Bazel version mismatch Check .bazelversion in the repo and install the required Bazel version.
libtensorflow_cc.so: cannot open shared object file Run source zentf_cc_api_setup.sh and verify LD_LIBRARY_PATH includes the TensorFlow library directory.
Linker errors during g++ compilation Ensure the -rpath and -L flags point to the correct directory containing the TensorFlow .so files.
Poor inference performance Review and tune OMP_NUM_THREADS, KMP_AFFINITY, and ZENDNNL_MATMUL_ALGO in zentf_env_setup.sh for your CPU.