Installing and deploying ROCm Compute Profiler

Installing and deploying ROCm Compute Profiler#

ROCm Compute Profiler core installation
- Provides the core application profiling capability.
- Allows the collection of performance counters, filtering by hardware block, dispatch, kernel, and more.
- Provides a CLI-based analysis mode.
- Provides a standalone web interface for importing analysis metrics.

Core installation#

The core ROCm Compute Profiler application requires the following basic software dependencies. As of ROCm 6.2, the core ROCm Compute Profiler is included with your ROCm installation.

Python >= 3.8
CMake >= 3.19
ROCm >= 5.7.1

Note

ROCm Compute Profiler will use the first version of python3 found in your system’s PATH. If the default version of Python is older than 3.8, you may need to update your system’s PATH to point to a newer version.

ROCm Compute Profiler depends on a number of Python packages documented in the top-level requirements.txt file. Install these before configuring ROCm Compute Profiler.

Tip

If looking to build ROCm Compute Profiler as a developer, consider these additional requirements.

`docs/sphinx/requirements.txt`	Python packages required to build this documentation from source.
`requirements-test.txt`	Python packages required to run ROCm Compute Profiler’s CI suite using PyTest.

The recommended procedure for ROCm Compute Profiler usage is to install into a shared file system so that multiple users can access the final installation. The following steps illustrate how to install the necessary Python dependencies using pip and ROCm Compute Profiler into a shared location controlled by the INSTALL_DIR environment variable.

Tip

To always run ROCm Compute Profiler with a particular version of Python, you can create a bash alias. For example, to run ROCm Compute Profiler with Python 3.10, you can run the following command:

alias rocprof-compute-mypython="/usr/bin/python3.10 /opt/rocm/bin/rocprof-compute"

Configuration variables#

The following installation example leverages several CMake project variables defined as follows.

CMake variable	Description
`CMAKE_INSTALL_PREFIX`	Controls the install path for ROCm Compute Profiler files.
`PYTHON_DEPS`	Specifies an optional path to resolve Python package dependencies.
`MOD_INSTALL_PATH`	Specifies an optional path for separate ROCm Compute Profiler modulefile installation.
`rocprofiler-sdk_DIR`	Specifies the path to the rocprofiler-sdk CMake package configuration directory used to build the rocprofiler-compute counter collection tool. This directory should contain `rocprofiler-sdkConfig.cmake` (for example, `<rocprofiler-sdk-install-path>/lib/cmake/rocprofiler-sdk`).
`STANDALONEBINARY_EXTRACT_DIR`	Specifies an optional temporary path to be used for extraction by the ROCm Compute Profiler standalone binary.
`STANDALONEBINARY`	Should be ON to enable the build of a standalone binary for ROCm Compute Profiler.
`TEST_FROM_INSTALL`	Should be ON to enable testing from the installation location without dependency on the source directory.

Install from the TheRock nightly releases#

For detailed instructions on installing TheRock nightly release artifacts, refer to TheRock/Release.

Install from the source#

Sparse clone the repository ROCm/rocm-systems to get the ROCm Compute Profiler source code.

git clone --no-checkout --filter=blob:none https://github.com/ROCm/rocm-systems.git
cd rocm-systems
git sparse-checkout init --cone
git sparse-checkout set projects/rocprofiler-compute
git checkout develop

Navigate to the rocprofiler-compute project root.
```
cd projects/rocprofiler-compute
```

Install Python dependencies in a virtual environment, complete the ROCm Compute Profiler configuration and install process.

# define top-level install path
export INSTALL_DIR=<your-top-level-desired-install-path>

# install python deps
python3 -m pip install -t ${INSTALL_DIR}/python-libs -r requirements.txt

# configure ROCm Compute Profiler for shared install
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=${INSTALL_DIR}/3.5.0 \
        -DPYTHON_DEPS=${INSTALL_DIR}/python-libs \
        -DMOD_INSTALL_PATH=${INSTALL_DIR}/modulefiles/rocprofiler-compute ..

# install
make -j$(nproc) install

Tip

You might need to sudo the final installation step if you don’t have write access for the chosen installation path.

Upon successful installation, your top-level installation directory should look like this.
```
$ ls $INSTALL_DIR
modulefiles  3.5.0  python-libs
```

Install from the tarball#

Download the rocprofiler-compute specific tarball for the latest release from ROCm/rocm-systems.
Untar the downloaded tarball and navigate to the rocprofiler-compute directory.
Follow the installation steps under Install from the source.

Execution using modulefiles#

The installation process includes the creation of an environment modulefile for use with Lmod. On systems that support Lmod, you can register the ROCm Compute Profiler modulefile directory and setup your environment for execution of ROCm Compute Profiler as follows.

$ module use $INSTALL_DIR/modulefiles
$ module load rocprofiler-compute
$ which rocprof-compute
/opt/apps/rocprofiler-compute/3.5.0/bin/rocprof-compute

$ rocprof-compute --version
ROC Profiler:   /opt/rocm-5.1.0/bin/rocprof

rocprofiler-compute (v3.5.0)

Tip

If you’re relying on an Lmod Python module locally, you may wish to customize the resulting ROCm Compute Profiler modulefile post-installation to include extra module dependencies.

Execution without modulefiles#

To use ROCm Compute Profiler without the companion modulefile, update your PATH settings to enable access to the command line binary. If you installed Python dependencies in a shared location, also update your PYTHONPATH configuration.

export PATH=$INSTALL_DIR/3.5.0/bin:$PATH
export PYTHONPATH=$INSTALL_DIR/python-libs

Configuring the environment for ROCprofiler-SDK#

ROCm Compute Profiler profiling process relies on ROCprofiler-SDK’s rocprofiler-sdk library. Optionally, a rocprofv3 binary can be used in substitution of rocprofiler-sdk library when ROCPROF environment variable is set to rocprofv3 or to the path of rocprofv3 binary.