Installation¶
The Kernel Tuner requires several packages to be installed. First of all, you need a working Python version, several Python packages, and optionally CUDA and/or OpenCL installations. All of this is explained in detail in this guide.
For comprehensive step-by-step instructions on setting up a development environment, see Development Environment.
Python¶
You need a Python installation. We recommend using Python 3 and installing it with Miniconda. Linux users could type the following to download and install Python 3 using Miniconda:
wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
You are of course also free to use your own Python installation, and the Kernel Tuner is developed to be fully compatible with Python 3.8 and newer.
Installing Python Packages¶
Note that when you are using a native Python installation, the pip command used to install Kernel Tuner and its dependencies requires sudo rights for system wide installation.
Sudo rights are typically not required when using Miniconda or virtual environments. You could also use e.g. the –user or –prefix option of pip to install into your home directory, this requires that your home directory is on your $PYTHONPATH environment variable (see for further details the pip documentation).
The following command will install Kernel Tuner together with the required dependencies:
pip install kernel_tuner
There are also optional dependencies, explained below.
CUDA and PyCUDA¶
Installing CUDA and PyCUDA is optional, because you may want to only use Kernel Tuner for tuning OpenCL or C kernels.
If you want to use the Kernel Tuner to tune CUDA kernels you will first need to install the CUDA toolkit (https://developer.nvidia.com/cuda-toolkit). A recent version of the CUDA toolkit (and the PyCUDA Python bindings for CUDA) are recommended (older version may work, but may not support all features of Kernel Tuner).
It’s very important that you install the CUDA toolkit before trying to install PyCuda.
You can install PyCuda manually using:
pip install pycuda
Or you could install Kernel Tuner and PyCUDA together if you haven’t done so already:
pip install kernel_tuner[cuda]
If you run into trouble with installing PyCuda, make sure you have CUDA installed first. Also make sure that the Python package Numpy is already installed, e.g. using pip install numpy.
If you retry the pip install pycuda command, you may need to use the
--no-cache-dir option to ensure the pycuda installation really starts over and not continues
from an installation that is failing.
If this fails, I recommend to see the PyCuda installation guide (https://wiki.tiker.net/PyCuda/Installation)
Other CUDA Backends¶
Kernel Tuner can also be used with CuPy (https://cupy.dev/) or Nvidia’s CUDA Python bindings (https://nvidia.github.io/cuda-python/). Please see the installation instructions of those projects for how the required Python packages.
Please refer to the documentation on backends on how to use and select these backends.
OpenCL and PyOpenCL¶
Before we can install PyOpenCL you’ll need an OpenCL compiler. There are several OpenCL compilers available depending on the OpenCL platform you want to your code to run on.
You can also look at this OpenCL Installation Guide for PyOpenCL.
As with the CUDA toolkit, recent versions of one or more of the above OpenCL SDK’s and PyOpenCL are recommended to support all features of the Kernel Tuner.
After you’ve installed your OpenCL compiler of choice you can install PyOpenCL using:
pip install pyopencl
Or you could install Kernel Tuner and PyOpenCL together if you haven’t done so already:
pip install kernel_tuner[opencl]
If this fails, please see the PyOpenCL installation guide (https://wiki.tiker.net/PyOpenCL/Installation)
HIP and PyHIP¶
Before we can install PyHIP, you’ll need to have the HIP runtime and compiler installed on your system. The HIP compiler is included as part of the ROCm software stack. Here is AMD’s installation guide:
After you’ve installed HIP, you will need to install PyHIP. Run the following command in your terminal to install:
pip install pyhip-interface
Alternatively, you can install PyHIP from the source code. First, clone the repository from GitHub:
git clone https://github.com/jatinx/PyHIP
Then, navigate to the repository directory and run the following command to install:
python setup.py install
Installing the git version¶
You can also install from the git repository. This way you also get the examples. Please note that this will install all required dependencies in the current environment. For step-by-step instructions on setting up a development environment, see Development Environment.
git clone https://github.com/benvanwerkhoven/kernel_tuner.git
cd kernel_tuner
curl -sSL https://install.python-poetry.org | python3 -
poetry install
You can install Kernel Tuner with several optional dependencies.
In this we differentiate between development and runtime dependencies.
The development dependencies are test and docs, and can be installed by appending e.g. --with test,docs.
The runtime dependencies are:
cuda: install pycuda along with kernel_tuner
opencl: install pycuda along with kernel_tuner
hip: install pyhip along with kernel_tuner
tutorial: install packages required to run the guides
These can be installed by appending e.g. -E cuda -E opencl -E hip.
If you want to go all-out, use --all-extras.
For example, use: .. code-block:: bash
poetry install –with test,docs -E cuda -E opencl
To install Kernel Tuner along with all the packages required for development.
Dependencies for the guides¶
Some addition Python packages are required to run the Jupyter notebook guides. These packages are commonly used and chances are that you already have these installed.
However, to install Kernel Tuner along with the dependencies to run the guides, you could use:
pip install kernel_tuner[tutorial,cuda]
Or if you have already installed Kernel Tuner and PyCUDA, just use:
pip install jupyter matplotlib pandas