misc.rst

Installation

Installing from Conda Forge

Installing PyOpenCL

By far the easiest way to install PyOpenCL is to use the packages available in Conda Forge. Conda Forge is a repository of community-maintained packages for the Conda package manager. The following instructions are aimed at Linux and macOS. The analogous steps for Windows should also work.

Install a version of miniforge that fits your system:

curl -L -O "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash ./Miniforge3-*.sh
# (answer questions, pick install location)

Then run:

source /WHERE/YOU/INSTALLED/MINIFORGE/bin/activate root
conda install pyopencl

You can install these pieces of software in your user account and do not need root/administrator privileges.

Enabling access to CPUs and GPUs via (Py)OpenCL

Note that PyOpenCL is no fun (i.e. cannot run code) without an OpenCL device driver (a so-called "ICD", for "installable client driver") that provides access to hardware through OpenCL. If you get an error message like pyopencl._cl.LogicError: clGetPlatformIDs failed: PLATFORM_NOT_FOUND_KHR, that means PyOpenCL installed successfully, but you have no OpenCL drivers installed.

Note that drivers (ICDs) are separate pieces of software from PyOpenCL. They might be provided by your hardware vendor (e.g. for Nvidia or AMD GPUs). If you have such hardware, see below for instructions on how to make those work with PyOpenCL from Conda Forge.

It is important to note that OpenCL is not restricted to GPUs. In fact, no special hardware is required to use OpenCL for computation--your existing CPU is enough. On Linux or macOS, type:

conda install pocl

to install a CPU-based OpenCL driver. On macOS, pocl can offer a marked robustness (and, sometimes, performance) improvement over the OpenCL drivers built into the operating system.

On Linux and Windows, you can use Intel's CPU OpenCL runtime:

conda install intel-opencl-rt

On Linux Intel Broadwell or newer processors with an Intel graphics card, you can use NEO:

conda install intel-compute-runtime

On Linux Intel Sandybridge or newer processors with an Intel graphics card, you can use Beignet:

conda install beignet

On Linux, Windows and macOS, you can use Oclgrind to detect memory access errors:

conda install oclgrind

You are now ready to run code based on PyOpenCL, such as the code examples.

Using vendor-supplied OpenCL drivers (mainly on Linux)

The instructions above help you get a basic OpenCL environment going that will work independently of whether you have specialized hardware (such as GPUs or FPGAs) available. If you do have such hardware, read on for how to make it work.

On Linux, PyOpenCL finds which drivers are installed by looking for files with the extension .icd in a directory. PyOpenCL as installed from Conda will look for these files in :file:`/WHERE/YOU/INSTALLED/MINICONDA/etc/OpenCL/vendors`. They are just simple text files containing either just the file names or the fully qualified path names of the shared library providing the OpenCL driver.

Note

If you ran the commands above in a Conda environment (i.e. if the environment indicator on your command line prompt says anything other than (root)), then you may need to use a path like the following instead:

:file:`/WHERE/YOU/INSTALLED/MINICONDA/envs/ENVIRONMENTNAME/etc/OpenCL/vendors`

Note that you should replace ENVIRONMENTNAME with the name of your environment, shown between parentheses on your command line prompt. This path (for the currently-active conda environment) can be obtained from the environment variable CONDA_PREFIX, i.e.,

:file:`$CONDA_PREFIX/etc/OpenCL/vendors` (once the Conda environment is activated).

On Linux, if you have other OpenCL drivers installed (such as for your GPU), those will be in :file:`/etc/OpenCL/vendors`. You can make them work with PyOpenCL from Conda Forge by using the command:

conda install ocl-icd-system

will make sure these system-wide ICDs are also visible in your conda environment. As an alternative, one may manually copy ICD files from :file:`/etc/OpenCL/vendors` into, e.g., :file:`$CONDA_PREFIX/etc/OpenCL/vendors`.

If you are looking for more information, see ocl-icd and its documentation. Ocl-icd is the "ICD loader" used by PyOpenCL when installed from Conda Forge on Linux. It represents the code behind :file:`libOpenCL.so`.

On macOS, using the command:

conda install ocl_icd_wrapper_apple

will make sure that the Apple provided CPU and GPU implementations are available.

On Windows, the packaging of PyOpenCL for Conda Forge relies on the Khronos ICD Loader, and it is packaged so that the OpenCL drivers that are registered in the OS using registry keys are automatically available.

Installing from PyPI wheels

PyOpenCL distributes wheels for most popular OSs and Python versions. To check available versions please visit PyPI page for PyOpenCL.

On Linux, the wheels come with OCL-ICD bundled and configured to use any OpenCL implementation supporting the ICD interface and listed in :file:`/etc/OpenCL/vendors`. Wheels for Windows and MacOS are built using the ICD Loader from the Khronos Group.

To install, type:

pip install pyopencl

You can also install the following CPU based OpenCL implementation using pip shipped as binary wheels. Note that pyopencl has to be installed using a wheel for pyopencl to recognize these wheels.

To install pyopencl with pocl, a CPU based implementation do:

pip install pyopencl[pocl]

To install pyopencl with oclgrind, an OpenCL debugger do:

pip install pyopencl[oclgrind]

Note

Avoid mixing components installed from Conda Forge and PyPI. For example, installing PyOpenCL from pip followed by OCL-ICD from Conda Forge can redirect the ICD loader, removing access to system-wide ICDs.

Installing via Christoph Gohlke's binary wheels (Windows)

Christoph Gohlke distributes binary wheels for PyOpenCL on Windows.

Installing from source

Information on how to install PyOpenCL from source is still available on the Former PyOpenCL Wiki, but that should mostly not be necessary unless you have very specific needs or would like to modify PyOpenCL yourself.

Tips

Syntax highlighting

You can obtain Vim syntax highlighting for OpenCL C inlined in Python by checking this file.

Note that the triple-quoted strings containing the source must start with """//CL// ...""".

IPython integration

PyOpenCL comes with IPython integration, which lets you seamlessly integrate PyOpenCL kernels into your IPython notebooks. Simply load the PyOpenCL IPython extension using:

%load_ext pyopencl.ipython_ext

and then use the %%cl_kernel 'cell-magic' command. See this notebook (which ships with PyOpenCL) for a demonstration.

You can pass build options to be used for building the program executable by using the -o flag on the first line of the cell (next to the %%cl_kernel directive). For example:

%%cl_kernel -o "-cl-fast-relaxed-math"

There are also line magics: cl_load_edit_kernel which will load a file into the next cell (adding cl_kernel to the first line) and cl_kernel_from_file which will compile kernels from a file (as if you copy-and-pasted the contents of the file to a cell with cl_kernel). Both of these magics take options -f to specify the file and optionally -o for build options.

Guidelines

API Stability

I consider PyOpenCL's API "stable". That doesn't mean it can't change. But if it does, your code will generally continue to run. It may however start spewing warnings about things you need to change to stay compatible with future versions.

Deprecation warnings will be around for a whole year, as identified by the first number in the release name. (the "2014" in "2014.1") I.e. a function that was deprecated in 2014.n will generally be removed in 2015.n (or perhaps later). Further, the stability promise applies for any code that's part of a released version. It doesn't apply to undocumented bits of the API, and it doesn't apply to unreleased code downloaded from git.

Relation with OpenCL's C Bindings

We've tried to follow these guidelines when binding the OpenCL's C interface to Python:

• Remove the cl_, CL_ and cl prefix from data types, macros and function names.
• Follow PEP 8, i.e.
  • Make function names lowercase.
  • If a data type or function name is composed of more than one word, separate the words with a single underscore.
• get_info functions become attributes.
• Object creation is done by constructors, to the extent possible. (i.e. minimize use of "factory functions")
• If an operation involves two or more "complex" objects (like e.g. a kernel enqueue involves a kernel and a queue), refuse the temptation to guess which one should get a method for the operation. Instead, simply leave that command to be a function.

Interoperability with other OpenCL software

Just about every object in :mod:`pyopencl` supports the following interface (here shown as an example for :class:`pyopencl.MemoryObject`, from which :class:`pyopencl.Buffer` and :class:`pyopencl.Image` inherit):

• :meth:`pyopencl.MemoryObject.from_int_ptr`
• :attr:`pyopencl.MemoryObject.int_ptr`

This allows retrieving the C-level pointer to an OpenCL object as a Python integer, which may then be passed to other C libraries whose interfaces expose OpenCL objects. It also allows turning C-level OpenCL objects obtained from other software to be turned into the corresponding :mod:`pyopencl` objects.

User-visible Changes

Unreleased

Note

This version is currently under development. You can get snapshots from PyOpenCL's git repository.

Version 2022.2

• Added :ref:`opaque-style SVM <opaque-svm>` and :class:`pyopencl.SVMPointer`.
• Added :class:`pyopencl.tools.SVMPool`.
• Added automatic queue-synchronized deallocation of SVM.

Version 2020.2

• Drop Python 2 support.
• Add allow_empty_ndrange to kernel enqueue.
• Bug fixes.

Version 2018.2

• Use pybind11.
• Many bug fixes.
• Support arrays with offsets in scan kernels.

Version 2018.1

• Introduce eliminate_empty_output_lists argument of :class:`pyopencl.algorithm.ListOfListsBuilder`.
• Many bug fixes.

Version 2017.2

• Many bug fixes.

Version 2017.1

• Introduce :mod:`pyopencl.cltypes`

Version 2016.2

• Deprecate RANLUXCL. It will be removed in the 2018.x series of PyOpenCL.
• Introduce Random123 random number generators. See :mod:`pyopencl.clrandom` for more information.
• Add support for range and slice kwargs and data-less reductions to :class:`pyopencl.reduction.ReductionKernel`.
• Add support for SPIR-V. (See :class:`pyopencl.Program`.)
• Add support for :ref:`svm`.
• :class:`pyopencl.MemoryMap` is usable as a context manager.

Version 2016.1

• The from_int_ptr methods now take a retain parameter for more convenient ownership management.
• Kernel build options (if passed as a list) are now properly quoted. (This is a potentially compatibility-breaking change.)
• Many bug fixes. (GL interop, Windows, event callbacks and more)

Version 2015.2.4

• Fix building on Windows, using mingwpy and VS 2015.

Version 2015.2.3

• Fix one more Ubuntu 14.x build issue.

Version 2015.2.2

• Fix compatibility with CL 1.1
• Fix compatibility with Ubuntu 14.x.
• Various bug fixes

Version 2015.2.1

• Fix global_offset kernel launch parameter

Version 2015.2

• [INCOMPATIBLE] Changed PyOpenCL's complex numbers from float2 and double2 OpenCL vector types to custom struct. This was changed because it very easily introduced bugs where

  • complex*complex
  • real+complex

  look like they may do the right thing, but silently do the wrong thing.

• Rewrite of the wrapper layer to be based on CFFI

• Pypy compatibility

• Faster kernel invocation through Python launcher code generation

• POCL compatibility

Version 2015.1

• Support for new-style buffer protocol
• Numerous fixes

Version 2014.1

• :ref:`ipython-integration`
• Bug fixes

Version 2013.2

• Add :meth:`pyopencl.array.Array.map_to_host`.
• Support strides on :func:`pyopencl.enqueue_map_buffer` and :func:`pyopencl.enqueue_map_image`.
• :class:`pyopencl.ImageFormat` was made comparable and hashable.
• :mod:`pyopencl.reduction` supports slicing (contributed by Alex Nitz)
• Added :ref:`interoperability`
• Bug fixes

Version 2013.1

• Vastly improved :ref:`custom-scan`.

• Add :func:`pyopencl.tools.match_dtype_to_c_struct`, for better integration of the CL and :mod:`numpy` type systems.

• More/improved Bessel functions. See the source.

• Add :envvar:`PYOPENCL_NO_CACHE` environment variable to aid debugging. (e.g. with AMD's CPU implementation, see their programming guide)

• Deprecated :func:`pyopencl.tools.register_dtype` in favor of :func:`pyopencl.tools.get_or_register_dtype`.

• Clean up the :class:`pyopencl.array.Array` constructor interface.

• Deprecate pyopencl.array.DefaultAllocator.

• Deprecate pyopencl.tools.CLAllocator

• Introduce :class:`pyopencl.tools.DeferredAllocator`, :class:`pyopencl.tools.ImmediateAllocator`.

• Allow arrays whose beginning does not coincide with the beginning of their :attr:`pyopencl.array.Array.data` :class:`pyopencl.Buffer`. See :attr:`pyopencl.array.Array.base_data` and :attr:`pyopencl.array.Array.offset`. Note that not all functions in PyOpenCL support such arrays just yet. These will fail with :exc:`pyopencl.array.ArrayHasOffsetError`.

• Add :meth:`pyopencl.array.Array.__getitem__` and :meth:`pyopencl.array.Array.__setitem__`, supporting generic slicing.

  It is possible to create non-contiguous arrays using this functionality. Most operations (elementwise etc.) will not work on such arrays.

  Note also that some operations (specifically, reductions and scans) on sliced arrays that start past the beginning of the original array will fail for now. This will be fixed in a future release.

• :class:`pyopencl.CommandQueue` may be used as a context manager (in a with statement)

• Add :func:`pyopencl.clmath.atan2`, :func:`pyopencl.clmath.atan2pi`.

• Add :func:`pyopencl.array.concatenate`.

• Add :meth:`pyopencl.Kernel.capture_call`.

Note

The addition of :meth:`pyopencl.array.Array.__getitem__` has an unintended consequence due to numpy bug 3375. For instance, this expression:

numpy.float32(5) * some_pyopencl_array

may take a very long time to execute. This is because :mod:`numpy` first builds an object array of (compute-device) scalars (!) before it decides that that's probably not such a bright idea and finally calls pyopencl.array.Array.__rmul__.

Note that only left arithmetic operations of :class:`pyopencl.array.Array` by :mod:`numpy` scalars are affected. Python's number types (:class:`float` etc.) are unaffected, as are right multiplications.

If a program that used to run fast suddenly runs extremely slowly, it is likely that this bug is to blame.

Here's what you can do:

• Use Python scalars instead of :mod:`numpy` scalars.
• Switch to right multiplications if possible.
• Use a patched :mod:`numpy`. See the bug report linked above for a pull request with a fix.
• Switch to a fixed version of :mod:`numpy` when available.

Version 2012.1

• Support for complex numbers.
• Support for Bessel functions. (experimental)
• Numerous fixes.

Version 2011.2

• Add :func:`pyopencl.enqueue_migrate_mem_objects`.
• Add :func:`pyopencl.image_from_array`.
• IMPORTANT BUGFIX: Kernel caching was broken for all the 2011.1.x releases, with severe consequences on the execution time of :class:`pyopencl.array.Array` operations. Henrik Andresen at a PyOpenCL workshop at DTU first noticed the strange timings.
• All comparable PyOpenCL objects are now also hashable.
• Add pyopencl.tools.context_dependent_memoize to the documented functionality.
• Base :mod:`pyopencl.clrandom` on RANLUXCL (https://bitbucket.org/ivarun/ranluxcl>), add functionality.
• Add :class:`pyopencl.NannyEvent` objects.
• Add :mod:`pyopencl.characterize`.
• Ensure compatibility with OS X Lion.
• Add :func:`pyopencl.tools.register_dtype` to enable scan/reduction on struct types.
• :func:`pyopencl.enqueue_migrate_mem_objects` was renamed pyopencl.enqueue_migrate_mem_objects_ext. :func:`pyopencl.enqueue_migrate_mem_objects` now refers to the OpenCL 1.2 function of this name, if available.
• :meth:`pyopencl.Device.create_sub_devices` was renamed pyopencl.Device.create_sub_devices_ext. :meth:`pyopencl.Device.create_sub_devices` now refers to the OpenCL 1.2 function of this name, if available.
• Alpha support for OpenCL 1.2.

Version 2011.1.2

• More bug fixes.

Version 2011.1.1

• Fixes for Python 3 compatibility. (with work by Christoph Gohlke)

Version 2011.1

• All is_blocking parameters now default to True to avoid crashy-by-default behavior. (suggested by Jan Meinke) In particular, this change affects pyopencl.enqueue_read_buffer, pyopencl.enqueue_write_buffer, pyopencl.enqueue_read_buffer_rect, pyopencl.enqueue_write_buffer_rect, pyopencl.enqueue_read_image, pyopencl.enqueue_write_image, pyopencl.enqueue_map_buffer, pyopencl.enqueue_map_image.
• Add :mod:`pyopencl.reduction`.
• Add :ref:`reductions`.
• Add :mod:`pyopencl.scan`.
• Add :meth:`pyopencl.MemoryObject.get_host_array`.
• Deprecate context arguments of :func:`pyopencl.array.to_device`, :func:`pyopencl.array.zeros`, :func:`pyopencl.array.arange`.
• Make construction of :class:`pyopencl.array.Array` more flexible (cqa argument.)
• Add :ref:`memory-pools`.
• Add vector types, see :class:`pyopencl.array.vec`.
• Add :attr:`pyopencl.array.Array.strides`, :attr:`pyopencl.array.Array.flags`. Allow the creation of arrays in C and Fortran order.
• Add :func:`pyopencl.enqueue_copy`. Deprecate all other transfer functions.
• Add support for numerous extensions, among them device fission.
• Add a compiler cache.
• Add the 'g_times_l' keyword arg to kernel execution.

Version 0.92

• Add support for OpenCL 1.1.
• Add support for the cl_khr_gl_sharing extension, leading to working GL interoperability.
• Add :meth:`pyopencl.Kernel.set_args`.
• The call signature of :meth:`pyopencl.Kernel.__call__` changed to emphasize the importance of local_size.
• Add :meth:`pyopencl.Kernel.set_scalar_arg_dtypes`.
• Add support for the cl_nv_device_attribute_query extension.
• Add :meth:`pyopencl.array.Array` and related functionality.
• Make build not depend on Boost C++.

Version 0.91.5

• Add :attr:`pyopencl.ImageFormat.channel_count`, :attr:`pyopencl.ImageFormat.dtype_size`, :attr:`pyopencl.ImageFormat.itemsize`.
• Add missing pyopencl.enqueue_copy_buffer.
• Add :func:`pyopencl.create_some_context`.
• Add :func:`pyopencl.enqueue_barrier`, which was previously missing.

Version 0.91.4

A bugfix release. No user-visible changes.

Version 0.91.3

• All parameters named host_buffer were renamed hostbuf for consistency with the :class:`pyopencl.Buffer` constructor introduced in 0.91. Compatibility code is in place.
• The :class:`pyopencl.Image` constructor does not need a shape parameter if the given hostbuf has hostbuf.shape.
• The :class:`pyopencl.Context` constructor can now be called without parameters.

Version 0.91.2

• :meth:`pyopencl.Program.build` now captures build logs and adds them to the exception text.
• Deprecate pyopencl.create_context_from_type in favor of second form of :class:`pyopencl.Context` constructor
• Introduce :class:`pyopencl.LocalMemory`.
• Document kernel invocation and :meth:`pyopencl.Kernel.set_arg`.

Version 0.91.1

• Fixed a number of bugs, notably involving :class:`pyopencl.Sampler`.
• :class:`pyopencl.Device`, :class:`pyopencl.Platform`, :class:`pyopencl.Context` now have nicer string representations.
• Add :attr:`pyopencl.Image.shape`. (suggested by David Garcia)

Version 0.91

• Add :ref:`gl-interop`.
• Add a test suite.
• Fix numerous get_info bugs. (reports by David Garcia and the test suite)
• Add :meth:`pyopencl.ImageFormat.__repr__`.
• Add :meth:`pyopencl.addressing_mode.to_string` and colleagues.
• The pitch arguments to pyopencl.create_image_2d, pyopencl.create_image_3d, pyopencl.enqueue_read_image, and pyopencl.enqueue_write_image are now defaulted to zero. The argument order of enqueue_{read,write}_image has changed for this reason.
• Deprecate pyopencl.create_image_2d, pyopencl.create_image_3d in favor of the :class:`pyopencl.Image` constructor.
• Deprecate pyopencl.create_program_with_source, pyopencl.create_program_with_binary in favor of the :class:`pyopencl.Program` constructor.
• Deprecate pyopencl.create_buffer, pyopencl.create_host_buffer in favor of the :class:`pyopencl.Buffer` constructor.
• :meth:`pyopencl.Image.get_image_info` now actually exists.
• Add :attr:`pyopencl.Image.info`.
• Fix API tracing.
• Add constructor arguments to :class:`pyopencl.ImageFormat`. (suggested by David Garcia)

Version 0.90.4

• Add build fixes for Windows and OS X.

Version 0.90.3

• Fix a GNU-ism in the C++ code of the wrapper.

Version 0.90.2

• Fix :meth:`pyopencl.Platform.get_info`.
• Fix passing properties to :class:`pyopencl.CommandQueue`. Also fix related documentation.

Version 0.90.1

• Fix building on the Mac.

Version 0.90

• Initial release.

License

Frequently Asked Questions

The FAQ is maintained collaboratively on the Wiki FAQ page.

Citing PyOpenCL

We are not asking you to gratuitously cite PyOpenCL in work that is otherwise unrelated to software. That said, if you do discuss some of the development aspects of your code and would like to highlight a few of the ideas behind PyOpenCL, feel free to cite this article:

Andreas Klöckner, Nicolas Pinto, Yunsup Lee, Bryan Catanzaro, Paul Ivanov, Ahmed Fasih, PyCUDA and PyOpenCL: A scripting-based approach to GPU run-time code generation, Parallel Computing, Volume 38, Issue 3, March 2012, Pages 157-174.

Here's a Bibtex entry for your convenience:

@article{kloeckner_pycuda_2012,
   author = {{Kl{\"o}ckner}, Andreas
        and {Pinto}, Nicolas
        and {Lee}, Yunsup
        and {Catanzaro}, B.
        and {Ivanov}, Paul
        and {Fasih}, Ahmed },
   title = "{PyCUDA and PyOpenCL: A Scripting-Based Approach to GPU Run-Time Code Generation}",
   journal = "Parallel Computing",
   volume = "38",
   number = "3",
   pages = "157--174",
   year = "2012",
   issn = "0167-8191",
   doi = "10.1016/j.parco.2011.09.001",
}

Acknowledgments

Contributors

Too many to list. Please see the commit log for detailed acknowledgments.

Funding

Work on pytential was supported in part by

• the US National Science Foundation under grant numbers DMS-1418961, DMS-1654756, SHF-1911019, and OAC-1931577, and
• the Department of Energy, National Nuclear Security Administration, under Award Number DE-NA0003963.

AK also gratefully acknowledges a hardware gift from Nvidia Corporation.

The views and opinions expressed herein do not necessarily reflect those of the funding agencies.

Documentation Cross-References

Numpy

See :class:`numpy.generic`.

See :class:`numpy.generic`.

See :class:`numpy.generic`.

OpenCL Specification

Internal Types

See :class:`pyopencl.Platform`.

See :class:`pyopencl.Device`.

See :class:`pyopencl.CommandQueue`.

See :class:`pyopencl.Context`.

See :class:`pyopencl.Event`.

See :class:`pyopencl.SVMAllocation`.

See :class:`pyopencl.MemoryMap`.

See :class:`pyopencl.Sampler`.

See :class:`pyopencl.Program`.

See :class:`pyopencl.Program`.

See :class:`pyopencl.Kernel`.