misc.rst

Installation

Installing from Conda Forge

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.

On Linux or OS X, the following set of instructions should work:

1. Install a version of miniconda that fits your system. Both Python 2 and Python 3 work. You can install these pieces of software in your user account and do not need root/administrator privileges.

   Note that if you already have Continuum Anaconda installed on your system, you may just use that and do not need to install Miniconda.

2. source /WHERE/YOU/INSTALLED/MINICONDA/bin/activate root

3. conda config --add channels conda-forge

4. conda install pyopencl

The analogous steps for Windows should also work.

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.cffi_cl.LogicError: clGetPlatformIDs failed: <unknown error -1001>, that means 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:

1. conda install pocl

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

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

1. conda install oclgrind

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

1. conda install intel-compute-compute

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

1. conda install beignet

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.

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.

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, the packaging of PyOpenCL for Conda Forge relies on the Khronos ICD Loader, and it is packaged so that the OpenCL drivers built into the operating system are automatically available, in addition to other ICDs installed manually.

Installing from PyPI with Linux wheels

PyOpenCL distributes manylinux1 wheels in PyPI. These wheels are compatible with GLIBC>=2.5 based distributions.

On Linux, type

1. pip install pyopencl

The wheels comes with OCL-ICD bundled and configured to use any OpenCL implementation supporting ICD interface installed in :file:`/etc/OpenCL/vendors`

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,

1. pip install pyopencl[pocl]

To install pyopencl with oclgrind, an OpenCL debugger do,

1. pip install pyopencl[oclgrind]

Installing from source

Information on how to install PyOpenCL from source is maintained collaboratively on the 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

Version 2018.2

Note

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

• 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 :class:`pyopencl.array.DefaultAllocator`.

• Deprecate :class:`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 :meth:`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_object`.
• 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 :func:`pyopencl.tools.context_dependent_memoize` to the documented functionality.
• Base :mod:`pyopencl.clrandom` on 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_object` was renamed :func:`pyopencl.enqueue_migrate_mem_object_ext`. :func:`pyopencl.enqueue_migrate_mem_object` now refers to the OpenCL 1.2 function of this name, if available.
• :func:`pyopencl.create_sub_devices` was renamed :func:`pyopencl.create_sub_devices_ext`. :func:`pyopencl.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 :func:`pyopencl.enqueue_read_buffer`, :func:`pyopencl.enqueue_write_buffer`, :func:`pyopencl.enqueue_read_buffer_rect`, :func:`pyopencl.enqueue_write_buffer_rect`, :func:`pyopencl.enqueue_read_image`, :func:`pyopencl.enqueue_write_image`, :func:`pyopencl.enqueue_map_buffer`, :func:`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 :func:`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 :func:`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:`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 :func:`pyopencl.create_image_2d`, :func:`pyopencl.create_image_3d`, :func:`pyopencl.enqueue_read_image`, and :func:`pyopencl.enqueue_write_image` are now defaulted to zero. The argument order of enqueue_{read,write}_image has changed for this reason.
• Deprecate :func:`pyopencl.create_image_2d`, :func:`pyopencl.create_image_3d` in favor of the :class:`pyopencl.Image` constructor.
• Deprecate :func:`pyopencl.create_program_with_source`, :func:`pyopencl.create_program_with_binary` in favor of the :class:`pyopencl.Program` constructor.
• Deprecate :func:`pyopencl.create_buffer`, :func:`pyopencl.create_host_buffer` in favor of the :class:`pyopencl.Buffer` constructor.
• :meth:`pyopencl.MemoryObject.get_image_info` now actually exists.
• Add :attr:`pyopencl.MemoryObject.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.