pyopencl-extension 0.2.9

This package extends PyOpenCl by providing an object-oriented kernel programming framework and debugging capabilities.

This package extends PyOpenCl by providing an object-oriented kernel programming framework and debugging capabilities.

What makes pyopencl-extension special?

• Build on top of PyOpenCl which can increase performance significantly.

• Usage of this framework forces consistent code when programming for GPU.

• Allows debugging of OpenCl-Programs through kernel emulation in Python using a visual debugger (tested with Pycharm).

• OpenCl emulation allows to find out-of-bounds array indexing more easily.

• Integrated profiling features give quick overview over performance bottlenecks.

The project is in an early development stage and actively maintained. For any feature requests/feedback/etc. you can get in touch via Github or by E-Mail ([e-mail]).

Installation

Install this library with pip install pyopencl-extension.

When automatic installation of PyOpenCl fails (happens when using Windows):

1. Download appropriate .whl binary from PyOpenCl binaries for Windows

2. Make sure to have proper OpenCl driver for your graphics card installed.

3. Run pip install "pyopencl-X-cpX-cpX-X.whl" in terminal, where X must be replaced with corresponding version.

Usage

One very simple example is given below.

from pyopencl_extension import Kernel, Global, Scalar, Types, zeros
import numpy as np

ary = zeros((10,), Types.short)

knl = Kernel('some_operation',
             {'ary': Global(ary),  # notice that ary is set as default argument
              'number': Scalar(Types.short)},
             """
                ary[get_global_id(0)] = get_global_id(0) + number;
             """,
             global_size=ary.shape).compile(emulate=True)
knl(number=3)
assert np.allclose(ary.get(), np.arange(10) + 3)

By setting the argument ‘emulate=True’ the kernel will be compiled in emulation mode. This mode creates a file ‘some_operation.py’, which can be inspected using a visual debugger:

More advanced examples can by found in the tests hosted on Github.