.. blogpost::
    :title: GPU and pycuda or pyopencl on Windows
    :keywords: Windows, pycuda
    :date: 2017-01-14
    :categories: module

    `pycuda <https://developer.nvidia.com/pycuda>`_
    is a python module implemented to use GPU.
    On Windows, the best way to install it is to get it
    from `Unofficial Windows Binaries for Python Extension Packages <http://www.lfd.uci.edu/~gohlke/pythonlibs/>`_.
    The module requires `Visual Studio Community Edition <https://www.visualstudio.com/vs/community/>`_ and
    `CUDA Toolkit <https://developer.nvidia.com/accelerated-computing-toolkit>`_.
    Once it is done, you can try the following program to test
    your installation:

    ::

        from pycuda.compiler import SourceModule
        mod = SourceModule("""
        __global__ void multiply_them(float *dest, float *a, float *b)
        {
          const int i = threadIdx.x;
          dest[i] = a[i] * b[i];
        }
        """)

        multiply_them = mod.get_function("multiply_them")

        a = numpy.random.randn(400).astype(numpy.float32)
        b = numpy.random.randn(400).astype(numpy.float32)

        dest = numpy.zeros_like(a)
        multiply_them(
                drv.Out(dest), drv.In(a), drv.In(b),
                block=(400,1,1), grid=(1,1))

        print(dest-a*b)

    If it fails due to the following error:

    ::

        ImportError: DLL load failed: Le module spécifié est introuvable.

    It probably means the DLL ``nvcuda.dll`` is missing.
    You can use `dependency walker <http://www.dependencywalker.com/>`_
    to look at the DLL dependencies for a specific DLL (files *.dll
    or *.pyd if they are compiled with Python).
    In that case, this means the drivers for the graphics card is missing.
    You first need to check your card. If you realise your card is not
    NVidia, you can stop: pycuda only works with a NVidia card.
    Mine is a *Intel Graphics 4000* and I downloaded the drivers from this location:
    `Intel® Graphics Driver for Windows® 10 and Windows 7*/8.1* [15.33] <https://downloadcenter.intel.com/download/25977/Intel-Graphics-Driver-for-Windows-10-and-Windows-7-8-1-15-33-?product=81499>`_.
    In that case, you will need to switch to
    `pyopencl <https://documen.tician.de/pyopencl/>`_.
    You can still get it from
    `Unofficial Windows Binaries for Python Extension Packages <http://www.lfd.uci.edu/~gohlke/pythonlibs/>`_.

    .. https://software.intel.com/en-us/articles/opencl-drivers#win64.

    Finally, you should get the same OpenCL library used to compile
    *pyopencl* which `Intel OpenCL <https://software.intel.com/en-us/intel-opencl>`_.
    You might also need
    `Windows 10 SDK <Khttps://developer.microsoft.com/en-us/windows/downloads/windows-10-sdk>`_.
    Finally, you need to choose the backend for *pyopencl*:

    ::

        import pyopencl as cl
        ctx = cl.create_some_context()

    Follow the questions and choose the graphics card as the backend.
    Setup the following environment variable to skip the questions next time:

    ::

        import os
        os.environ["PYOPENCL_CTX"] = '0:1'

    **theano**

    If you want to use `theano <http://deeplearning.net/software/theano/>`_ on GPU,
    you need `TDM-GCC 64bit <http://tdm-gcc.tdragon.net/>`_ but
    `theano <http://deeplearning.net/software/theano/>`_ only
    supports NVidia cards. The documentation
    tells it will support *opencl* soon:
    `GpuArray Backend <http://deeplearning.net/software/theano/tutorial/using_gpu.html#gpuarray>`_.