.. index::
      pair: Tutorial; Index

.. _GPU_compiling:

================================
CUDA and OpenACC Compilers
================================

This page contains information about compiling GPU-based codes with NVidia's CUDA compiler and PGI's OpenACC compiler directives.  For information on how to run GPU Computing jobs in Midway, see :ref:`GPU_computing`

.. contents:: :local:

Compiling CUDA GPU code on Midway
---------------------------------

To view available CUDA versions on Midway, use the command:

.. code:: bash
  
  module avail cuda

At time of writing, the available CUDA versions are::
  
  cuda/4.2(default) 
  cuda/5.0          
  cuda/5.5 


A very basic CUDA example code is provided below :download:`cudamemset.cu`:

.. literalinclude:: cudamemset.cu
  :language: bash

CUDA code must be compiled with Nvidia's nvcc compiler which is part of the cuda software module.  To build a CUDA executable, first load the desired CUDA module and compile with:

.. code:: bash

  nvcc source_code.cu


Compiling OpenACC GPU code on Midway
------------------------------------

OpenACC is supported on Midway through the PGI 2013 compiler suite.  To load the OpenACC compiler, use the command:

.. code:: bash

  module load pgi/2013


A very basic OpenACC example code is provided below :download:`stencil.c`:

.. literalinclude:: stencil.c
  :language: bash

OpenACC code targeted at an Nvidia GPU must be compiled with the PGI compiler using at least the following options:

.. code:: bash

  pgcc source_code.c -ta=nvidia -acc