CUDA backend for ODL

This is a default CUDA backend for ODL. It contains a default implementation of a CUDA based Rn space with associated methods.

By installing this package, CUDA spaces become available to rn, uniform_discr and related spaces in ODL through the impl='cuda' option.

The project uses CMake to enable builds. Using cmake-gui is recommended

Cmake webpage

The binaries are typically built in a build folder. Starting at the repository's top level directory, do

mkdir build
cd build
cmake-gui ../

And set the required variables.

To build and install the package to your python installation, run (as root)

make pyinstall

To verify your installation, run (in the odlcuda root directory)

py.test

This requires the pytest package.

To build on windows, open the CMake gui, run configure-generate and set the required variables. Then open the project with Visual Studio and build pyinstall.

The code is written in C++11/14.

The code is intended to be usable with all major compilers. Current status (2015-06-22) is

The code is formatted using CLang format as provided by the LLVM project. The particular style used is defined in the formatting file.

Current external dependencies are

#####Python The building block of ODL, odlcuda needs access to both python and numpy header files and compiled files to link against.

#####Boost General library with C++ code. This project specifically uses Boost.Python to handle the python bindings.

Boost webpage Prebuilt boost (windows), this version uses python 2.7

#####CUDA Used for GPU accelerated versions of algorithms. The code uses C++11 features in device code, so CUDA 7.0 is required. CUDA 6.5 may work on some platforms (notably windows). It should as of now compile with compute capbability >2.0, in the future certain features require higher compute capability may be added.

CUDA

There are a few common errors encountered, this is the solution to some of these

• When compiling if, you get a error like

    NumpyConfig.h not found
  

  then it is likely that the variable PYTHON_NUMPY_INCLUDE_DIR is not correctly set.

• Compiling

    error C1083: Cannot open include file: 'Eigen/Core': No such file or directory.
  

  you have tried to build the default target, instead build target "pyinstall".

• Compiling

    [ 20%] Building NVCC (Device) object odlcuda/CMakeFiles/cuda.dir//./cuda_generated_cuda.cu.o /usr/include/c++/4.9.2/bits/alloc_traits.h(248): error: expected a ">"
  

  It may be that you are trying to compile with CUDA 6.5 and GCC 4.9, this combination is not supported by CUDA.

• Compiling

    error LNK1112: module machine type 'x64' conflicts with target machine type 'X86'
  

  You have a 64-bit library on your path (Boost for instance) while trying to build 32-bit odlcuda. Either change the lib, or configure to build 64-bit. On Windows, if you are using Visual Studio to compile use Configuration Manager to set platform to x64, if you are compiling on command line via CMake ensure that the Script Generator is for instance "Visual Studio 12 2013 Win64" (note the Win64 at the end).

• If you get a error like

    Error	5	error LNK2019: unresolved external symbol "__declspec(dllimport) struct _object * __cdecl boost::python::detail::init_module(struct PyModuleDef &,void (__cdecl*)(void))" (__imp_?init_module@detail@python@boost@@YAPEAU_object@@AEAUPyModuleDef@@P6AXXZ@Z) referenced in function PyInit_utils	C:\Programming\Projects\odlcuda_bin\odlcuda\utils.obj	utils
  

  then it is likely that you are trying to build against unmatched python header files and boost python version

• If, when running the tests in python, you get an error like

    RuntimeError: function_attributes(): after cudaFuncGetAttributes: invalid device function
  

  It may be that the compute version used is not supported by your setup, try changing the cmake variable CUDA_COMPUTE.

• If, when running the test in python, you encounter an error like

    ImportError: No module named odlcuda
  

  It may be that you have not installed the package, run (as root) make pyinstall or equivalent.

GPL Version 3. See LICENSE file.

If you would like to get the code under a different license, please contact the developers.

• Jonas Adler (jonas-<ätt>-kth--se)
• Holger Kohr (kohr-<ätt>-kth--se)