trick-17/backends

Interchangeable backends in C++, OpenMP, CUDA, OpenCL, OpenACC

backends

Interchangeable backends:

• C++
• CUDA
• OpenCL
• OpenACC

Project Health

Service	System	Compiler	Status
Travis-CI	Ubuntu 14.04 OSX	GCC 6 Clang
AppVeyor	Windows	MSVC12 MSVC14 MSVC14.1

Including this library in your code

TODO...

Library usage

Filling a vector with a single value:

auto myVec = backends::vector<double>(2000);
backends::fill(myVec, 12.0);

Typical reductions:

auto myVec = backends::vector<double>{1, 2, 3, 4};
double sum  = backends::sum(myVec);
double mean = backends::mean(myVec);
// The sum should be 12, the mean should be 2.5

backends's Issues

Implement testing for cuda and opencl

Can we somehow do this abstractly (i.e. only once for cpp cu and cl at once) or do we have to write a test for each version?

An idea might be to put a switch in the CMakeLists to cpp ´cudaoropencl, which then changes which version is included with a #define`

Do we add a single file include?

We might do it like Eigen and make it possible to e.g. #include "backends/cpp" by placing files cpp cuda and opencl (note the missing file extensions), which contain the necessary includes, into the backends folder.

Document exact version requirements (OpenMP, CUDA, ...)

• C++ ? (nvcc only supports 11 or 14 I think... probably we therefore cannot use C++17?)
• OpenMP 4.5 for GPU offloading (not yet implemented)
• CUDA 8.0 for unified memory
• OpenCL ?
• OpenACC ?

Maybe we should even write down which compilers exactly support each backend due to these requirements?

Document library usage and write a few examples

Need reduction backends

For OpenCL see here
For CUDA see here

First functions could be:

• sum (and thereby also mean)
• min, max and maybe minmax

Implement setup and teardown (currently just for OpenCL)

I see two options:

• use a function
• use a define

Do we add Eigen-specific functions?

Since we need modern compilers anyways and Eigen is quite small and header-only, I don't see a particular reason not to.
Examples would be:

• dot which could take two vector fields and output a scalar field
• cross
• mean which would give a mean vector instead of a scalar

@NOhs does the usage of Eigen (e.g. std::vector<Eigen::Vector3f> correspond to your use-case?

Note the mean function could not be applied by using the templated mean, as it requires atomics. Thus it would have to be specialized.