Hi, Laszlo, Thanks for your interest in using the CUDA package.
Here are the brief answers to your questions: (1) libcuda is the shared library that supports CUDA driver, which the CUDA.jl package relies on to function. (2) most CUDA driver functions are from this library (see http://docs.nvidia.com/cuda/cuda-driver-api/#axzz2zHGOfEmd for the documentation) (3) Currently, this package requires one to write CUDA kernels in CUDA C and compiles them into ptx files. The package can load the ptx modules and launch them. Just like the example provided in the project page. This package is supposed to provide low-level support. I think some people may build higher-level (i.e. more user friendly) packages as GSoC projects. This package has only been tested under Ubuntu and Mac OS X. Since I don't have windows machines, I won't be able to test them under windows anytime soon. However, patches that can make it work under windows are definitely welcomed. Dahua On Friday, April 18, 2014 3:39:44 PM UTC-5, Laszlo Hars wrote: > > Jake: > > I also have volunteered some of my time to this community. E.g. I had to > adapt Gaston to Windows systems, and documented the effort. > > The description of a package listed in the Julia website ought to contain > notes that a package was only tested under a certain OS, and should not be > assumed to work elsewhere. I spent many hours with packages, which do not > tell this, and just don't work in my system. > > In the CUDA case: what is libcuda? Is it the CUDA runtime shared library? > What functions are expected and used from it? Adding these notes should not > take more than a few minutes of the author's time, but saves a lot of > frustration and time of a potential user. > > The lack of responses to my original question indicates that there are no > Windows users of CUDA, even though having a Windows CUDA package could have > been really useful: many of us have modern NVIDIA graphic cards, which > support CUDA, and speed up certain algorithms by orders of magnitude. Now I > may end up writing my own CUDA module, which is duplicating work. >
