Windows + Visual Studio port

[tbennun]

Here is my attempt to port nccl to Windows.
In the process, some concepts (e.g., atomic variables) were converted to C++11.

Unsupported features in Windows:

• CUDA IPC memory (and thus MPI applications)
• NVML "set processor affinity" functionality

Untested features:

• shmOpen/Unlink/Unmap (because GPU IPC memory pointers aren't supported anyway)

[happynear]

Hi @tbennun ,

Could you update your repository to the latest version of the official repo?

[tbennun]

Sure, I'll try to do it over the weekend.

[tbennun]

@happynear Done, tested on Windows. Please make sure everything still works prior to the merge, I didn't get a chance to test this on a Linux machine.

[happynear]

Thanks:)

[pipipopo]

Thanks for this contribution! I met a problem when compiling using VS2013 and CUDA 8.0, reads below:
/nccl-windows-master/src/all_gather.cu(180): error C2440: “type conversion”: cannot convert from “void (__cdecl *)(const KernelArgs)” to “void *”
May I know what's wrong?

[tbennun]

@pipipopo I'm using VS2015 and don't observe this error.

The error seems to stem from src/common_coll.h, line 111. Try to play with the casting there (reinterpret_cast?) until it compiles.

[ZoroDerVonCodier]

Great work! works great using VS2015. The only suggestion that I would have is to add a VERSION resource to the project so as to better track versioning info with the master project.