tree 896756920c9921d47b421a4565fb61ce14cbefe7
parent d93328711457c38a77506790764e9e5ed2c8dc47
author Pieter Noordhuis <pietern@fb.com> 1490911468 -0700
committer Facebook Github Bot <facebook-github-bot@users.noreply.github.com> 1490911563 -0700

Move CUDA collectives to cuda_collectives.h

Summary:
The CUDA algorithms all had their own version of local reduction and
broadcast. This commit consolidates them and allows all CUDA
algorithms to work with CudaDevicePointer instances.

Reviewed By: andrewwdye

Differential Revision: D4797968

fbshipit-source-id: cccef39fce01905a2cd757ccbcffd29803411409