vspetrov/nccl

Optimized primitives for collective multi-GPU communication

CudaNOASSERTION