NCCL_P2P_DIABLE vs. cudaDeviceEnablePeerAccess

NCCL is a collective library, based on CUDA. If I disable P2P access using cuda_runtime lib, whether the P2P access of NCCL will be influenced.

Consider this hypothesis：

1. I use cudaDeviceDisablePeerAccess in my code.
2. Then I use NCCL to do some collective operation (export NCCL_P2P_DISABLE=0).

Anyone can tell me whether these collective operations through p2p or not?