Why group calls (ncclGroupStart() and ncclGroupEnd()) are invoked in ncclSend() and ncclRecv()

[ZhiyiHu1999]

From nccl documents, we can know group calls have three functions:

1. Managing multiple GPUs from one thread (to avoid deadlocks);
2. Aggregating communication operations to improve performance;
3. Merging multiple send/receive point-to-point operations

I wonder which function the group calls invoked in ncclSend() and ncclRecv() corresponds to. Thanks a lot!