HcclCommInitAll
Applicability
Product |
Supported |
|---|---|
Atlas 350 Accelerator Card |
√ |
√ |
|
√ |
|
☓ |
|
√ |
|
√ |
For
For the
Function
In the single-server communication scenario, a process is used to create a communicator for multiple devices (one device corresponds to one thread). During the initialization of the communicator, devices[0] functions as the root rank to automatically collect cluster information.
Prototype
1 | HcclResult HcclCommInitAll(uint32_t ndev, int32_t* devices, HcclComm* comms) |
Parameters
Parameter |
Input/Output |
Description |
|---|---|---|
ndev |
Input |
Number of devices in the communicator. |
devices |
Input |
List of devices in the communicator. The values are logical IDs of the devices, which can be queried by running the npu-smi info -m command. HCCL creates communicators in the sequence specified by the devices parameter. Note that the entered device list cannot contain duplicate device IDs. |
comms |
Output |
Array of generated communicator handles. Its size is ndev * sizeof(HcclComm). For details about the definition of the HcclComm type, see HcclComm. |
Returns
HcclResult: HCCL_SUCCESS on success, or else failure.
Constraints
- This interface applies only to the single-server communication scenario.
- When multiple threads call collective operation APIs (such as HcclAllReduce), ensure that the time difference between collective operation API calls in different threads does not exceed the value of the environment variable HCCL_CONNECT_TIMEOUT to avoid link setup timeout.
- One device cannot call multiple collective operation APIs at a time.
Call Example
1 2 3 4 5 6 7 8 9 | uint32_t rankSize = 2; int32_t devices[rankSize] = {0, 1}; HcclComm comms[rankSize]; // Initialize the communicator. HcclCommInitAll(rankSize, devices, comms); // Destroy the communicator. for (uint32_t i = 0; i < rankSize; i++) { HcclCommDestroy(comms[i]); } |