HcclThreadAcquire

Applicability

For Atlas A2 training product/Atlas A2 inference product, only the Atlas 800T A2 training server, Atlas 900 A2 PoD cluster basic unit, and Atlas 200T A2 Box16 heterogeneous subrack are supported.

Function

Obtains the communication thread based on the communicator.

Prototype

1

HcclResult HcclThreadAcquire(HcclComm comm, CommEngine engine, uint32_t threadNum, uint32_t notifyNumPerThread, ThreadHandle *threads)

Parameters

1

typedef void *HcclComm;

Returns

HcclResult: HCCL_SUCCESS on success, or else failure.

Restrictions

The returned communication threads and synchronization resources are managed by the library and cannot be released by the caller.

Example

The following is an example of creating thread resources:

1
2
3
4
5

HcclComm comm;
CommEngine engine = COMM_ENGINE_AICPU_TS;
ThreadHandle threads[5];
// Allocate five threads, each with two Notify resources.
HcclThreadAcquire(comm, engine, 5, 2, threads);

The following is an example of synchronizing host thread resources:

// Allocate one host stream.
aclrtStream stream;
aclrtCreateStream(&stream);
// Create a thread of the CPU_TS type based on the allocated stream.
ThreadHandle cpuThread;
HcclThreadAcquireWithStream(comm, COMM_ENGINE_CPU_TS, streams, 1, &cpuThread);
// Task orchestration
// ...
// Stream synchronization
aclrtSynchronizeStream(stream);

The following is an example of synchronizing AICPU thread resources:

// --Host-side call process--
// Allocate one host stream.
aclrtStream stream;
aclrtCreateStream(&stream);
// Create a thread of the CPU_TS type based on the allocated stream.
ThreadHandle cpuThread;
HcclThreadAcquireWithStream(comm, COMM_ENGINE_CPU_TS, streams, 1, &cpuThread);

// Create a thread of the AICPU_TS type.
ThreadHandle aicpuThread;
HcclThreadAcquire(comm, COMM_ENGINE_AICPU_TS, 1, 1, &aicpuThread);

// Export the created thread of the AICPU_TS type as a thread available on the CPU.
ThreadHandle exportedCpuThread;
HcclThreadExportToCommEngine(comm, 1, &aicpuThread, COMM_ENGINE_CPU_TS, &exportedCpuThread);
// Export the created thread of the CPU type as a thread available on the AICPU.
ThreadHandle exportedAicpuThread;
HcclThreadExportToCommEngine(comm, 1, &cpuThread, COMM_ENGINE_AICPU_TS, &exportedAicpuThread);

// Send a synchronization signal.
HcommThreadNotifyRecordOnThread(cpuThread, exportedCpuThread, 0);
// Launch the kernel and export the exportedAicpuThread and aicpuThread to the AICPU.
// ...
uint32_t timeout = 1;
// Wait for a synchronization signal.
HcommThreadNotifyWaitOnThread(cpuThread, 0, timeout);

// --Call process on the AICPU--
// Wait for a synchronization signal.
uint32_t timeout = 1;
HcommThreadNotifyWaitOnThread(aicpuThread, 0, timeout);
// Orchestrate the task and dispatch it to the aicpuThread.
// ...
// Send a synchronization signal.
HcommThreadNotifyRecordOnThread(aicpuThread, exportedAicpuThread, 0);

// Stream synchronization
aclrtSynchronizeStream(stream);