Function: set_exception_info_callback

Applicability

Description

Sets the exception callback function.

Prototype

• C Prototype

  1	aclError aclrtSetExceptionInfoCallback(aclrtExceptionInfoCallback callback)

• Python Function

  1	ret = acl.rt.set_exception_info_callback(fn)

Parameters

def exception_callback(exception_info)
"""
:exception_info: pointer address of aclrtExceptionInfo.
:return:
"""

Return Value

Restrictions

• Callback functions involve shared resources (such as locks). Exercise caution when using callback functions. Do not call APIs for resource application & release, stream synchronization, device synchronization, task delivery, and task termination in callback functions. Otherwise, errors or deadlocks may occur.
• Before executing an asynchronous task, you need to set an exception callback function. When an exception occurs during task execution on the device, the system transfers a pointer address of aclrtExceptionInfo that contains the task ID, stream ID, thread ID, device ID, and error code to the exception callback function, execute the callback function. You can call the acl.rt.get_task_id_from_exception_info, acl.rt.get_stream_id_from_exception_info, acl.rt.get_thread_id_from_exception_info, acl.rt.get_device_id_from_exception_info, and acl.rt.get_error_code_from_exception_info APIs to obtain the task ID, stream ID, thread ID, device ID, and error code to locate the fault.

  Application scenario Example: Before calling the acl.op.execute_v2 API, call the acl.rt.set_exception_info_callback API to set the exception callback function. When an exception occurs during operator execution on the device, the system passes a pointer address of aclrtExceptionInfo that contains the task ID, stream ID, thread ID, device ID, and error code to the exception callback function, execute the callback function.

• If the exception callback function is set repeatedly, the most recent setting applies.
• To clear the callback function, pass None or no parameter to the acl.rt.set_exception_info_callback call.

API Call Sequence

Application scenario: For example, if an AI Core error is reported during network inference (dynamic shape scenarios unsupported), you can call this API to obtain the description of the error operator and then perform further troubleshooting.

Sample Code

Add exception handling branches following the API calls. The sample code of key steps is as follows, which is for reference only. Do not directly copy and run the code.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

import acl
import numpy as np
# ......

# 1. Allocate runtime resources.
# ......

# 2. Load a model. After the model is successfully loaded, model_id that identifies the model is returned.
# ......

# 3. Create data of type aclmdlDataset to describe the inputs and outputs of the model.
# ......

# 4. Implement the exception callback function.
def exception_callback(info):
    stream_id = acl.rt.get_stream_id_from_exception_info(info)
    device_id = acl.rt.get_device_id_from_exception_info(info)
    task_id = acl.rt.get_task_id_from_exception_info(info)
    # You can write the obtained operator information to a file, or start another thread to listen to the exception callback. When an error occurs, the thread handling function is triggered, and the operator information is printed to the screen.
    op_name, input_desc, num_inputs, output_desc, num_outputs, ret = \
        acl.mdl.create_and_get_op_desc(device_id, stream_id, task_id, 256)

    # You can call related tensor APIs provided by acl to obtain the operator information as required.
    for i in range(num_inputs):
        desc = acl.get_tensor_desc_by_index(input_desc, i)
        address = acl.get_tensor_desc_address(desc)
        num_dims = acl.get_tensor_desc_num_dims(desc)
        dim_0, ret = acl.get_tensor_desc_dim_v2(desc, 0)

    for i in range(num_outputs):
        desc = acl.get_tensor_desc_by_index(output_desc, i)
        address = acl.get_tensor_desc_address(desc)
        num_dims = acl.get_tensor_desc_num_dims(desc)
        dim_0, ret = acl.get_tensor_desc_dim_v2(desc, 0)

    acl.destroy_tensor_desc(input_desc)
    acl.destroy_tensor_desc(output_desc)

# 5. Set the exception callback function.
ret = acl.rt.set_exception_info_callback(exception_callback)

# 6. Execute the model.
ret = acl.mdl.execute(model_id, input, output)

# 7. Process the model inference result.
# ......

# 8. Destroy allocations such as the model inputs and outputs, free memory, and unload the model.
# ......

# 9. Deallocate runtime resources.
# ......