mstxRangeStartA

Function Description

Marks the start position of the mstx range capability.

Function Prototype

C/C++:

mstxRangeId mstxRangeStartA(const char *message, aclrtStream stream)

Python:

mstx.range_start(message, stream)

Parameters

**Table 1** Parameters
Parameter	Input/Output	Description
message	Input	message indicates the marked text, which carries the dotting information. Data type in C/C++: const char . In Python, message* is a character string. The default value is None. Length of the input message string: In MSPTI scenarios, the length cannot exceed 255 bytes. In scenarios other than MSPTI (such as msProf command line and Ascend PyTorch Profiler), the length cannot exceed 156 bytes. NOTE: message cannot be a null pointer.
stream	Input	stream indicates the thread that uses the mark. Data type in C/C++: aclrtStream. In Python, stream is an aclrtStream object. The default value is None. Marks only instantaneous events on the host when nullptr is configured. Identifies instantaneous events on the host and device when a valid stream is configured.

Returns

If 0 is returned, the operation fails.

Calling Example

C/C++ calling:

...
bool RunOp()
{
// create op desc
...
const char *message = "h1";
mstxRangeId id = mstxRangeStartA(message, NULL);
...
// Run op
if
(!opRunner.RunOp()) {
ERROR_LOG("Run
op failed");
return false;
}
mstxRangeEnd(id);
...
}

Python calling method 1:

The Python API is used to implement related API content in the C/C++ language and compile and generate .so files. The .so files can be directly referenced by Python in PYTHONPATH.

import mstx
mstx.range_start("aaa")
print(1)
mstx.range_end(1)
import torch
import torch_npu
a = torch.Tensor([1,2,3,4]).npu()
b = torch.Tensor([1,2,3,4]).npu()
hi_str = "hi"
hello_str = "hello"
hi_id = mstx.range_start(hi_str, None)
c = a + b
hello_id = mstx.range_start(hello_str, stream=None)
d = a - b
mstx.range_end(hi_id)
e = a * b
mstx.range_end(hello_id)

Python calling method 2:

Use Python for development, use ctypes.CDLL("libms_tools_ext.so") to reference the original .so file of mstx, and use the APIs provided in the file.

import mstx
import torch
import torch_npu
import acl
import sys
import ctypes
lib = ctypes.CDLL("libms_tools_ext.so")
# Define the parameter type and return type of the function.
lib.mstxRangeStartA.argtypes = [ctypes.c_char_p, ctypes.c_void_p]
lib.mstxRangeStartA.restype = ctypes.c_uint64
lib.mstxRangeEnd.argtypes = [ctypes.c_uint64]
lib.mstxRangeEnd.restype = None
a = torch.Tensor([1,2,3,4]).npu()
b = torch.Tensor([1,2,3,4]).npu()
# Create a ctypes.c_char_p pointer.
hi_str = b"hi"
hi_ptr = ctypes.c_char_p(hi_str)
hi_id = ctypes.c_uint64()
# Create a ctypes.c_char_p pointer.
hello_str = b"hello"
hello_ptr = ctypes.c_char_p(hello_str)
hello_id = ctypes.c_uint64()
# Call functions.
hi_id.value = lib.mstxRangeStartA(hi_ptr, None)
c = a + b
hello_id.value = lib.mstxRangeStartA(hello_ptr, None)
d = a - b
lib.mstxRangeEnd(hi_id)
e = a * b
lib.mstxRangeEnd(hello_id)