Reference collector receives NULL name for __itt_task_begin after first in PyTorch

[nchaimov]

I'm trying to write a tool which will use the __itt_task_begin and __itt_task_end calls emitted by the ITT instrumentation built into PyTorch. To that end, I was experimenting with the reference collector found in this repository, but am finding that it does not work correctly with PyTorch, even though VTune does.

Specifically, with the reference collector, I find that the first call to __itt_task_begin gets logged correctly. However, the second and all subsequent calls yields an Incorrect function call warning, which is due to the __itt_string_handle *name argument being NULL.

The simplest example which demonstrates the problem manually annotates two tasks:

import torch

torch.profiler.itt.range_push("Hello")
torch.profiler.itt.range_push("World")
torch.profiler.itt.range_pop()
torch.profiler.itt.range_pop()

If I run this example with Intel VTune 2023.1.0, the output correctly includes both tasks:

$ vtune -collect hotspots python pytorch_itt_manual.py

[...]
Top Tasks
Task Type  Task Time  Task Count  Average Task Time
---------  ---------  ----------  -----------------
Hello         0.000s           1             0.000s
World         0.000s           1             0.000s

However, if I use the reference collector, only the first task is correctly logged. The second one is invalid, because the name argument is NULL.

$ INTEL_LIBITTNOTIFY_LOG_DIR=$PWD INTEL_LIBITTNOTIFY64=$HOME/ittapi/src/ittnotify_refcol/libittrefcol.so  python pytorch_itt_manual.py
$ cat libittnotify_refcol_202555124144.log
[INFO] __itt_task_begin(...) - functions args: domain=PyTorch handle=Hello
[WARN] __itt_task_begin(...) - Incorrect function call
[INFO] __itt_task_end(...) - functions args: domain=PyTorch
[INFO] __itt_task_end(...) - functions args: domain=PyTorch

The second call to __itt_task_begin doesn't work, nor will any subsequent calls if there are more than two.

My assumption is that the problem is with the reference collector rather than PyTorch, since this works correctly with VTune, although I am not certain of this.

Does anyone have any advice about how I can get the reference collector to get the name handles for all __itt_task_begin annotations emitted by PyTorch?

[eparshut]

@nchaimov, Thank you for submitting the problem!

The problem happens because the reference collector does not include the __itt_string_handle_create() function implementation. This function is needed to create and store string handle object during the program's execution (task names in your case).

When the first __itt_task_begin() is called, the static ITT API library handles the creation of the string handle successfully. However, after this first call the reference collector has been loaded and it tries to handle subsequent calls of the __itt_string_handle_create(). Since this function is missing in the reference collector, it returns NULL, causing the second and later __itt_task_begin() calls to fail.

To fix this, need to add the __itt_string_handle_create() function implementation to the reference collector. This implementation should create and manage string handle objects to ensure they persist across the application's execution. This will ensure that all __itt_task_begin() calls work correctly.