How to write an injection library?

[martin-belanger]

Nvidia provides a lot of documentation on how to instrument code with NVTX. However, I have not seen any documentation (or examples) about writing a library to collect the data generated. I know that I could use nsys profile to look at the collected data. However, I would like to just collect specific data (ranges) and save it to a text file.

To be precise, I'm trying to collect statistics from NCCL (especially the type and size of data being processed). I know that NCCL is already instrumented with NVTX. And I would like to write a plugin for NCCL that can collect specific data. However, I have no idea what APIs need to be provided by the plugin and could not find any documentation about it.

Is there documentation or code examples that does that?