Exporting sources for C++ only operator/kernel implementations

Development
#1

For a custom operator I am currently implementing I generate C-code following the BYOC approach. As I have to include 3rd party C++ header files containing namespaces etc. I have not really found a way to compile my kernels (In my case src/lib2.c) without changing the file extension to .cc or .cpp.

Also it would be great to use C++ features in the kernel Implementations as well. Is this currently not supported (maybe due to the overhead this would have for embedded deployment) or have I overlooked anything?

My implementations should be mainly deployed in an uTVM environment I am exporting the output files for the target device using either of these approaches:

  1. following the (default) uTVM Flow using tvm.micro.build_static_runtime())
  2. Using the Model Library Format (tvm.micro.export_model_format()) from uTVM (and compile externally)

However I think my concern is not uTVM specific.

My Quick fix to generate lib*.cc files instead of lib*.c files without needing to rename them manually after the export is based on hardcoding the file extension in python/tvm/micro/model_library_format.py (for (2)) and python/tvm/runtime/module.py (for (1)).

Now I would like to find something more robust to contribute it to the TVM codebase but I am not sure on how to integrate it in a meaningful way. My main two question are:

  • How to tell TVM that C++ files need to be generated instead of using the default .c file extension. This maybe passed in as an optional argument by the user (for all CSourceModules that are going to be build together) or internally on a per-module basis (f.e. in by BYOC code)
  • Can I reuse the available CSourceModule and stick with the type_key "c" for C++ as well by just changing the filename extension or should I add another type key like "cc"/"cpp" with an separate SourceModuleNode.

I would appreciate a lot if someone could give me a hint on wether I have misunderstood anything and if my proposal sounds like an improvement to the TVM ecosystem, feel free to discuss my questions on the possible implementations as well.