Hello community, I am currently trying to deploy a partitioned Graph on multiple CPUs within our specialized hardware
- the board contains 152 Processing elements , which have arm processors.
- the software stack is built using c.
what I already did is the following :
I partitioned the graph using the pipeline_executor.py and I have them exported into lib.so and param files, from what is available , the pipeline executor relies on the graph executor,
on the other hand , I have created a template test project that was successfully deployed , this project is the implementation of a dummy network on a single core, it relies on the AOT executor and crt library.
so now i am trying to find a connection between the two sides : microtvm support and pipe-lining.
my current idea is to hard-code the process doing the following:
- create the Network
- divide it to submodules( e.g 2 submodules) using the graph_split function
- export each submodule independantly using the export_model_library_format()
- generate independant project for each exported library,
- manually add pipelining through interruptions ,
- compile and test.
the thing i want to point also is the presence of these files ,
microtvm_runtime.cc
microtvm_graph_executor.cc
threading_backend.cc
I will be thankful if I can get your help or any idea that could be relevant to the issue .