Hi @Tonylyc,
Welcome to these forums and thanks for your question. I should clarify a bit the efforts in my thesis, because they have moved quite a bit from deploying on this development board. We are using a single core from GAP8 paired with a custom accelerator to deploy NN’s on. As stated in this post: Feedback on TVM port to custom accelerator
I have no experience with enabling multicore on the GAP8, but I’d think the best thing you can do there is to reuse some arm code from the regular TVM flow (no BYOC). I also don’t know how the multiple cores are addressed by the GAP8 SDK, and how well this integrates with TVM.
IIRC some people at University of Bologna were working on a library for efficiently executing (Q)NN layers on PULP-like architectures (main authors of https://arxiv.org/pdf/2008.07127.pdf) , maybe the easiest way to get started with GAP8 is to use BYOC and integrate the library that way?
You should know that a lot has changed since a year in microTVM. A lot of those changes will be discussed in the following TVMcon https://www.tvmcon.org/ I’d suggest you go and look there if you’d like to get an update.
Anyway, I’m currently not looking into deployment specifically on GAP8 anymore, but do let me know if you have any further questions!