TVM integration for HANNAH

To get the basic installation use:

1. Install git-lfs on ubuntu apt install git-lfs
2. checkout submodules: git submodule update --init --recursive
3. poetry install -E automate -E dash -E micro
4. poetry shell

For the tvm installation there are the following installation options.

MicroTVM installation

This installation installs tvm with microtvm support without llvm backend

./scripts/install_micro.sh
./scripts/install_zephyr.sh
source env

Zephyr installation is optional but is currently needed for host driven execution. For an example using host driven execution and the micro tvm zephyr runtime on stm32f429i discovery boards see samples/micro/zephyr_host_driven_stm32f429i.py.

MicroTVM with pulp-llvm support

This installation option activates the pulp-llvm based backend for direct vectorization on xpulpv targets. For pulp targets no zephyr support is needed at the moment.

./scripts/install_micro_pulp.sh

Full installation

The full installation uses the host provided llvm backend and activates cuda and llvm backends if available.

./scripts/install_full.sh

Common error reasons

1. Pythonpath not set when using automate runner on schrank boards

Add AcceptEnv LANG LC_* PYTHONPATH to /etc/ssh/sshd_config and restart server

Result Visualization

Currently a basic result visualization is available via hannah-tvm-dashboard