Sample Zephyr App using Kenning Zephyr Runtime for inference

October 3, 2025 · View on GitHub

This is a sample standalone Zephyr app that uses Kenning Zephyr Runtime library for inference of any model.

Building the project

Cloning the project

First, clone the project to provide all the necessary files for building the application and Docker image:

git clone git@github.com:antmicro/kenning-zephyr-runtime-example-app.git
cd kenning-zephyr-runtime-example-app

Installing dependencies

One way to obtain dependencies is to use the Docker image:

docker build -t kenning-zephyr-runtime-app -f environments/Dockerfile .
docker run --rm -it -v $(pwd):$(pwd) -w $(pwd) kenning-zephyr-runtime-app /bin/bash

To set things locally, install:

Zephyr dependencies
jq
curl
west
patch
CMake

On Debian-based systems, run:

sudo apt-get update -y
sudo apt-get install -y --no-install-recommends \
  ccache cmake colorized-logs curl device-tree-compiler dfu-util file fonts-lato \
  g++-multilib gcc gcc-multilib git jq libmagic1 libsdl2-dev make ninja-build \
  python3-dev python3-pip python3-setuptools python3-tk python3-wheel python3-venv \
  mono-complete wget xxd xz-utils patch

python3 -m venv .venv
source .venv/bin/activate
pip3 install west pyelftools

Downloading Zephyr SDK and necessary build dependencies

With all the system dependencies satisfied, you need to obtain all Zephyr repositories and additional tools. To do so, initialize west and install Kenning Zephyr Runtime dependencies:

west init -l app
west update
west zephyr-export

Then, prepare KZR modules and dependencies:

python3 -m pip install -r zephyr/scripts/requirements.txt
python3 -m pip install -r kenning-zephyr-runtime/requirements.txt
pushd ./kenning-zephyr-runtime
./scripts/prepare_modules.sh
popd

In the end, set up the Zephyr SDK (this step can be omitted in the Docker image).

ZEPHYR_SDK_ONLY=1 ./kenning-zephyr-runtime/scripts/prepare_zephyr_env.sh

Building the application

In addition to providing a platform, the app requires:

A runtime - -DEXTRA_CONF_FILE option, can be tvm.conf or tflite.conf
A model - -DCONFIG_KENNING_MODEL_PATH option, can be URL to model or existing local path.

To build an app using TVM as backend, run:

west build \
  -p always \
  -b max32690evkit/max32690/m4 app -- \
  -DEXTRA_CONF_FILE="tvm.conf" \
  -DCONFIG_KENNING_MODEL_PATH=\"https://dl.antmicro.com/kenning/models/classification/magic_wand.h5\"
west build -t board-repl

To build an app using TFLite as backend, run:

west build \
  -p always \
  -b max32690evkit/max32690/m4 app -- \
  -DEXTRA_CONF_FILE="tflite.conf" \
  -DCONFIG_KENNING_MODEL_PATH=\"https://dl.antmicro.com/kenning/models/classification/magic_wand.h5\"
west build -t board-repl

Running the app

To run the app on a simulated platform, first download Renode portable (this step can be omitted in the Docker container):

source kenning-zephyr-runtime/scripts/prepare_renode.sh

To start the simulation, run:

python3 kenning-zephyr-runtime/scripts/run_renode.py --no-kcomm

To stop the simulation, use Ctrl-C.

Profiling the app with Zephelin

Zephelin is a library for profiling Zephyr applications.

To use it, first, install dependencies and prepare modules:

python3 -m pip install -r zephelin/requirements.txt
west forall -c "git reset HEAD --hard && git clean -fd"
west patch -sm zephelin apply

Then, build the app with additional zpl.conf configuration file and a flag enabling tracing for a specific framework:

west build \
  -p always \
  -b max32690fthr/max32690/m4 app -- \
  -DEXTRA_CONF_FILE="tflite.conf;zpl.conf" \
  -DCONFIG_KENNING_MODEL_PATH=\"https://dl.antmicro.com/kenning/models/classification/magic_wand.h5\" \
  -DCONFIG_ZPL_TFLM_PROFILER=y
west build -t board-repl

If the platform is simulated, ensure Renode is prepared (this step is not required if you did it previously in the current environment):

source kenning-zephyr-runtime/scripts/prepare_renode.sh

For example, capturing the trace for 10 seconds can be using the following command:

python zephelin/scripts/run_renode.py --trace-output ./channel0_0 --timeout 10 --repl build/max32690fthr.repl

Using Zephelin CLI, you can create a TEF trace from the output:

west zpl-prepare-trace ./channel0_0 \
  --tflm-model-path build/modules/kenning-zephyr-runtime/lib/kenning_inference_lib/runtimes/tflite/generated/model.tflite \
  -o ./tef_tflm_profiler.json

This trace can be visualized in Zephelin Trace Viewer.