Bare-metal IREE runtime simulated using Renode

This tutorial shows how to compile models with IREE and evaluate them using the Kenning bare-metal IREE runtime working in Renode.

The platform used for testing is a Springbok RISC-V AI accelerator with V Extension. For more details, check:


For quick setup, use docker to open the environment with all necessary dependencies installed. Alternatively, install them manually by following the instructions in the runtime’s

docker run --rm -it

Kenning uses pyrenode3 package, which requires Renode to work. To install it, download the latest Renode Arch package and store its location in PYRENODE_PKG:

export PYRENODE_PKG=$(realpath renode-latest.pkg.tar.xz)

For other configuration options check pyrenode3

Evaluating the model in Kenning

Kenning can evaluate the runtime running on a device simulated in Renode. This allows us to:

  • Analyze model behavior without the need for physical hardware

  • Check model and runtime performance and quality on a simulated device in Continuous Integration pipelines

  • Obtain detailed metrics regarding device usage (e.g. histogram of instructions)

Creating the scenario

The following scenario is used for evaluating the model on Springbok AI accelerator in Renode:

    "dataset": {
        "type": "kenning.datasets.magic_wand_dataset.MagicWandDataset",
        "parameters": {
            "dataset_root": "./build/MagicWandDataset"
    "model_wrapper": {
        "type": "kenning.modelwrappers.classification.tflite_magic_wand.MagicWandModelWrapper",
        "parameters": {
            "model_path": "kenning:///models/classification/magic_wand.h5"
            "type": "kenning.optimizers.iree.IREECompiler",
                "compiled_model_path": "./build/tflite-magic-wand.vmfb",
                "backend": "llvm-cpu",
                "model_framework": "keras",
                "compiler_args": [
    "runtime": {
        "type": "kenning.runtimes.renode.RenodeRuntime",
        "parameters": {
            "runtime_binary_path": "kenning:///renode/springbok/iree_runtime",
            "platform_resc_path": "gh://antmicro:kenning-bare-metal-iree-runtime/sim/config/springbok.resc;branch=main",
            "resc_dependencies": [
            "post_start_commands": [
                "sysbus.vec_controlblock WriteDoubleWord 0xc 0"
            "runtime_log_init_msg": "Runtime started",
            "profiler_dump_path": "build/profiler.dump"
    "protocol": {
        "type": "kenning.protocols.uart.UARTProtocol",
        "parameters": {
            "port": "/tmp/uart",
            "baudrate": 115200,
            "endianness": "little"

The model used in the scenario is a classifier trained on Magic Wand dataset for accelerometer based gesture recognition:

  • dataset entry provides a class for managing the Magic Wand dataset (downloading the dataset, parsing data from files, an evaluating the model on Springbok by sending inputs and comparing outputs to ground truth).

  • model_wrapper entry provides the model to optimize, I/O specification and model-specific processing.

IREE compiler is enabled by adding it to the optimizers. It optimizes and compiles the model to Virtual Machine Flat Buffer format (.vmfb), which is later executed on a minimal IREE virtual machine. The additional flags provided to the compiler specify the RISC-V target architecture and V Extension features compatible with the Springbok accelerator.

Renode simulation is enabled by specifying the RenodeRuntime in the runtime entry and setting the following parameters:

  • runtime_binary_path - path to the runtime binary (in this case IREE runtime)

  • platform_resc_path - path to the Renode script (.resc file) used for setting up the emulation

  • resc_dependencies - additional dependencies for RESC files

  • post_start_commands - Renode monitor commands executed after emulation starts

  • profiler_dump_path - path for the profiler dump, which is then used for report generation

UART is used for Kenning communications by selecting UARTProtocol in protocol entry.

Running the scenario

Evaluate the model in Renode using the created scenario, and generate a report with performance and quality metrics:

kenning optimize test report \
    --json-cfg \
    --measurements ./results.json \
    --report-path ./reports/ \
    --report-name v-extensions-riscv \
    --model-names magic_wand_fp32 \
    --verbosity INFO \

Kenning executes the scenario in the following steps:

  • Kenning loads the model with Keras and compiles it with IREE to ./build/magic-wand.vmfb

  • The machine with the Springbok AI accelerator is created in Renode

  • Connection to the simulated device is established via UART

  • Kenning sends the compiled model and I/O specification

  • Input data is sent in a loop, the bare-metal IREE runtime performs inference and sends the results back

When the scenario is finished Kenning stores runtime metrics, profiling results and model evaluation in results.json.

The human readable report generated from results.json will be available under reports/springbok-magic-wand/

Last update: 2024-05-01