Software Emulation with Renode

TensorFlow Lite Micro makes use of Renode to for software emulation.

Here, we document how Renode is used as part of the TFLM project. For more general use of Renode, please refer to the Renode documentation.

Installation

Renode can be installed and used in a variety of ways, as documented here. For the purpose of Tensorflow Lite Micro, we make use of a portable version for Linux.

  1. Download portable version of Renode for Linux:

    tensorflow/lite/micro/testing/download_renode.sh tensorflow/lite/micro/tools/make/downloads/renode

  2. Install the Renode test dependencies

    pip3 install -r tensorflow/lite/micro/tools/make/downloads/renode/tests/requirements.txt

At this point in time you will be ready to run TFLM tests with Renode.

Running Unit Tests

All the tests for a specific platform (e.g. bluepill) can be run with:

make -f tensorflow/lite/micro/tools/make/Makefile TARGET=bluepill test
  • This makes use of the robot framework from Renode.
  • Note that the tests can currently not be run in parallel.
  • It takes about 25 second to complete all tests, including around 3 seconds for suite startup/teardown and average 0.38 second per test.

Under the hood of the Testing Infrastructure

Describe how we wait for a particular string on the UART. Some pointers into the robot files as well as any relevant documentation from Renode.

A test failure is the absence of a specific string on the UART so the test will wait for a specific timeout period (configured in the .robot) file before failing.

  • What this means in practice is that a failing test will take longer to finish than a test that passes.

  • If needed, an optimization on this would be to have a specific failure message as well so that both success and failure can be detected quickly.

Running a non-test Binary with Renode

Renode can also be used to run and debug binaries interactively. For example, to debug kernel_addr_test on Bluepill platform, run Renode:

tensorflow/lite/micro/tools/make/downloads/renode/renode

and issue following commands:

mach create
machine LoadPlatformDescription @platforms/cpus/stm32f103.repl    
# Create additional semihosting interface peripheral
machine LoadPlatformDescriptionFromString "uartSemihosting: UART.SemihostingUart @ cpu"
# Open separate window for semihosting UART output
showAnalyzer sysbus.cpu.uartSemihosting
# Load ELF file
sysbus LoadELF @tensorflow/lite/micro/tools/make/gen/bluepill_cortex-m3/bin/kernel_add_test
# Start simulation
start

You can also connect GDB to the simulation. To do that, start the GDB server in Renode before issuing the start command:

machine StartGdbServer 3333

Than you can connect from GDB with:

target remote localhost:3333

For further reference please see the Renode documentation.