This repository collects all the repositories required to build and run Zephyr TensorFlow Lite Micro demos on either:
The hardware version of the magic wand
demo also requires a PmodACL to be connected to the JD
port of the Arty A7 board.
Clone the repository and submodules:
git clone https://github.com/antmicro/litex-vexriscv-tensorflow-lite-demo
cd litex-vexriscv-tensorflow-lite-demo
DEMO_HOME=`pwd`
git submodule update --init --recursive
Install prerequisites (tested on Ubuntu 18.04):
sudo apt update
sudo apt install cmake ninja-build gperf ccache dfu-util device-tree-compiler wget python python3-pip python3-setuptools python3-tk python3-wheel xz-utils file make gcc gcc-multilib locales tar curl unzip xxd
Install Zephyr prerequisites:
# update cmake to required version
sudo pip3 install cmake virtualenv
# install Zephyr SDK
wget https://github.com/zephyrproject-rtos/sdk-ng/releases/download/v0.11.2/zephyr-sdk-0.11.2-setup.run
chmod +x zephyr-sdk-0.11.2-setup.run
sudo ./zephyr-sdk-0.11.2-setup.run -- -d /opt/zephyr-sdk
Export Zephyr configuration:
export ZEPHYR_TOOLCHAIN_VARIANT=zephyr
export ZEPHYR_SDK_INSTALL_DIR=/opt/zephyr-sdk
Build the Hello World
demo with:
cd $DEMO_HOME/tensorflow
make -f tensorflow/lite/micro/tools/make/Makefile TARGET=zephyr_vexriscv hello_world_bin
The resulting binaries can be found in the tensorflow/lite/micro/tools/make/gen/zephyr_vexriscv_x86_64/hello_world/build/zephyr
folder.
Build the Magic Wand
demo with:
cd $DEMO_HOME/tensorflow
make -f tensorflow/lite/micro/tools/make/Makefile TARGET=zephyr_vexriscv magic_wand_bin
The resulting binaries can be found in the tensorflow/lite/micro/tools/make/gen/zephyr_vexriscv_x86_64/magic_wand/build/zephyr
folder.
Note: For this section, if you have not already updated your udev rules, follow the instructions at "Download & setup udev rules" -- you probably won't need to reboot.
The FPGA bitstream (gateware) can be built using Litex Build Environment. Building the gateware currently requires Xilinx's FPGA tooling, Vivado, to be installed in the system.
Build the gateware with:
cd $DEMO_HOME/litex-buildenv
# Some of LiteX Buildenv's scritps have problems when running in a git repository in detached state.
# Let's create a fake branch to avoid build errors.
git checkout -b tf_demo
export CPU=vexriscv
export CPU_VARIANT=full
export PLATFORM=arty
export FIRMWARE=zephyr
export TARGET=tf
./scripts/download-env.sh
source scripts/enter-env.sh
make gateware
Once you have synthesized the gateware, load it onto the FPGA with:
make gateware-load
With the FPGA programmed, you can load the Zephyr binary on the device using the flterm program provided inside the environment you just initialized above:
cd $DEMO_HOME/tensorflow
flterm --port=/dev/ttyUSB1 --kernel=tensorflow/lite/micro/tools/make/gen/zephyr_vexriscv_x86_64/magic_wand/build/zephyr/zephyr.bin --speed=115200
See the Litex Build Environment Wiki for more available options.
The renode
directory contains all necessary scripts and assets required to simulate the Magic Wand
demo.
Install Renode as detailed in its README file.
Build the Magic Wand
demo as described in the section above or use a prebuilt binary from the binares/magic_wand
directory.
Now you should have everything to run the simulation using the locally built binary:
cd $DEMO_HOME/renode
renode -e "s @litex-vexriscv-tflite.resc"
Or using the prebuilt one:
cd $DEMO_HOME/renode
renode -e "set zephyr_elf @../binaries/magic_wand/zephyr.elf; s @litex-vexriscv-tflite.resc"
You should see the following output on the UART (which will open as a separate terminal in Renode automatically):
*** Booting Zephyr OS build 0.6.0-86741-g626bb2c4d0bd ***
4 bytes lost due to alignment. To avoid this loss, please make sure the tensor_arena is 16 bytes aligned.
Got accelerometer, label: accel-0
RING:
*
* *
* *
* *
* *
* *
*
SLOPE:
*
*
*
*
*
*
*
* * * * * * * *
RING:
*
* *
* *
* *
* *
* *
*
SLOPE:
*
*
*
*
*
*
*
* * * * * * * *
In order to exit Renode, type quit
in Monitor (the window with the Renode logo and (machine-0)
prompt).
Refer to the Renode documentation for details on how to use Renode to implement more complex usage scenarios, and use its advanced debug capabilities, or set up CI testing your ML-oriented system. If you need commercial support, please contact us at support@renode.io.