This repo contains the source code for the userspace level runtime driver for Coral devices. This software is distributed in the binary form at coral.ai/software.
There are three ways to build libedgetpu:
- Docker + Bazel: Compatible with Linux, MacOS and Windows (via Dockerfile.windows and build.bat), this method ensures a known-good build enviroment and pulls all external depedencies needed.
- Bazel: Supports Linux, macOS, and Windows (via build.bat). A proper enviroment setup is required before using this technique.
- Makefile: Supporting only Linux and Native builds, this strategy is pure Makefile and doesn't require Bazel or external dependencies to be pulled at runtime.
For Debian/Ubuntu, install the following libraries:
$ sudo apt install docker.io devscripts
Build Linux binaries inside Docker container (works on Linux and macOS):
DOCKER_CPUS="k8" DOCKER_IMAGE="ubuntu:22.04" DOCKER_TARGETS=libedgetpu make docker-build
DOCKER_CPUS="armv7a aarch64" DOCKER_IMAGE="debian:bookworm" DOCKER_TARGETS=libedgetpu make docker-build
All built binaries go to the out directory. Note that the bazel-* are not copied to the host from the Docker container.
To package a Debian deb for arm64,armhf,amd64 respectively:
debuild -us -uc -tc -b -a arm64 -d
debuild -us -uc -tc -b -a armhf -d
debuild -us -uc -tc -b -a amd64 -d
The version of bazel needs to be the same as that recommended for the corresponding version of tensorflow. For example, it requires Bazel 6.5.0 to compile TF 2.16.1.
Current version of tensorflow supported is 2.16.1.
Build native binaries on Linux and macOS:
$ make
Required libraries for Linux:
$ sudo apt install python3-dev
Build native binaries on Windows:
$ build.bat
Cross-compile for ARMv7-A (32 bit), and ARMv8-A (64 bit) on Linux:
$ CPU=armv7a make
$ CPU=aarch64 make
To package a Debian deb:
debuild -us -uc -tc -b
NOTE for MacOS: Compilation with MacOS fails. Two requirements:
- install
flatbuffers(via macports) - after failure in compilation, add the following line to the temporary file that is created by bazel in
/var/tmp/_bazl_xxxxx/xxxxxxxxxxxxx/external/local_config_cc/BUILDline 48:
"darwin_x86_64": ":cc-compiler-darwin",
Repeat compilation.
If only building for native systems, it is possible to significantly reduce the complexity of the build by removing Bazel (and Docker). This simple approach builds only what is needed, removes build-time depenency fetching, increases the speed, and uses upstream Debian packages.
To prepare your system, you'll need the following packages (both available on Debian Bookworm, Bullseye or Buster-Backports):
sudo apt install libabsl-dev libflatbuffers-dev
Next, you'll need to clone the Tensorflow Repo at the desired checkout (using TF head isn't advised). If you are planning to use libcoral or pycoral libraries, this should match the ones in those repos' WORKSPACE files. For example, if you are using TF2.15, we can check that tag in the TF Repo get the latest commit for that stable release and then checkout that address:
git clone https://github.com/tensorflow/tensorflow
git checkout v2.16.1
To build the library:
TFROOT=<Directory of Tensorflow> make -f makefile_build/Makefile -j$(nproc) libedgetpu
If you have question, comments or requests concerning this library, please reach out to coral-support@google.com.
If you're using the Coral USB Accelerator, it may heat up during operation, depending on the computation workloads and operating frequency. Touching the metal part of the USB Accelerator after it has been operating for an extended period of time may lead to discomfort and/or skin burns. As such, if you enable the Edge TPU runtime using the maximum operating frequency, the USB Accelerator should be operated at an ambient temperature of 25°C or less. Alternatively, if you enable the Edge TPU runtime using the reduced operating frequency, then the device is intended to safely operate at an ambient temperature of 35°C or less.
Google does not accept any responsibility for any loss or damage if the device is operated outside of the recommended ambient temperature range.
Note: This issue affects only USB-based Coral devices, and is irrelevant for PCIe devices.