2.6 KiB
Installation
There are two primary ways to set up the testing environment.
Native Installation
To run the test scripts, you must have a non-static, user-mode installation
of QEMU available in your system's PATH. For example, to test on ARM, the
qemu-arm binary must be installed and accessible from your shell.
Docker Environment
As an alternative to installing dependencies on your host machine, you can use the provided Docker environment. For instructions, see the "Run locally for debugging" section in the Dockerfile.
Software Emulation with QEMU
TensorFlow Lite Micro uses QEMU to run tests that have been cross-compiled for architectures other than x86_64. This setup enables the rapid execution of unit tests for various hardware targets directly on a standard development machine.
Running Unit Tests
You can run all Cortex-M unit tests using a single build script:
tensorflow/lite/micro/tools/ci_build/test_cortex_m_qemu.sh
After this script completes the initial build, you can run an individual test target as follows:
make -f tensorflow/lite/micro/tools/make/Makefile \
TARGET=cortex_m_qemu \
TARGET_ARCH=cortex-m3 \
OPTIMIZED_KERNEL_DIR=cmsis_nn
<TEST_TARGET>
Debugging
You can debug failures either interactively with GDB or by analyzing the core file after a crash.
Live GDB Debugging
Terminal 1: Start QEMU, telling it to wait for a GDB connection on port 1234.
qemu-arm -cpu cortex-m3 -g 1234 <TEST_BINARY_PATH>
Terminal 2: Launch GDB and connect to the waiting QEMU session.
$ arm-none-eabi-gdb <TEST_BINARY_PATH>
(gdb) target remote :1234
Core Dump Analysis
Load the test executable and the generated core file into GDB to inspect the state at the time of the crash.
$ arm-none-eabi-gdb <TEST_BINARY_PATH> <CORE_FILE_PATH>
(gdb) bt
Useful External Links for QEMU
-
The current testing framework uses QEMU's user mode. Read doc
-
QEMU uses ARM semihosting to replace newlib system calls for specific boards with the host OS. Further documentation on how this works is contained in
cortex_m_qemu_makefile.inc.