N2D2-IP only: available upon request.
- Export type:
C++ export for STM32.
n2d2 MobileNet_ONNX.ini -seed 1 -w /dev/null -export CPP_STM32
This export inherit the properties and optimizations from the C++ export, but includes optimized kernels for the Cortex-M4 and the Cortex-M7. Please refer to the Export: C++ for the available export parameters.
SMLADintrinsic is used to do two 16-bit signed integers multiplications with accumulation. To extend the 8-bit data to the necessary 16-bit, the
XTB16intrinsic is used.
- Loop unrolling
The unrolling of the loops can be done with
#pragma GCC unroll NB_ITERATIONSbut it does not always perform as well as expected. Some loops are manually unrolled instead using C++ templates. This increases the size of the compiled binary further but it provides a faster inference.
- Usage of intrinsics
Intrinsics provided by ARM are preferred to normal library methods calls when possible. For example the
USATintrinsics are used to clamp the output value resulting in better results than a naive call to the std::clamp method.
n2d2 MobileNet_ONNX.ini -seed 1 -w /dev/null -export CPP_STM32 -fuse -nbbits 8 -calib -1 -db-export 100 -test
This command generates a C++ project in the sub-directory
This project is ready to be cross-compiled with a
Makefile, using the
GNU Arm Embedded Toolchain (which provides the
To cross-compile the project using the GNU Arm Embedded Toolchain. An ELF binary file is generated in
To flash the board using OpenOCD with the previously generated
bin/n2d2_stm32.elfbinary. In the provided Makefile, the default OpenOCD location is
/usr/local/bin/openocdand the default script is
stm32h7x3i_eval.cfg, for the STM32H7x3I evaluation board family. These can be changed in the first lines of the Makefile.