Skip to content

Latest commit

 

History

History
93 lines (60 loc) · 3.01 KB

development.md

File metadata and controls

93 lines (60 loc) · 3.01 KB
Raw

Developping onnx2c

Adding a new node/op implementation

If you are adding the implementation for a new ONNX operator:

  • create the implementation file for it in src/nodes/ (check out src/nodes/TEMPLATE for a starting point). (At the begining of the project implementations were all put into a .h file, but with the multitude of nodes, the compilation times slow down, so prefer using both .h and .cc)
  • add a case to the switch in Graph::findNode() in src/graph.cc and an #include in the lines above that function
  • add the new file to CMakeLists.txt
  • add onnx backed tests for the new node in test/CMakeLists.txt. Search for them in onnx/onnx/backend/test/data/node.

Testing

The main onnx2c testing is built ontop of the ONNX backend node tests.

On top of these there are two classes of benchmarking tests in the test suite

  • Google Benchmark based. This is the new version.
  • ONNX model zoo based. Old, and less useful.

And there is a limited support to test performance on top of embedded targets.

Unit tests

The onnx2c build runs onnx backend tests as unit/acceptance tests.

To run these, continue the build steps with:

make
make test

All test should and must pass.

Google Benchmark based tests

The benchmark binary is built in test/benchmarks as a part of the unit test framework.

Run it by executing the fake custom target run_benchmarks (e.g. make run_benchmarks).

The run_benchmarks is intended as a development tool. It is useful only when chaninging the generated code from those operators/nodes that are included in the benchmark suite.

Note, run_benchmarks is host computer specific, and must be first run with a clean master build to get a reference baseline. See the comments in test/benchmarks/host/benchmark_helper.sh for more info.

ONNX model zoo based tests

These are mostly deprecated, but the infrastructure is still left in place.

Run:

cd tests/onnx_model_zoo
./donwload.sh

and continue with a fresh build (i.e. re-run cmake).

Included are implementations of e.g. Squeezenet and Alexnet. Some of these take minutes to compile, so they are mostly interesting for onnx2c development.

On target testing

There is a build system in the scripts folder to complile and run any .onnx file on an embedded development board. It assumes LibOpenCM3, a arm-none-eabi-gcc compiler chain, and OpenOCD are available in PATH and the OPENCM3_DIR environment variable.

For now, the script hard-codes STM32F411 NUCLEO as the target (pull requests to remedy this welcome!).

To test if your own network file will fit, run scripts/measure_stm32f411_nucleo.sh [file.onnx]. This will compile, flash and report memory usage and runtime for the .onnx file. No checks of correctnes are made.

A collection of benchmarking tests can be run with make run_target_stm32f411_benchmarks.

Coding style

  • indent with tab, align with space