Skip to content

Latest commit

 

History

History
87 lines (61 loc) · 4.17 KB

README.md

File metadata and controls

87 lines (61 loc) · 4.17 KB
Raw

Developer Tools Examples

This directory contains examples of BundledProgram and ETDump generation.

Directory structure

examples/devtools
├── scripts                           # Python scripts to illustrate export workflow of bundled program.
├── executor_runner                   # Contains an example for both BundledProgram to verify ExecuTorch model, and generate ETDump for runtime results.
├── CMakeLists.txt                    # Example CMakeLists.txt for building executor_runner with Developer Tools support.
├── build_example_runner.sh           # A convenient shell script to build the example_runner.
├── test_example_runner.sh            # A convenient shell script to run the example_runner.
└── README.md                         # Current file

BundledProgram

We will use an example model (in torch.nn.Module) and its representative inputs, both from models/ directory, to generate a BundledProgram(.bpte) file using the script. Then we will use devtools/example_runner to execute the .bpte model on the ExecuTorch runtime and verify the model on BundledProgram API.

  1. Sets up the basic development environment for ExecuTorch by Setting up ExecuTorch from GitHub.

  2. Using the script to generate a BundledProgram binary file by retreiving a torch.nn.Module model and its representative inputs from the list of available models in the models/ dir.

cd executorch # To the top level dir

# To get a list of example models
python3 -m examples.devtools.scripts.export_bundled_program -h

# To generate a specific `.bpte` model
python3 -m examples.devtools.scripts.export_bundled_program -m mv2 # for MobileNetv2

# This should generate ./mv2_bundled.bpte file, if successful.
  1. Once we have the BundledProgram binary (.bpte) file, then let's run and verify it with ExecuTorch runtime and BundledProgram APIs using the devtools/example_runner.
   cd executorch
   ./examples/devtools/build_example_runner.sh
   ./cmake-out/examples/devtools/example_runner --bundled_program_path mv2_bundled.bpte --output_verification

ETDump

Getting Started

After exporting a BundledProgram, runtime profiling and debug data can be collected in an ETDump. An ETDump is a buffer containing data generated by hooks within the ExecuTorch runtime. We offer an example runner that accepts a BundledProgram (.bpte) and runs a single iteration over the first method defined.

Running the program will generate an ETDump file (.etdp) at the location specified by --etdump_path.

   ./cmake-out/examples/devtools/example_runner --bundled_program_path mv2_bundled.bpte --etdump_path mv2_etdump.etdp

Parsing ETDump

Once an ETDump has been generated, it can be viewed using the CLI inspector. This will print a tabular view of the data recorded in the ETDump.

   python3 -m devtools.inspector.inspector_cli --etdump_path mv2_etdump.etdp

ETDump C++ API

ETDump profiling can also be used in a custom C++ program. ETDumpGen is an implementation of the abstract EventTracer class. Include the header file located at devtools/etdump/etdump_flatcc.h. To initialize the ETDump generator, construct it before loading the method from the program.

   executorch::etdump::ETDumpGen etdump_gen;
   Result<Method> method =
      program->load_method(method_name, &memory_manager, &etdump_gen);

Since the EventTracer hooks are embedded within the runtime, profiling and debug data will be automatically recorded.

Once execution has completed, finalize the ETDump buffer. This returns an etdump_result, a struct with the finalized buffer and its size.

   etdump_result result = etdump_gen.get_etdump_data();
   if (result.buf != nullptr && result.size > 0) {
    FILE* f = fopen(FLAGS_etdump_path.c_str(), "w+");
    fwrite((uint8_t*)result.buf, 1, result.size, f);
    fclose(f);
    free(result.buf);
  }