This project demonstrates how to compile RISC-V 32-bit ML workloads via IREE, and deploy the workloads with IREE's c API to generate the bare-matal executables. The built artifacts are targeted for Springbok, a RISC-V 32-bit bare-metal platform, and can be simulated with Renode.

First install the system packages:

sudo apt install xxd cmake ninja-build wget
pip install -r python-requirements.txt

NOTE: IREE requires CMake >= 3.17. If your system's default cmake is in an older version, please update it manually.

To get you going we have pre-compiled an RV32 LLVM toolchain. This can be installed using:

./build_tools/install_toolchain.sh

The IREE compiler can be downloaded using:

./build_tools/download_iree_compiler.py

Finally Renode can be downloaded using:

./build_tools/download_renode.py

Make sure your ${HOME}/.local/bin is in your PATH:

export PATH=${HOME}/.local/bin:${PATH}

NOTE: If your python packages are installed with the virtual environment, please set PATH accordingly.

• build_tools: Utility scripts for the project
• cmake: CMake Macros for the project
• samples: Codegen and execution of ML models based on IREE
  • device: Device HAL driver library
  • float_model: float model examples
  • quant_model: quantized model examples
  • simple_vec_mul: Point-wise vector multiplication examples
  • util: Runtime utility library for model execution
• sim/config: Renode configuration and infrastructure
• springbok: Low-level code and linker scripts for the Springbok machine
• third_party/iree: IREE codebase

All sample models we've included can be built at once using this command:

./build_tools/build_riscv.sh

Elfs will land in build/build-riscv/samples/<sample_folder>/ and come in two flavors: <model_name>_bytecode_static and <model_name>_emitc_static. The bytecode executable uses the IREE VM, while the emitc executable compiles the VM commands into C.

To run a simulation, run ./build_tools/sim_springbok.sh with a path to a compiled executable. For example, to run MobileNet v1:

./build_tools/sim_springbok.sh build/build-riscv/samples/quant_model/mobilenet_v1_emitc_static

This project utilizes LLVM lit and FileCheck to test the ML executable performance. The tests are defined in the *_test.txt files under samples. To run the tests:

lit --path $(realpath build/iree_compiler/tests/bin) -a samples

Test times can be found at build/springbok_iree/tests/.lit_test_times.txt.