C Integer only BNN inference for RISC-V

This repository contains code for training different bayesian neural networks, perform optimized integer only inference on a simulated RISC-V CPU and analyze the obtained results.

This is the code of the paper Accelerating Bayesian Neural Networks on low-power edge RISC-V processors.

Requirements

• The requirements.txt file contains the Python packages required. Using a virtual python environment is recommended.
• GHDL 4.0.0 with LLVM backend.
• RISC-V GNU Compiler Toolchain with support for the following flags -march=rv32g -mabi=ilp32 -mno-div. A special patched version of the compiler is required for running some experiments.

Custom RISC-V instructions

The RISC-V cpu code of this repository has support for 2 new instructions for sampling a standard gaussian distribution. A patched version of the GNU Compiler Toolchain is required to run the last RISC-V Simulation tests, those are commented out on the test script and can be uncommented if the modified toolchain is present.

The custom_riscv_gcc directory contains the patched files of binutils. Those files must substitute binutils/opcodes/riscv-opc.c and binutils/include/opcodes/riscv-opc.h.

Components

The directory ext contains external linked github repositories. The directory libs contains the code of the libraries of this work

• BNN_for_hyperspectral_datasets_analysis. Python code to train bayesian neural networks for some of the most widely used open hyperspectral imaging datasets.

• bnnc. Python code to port Keras Tensorflow Probability models to C integer only inference code.

• riszeto. Python code that helps running simulations and compile C code for a RISC-V CPU written in VHDL. The simulations are run using GHDL.

Tests

The test directory contains the python scripts to run the paper experiments. The results are stored in a directory tmp that will be created in this repository base directory.

Hyperspectral

To evaluate and create the figures for the hyperspectral models run

python test/hyperspectral.py

Train Models

To train the LENET-5 and the B2N2 models for the CIFAR10 and MNIST dataset run

python test/train_models.py

Test Models

To test the trained models with different optimizations run

python test/test_models.py

RISC-V Simulation

To simulate the execution of a bayesian forward pass of the hyperspectral models inside the RISC-V CPU run

python test/test_rv.py