This project aims to implement Sumcheck protocol on GPU via CUDA, following the algorithm described from here. This project aims to minimize C++ code and provide easy to use Rust API for launching kernel by leveraging bindgen and cudarc.

Architecture

bindgen: generates Rust struct defined inside CUDA kernels written in C++
cudarc: safe abstraction for interop with GPU driver

Structure

.
├── Cargo.lock
├── Cargo.toml
├── build.rs
├── docs
│   └── design.md
└── src
    ├── cuda
    │   ├── includes
    │   │   ├── barretenberg
    │   │   ├── prime_field.h
    │   │   ├── test.cpp
    │   │   └── wrapper.h
    │   └── kernels
    │       ├── multilinear.cu
    │       ├── scalar_multiplication.cu
    │       └── sumcheck.cu
    ├── field.rs
    └── lib.rs

build.rs contains the build script for building the whole project
src/cuda/includes/barretenberg/ contains barretenberg lib (https://github.com/AztecProtocol/barretenberg/tree/master) for field implementation in C++
src/cuda/includes/prime_field.h declares FieldBinding struct in C representing barretenberg field struct. This wrapping is due to the lack of support for C++ language of bindgen crate
src/cuda/includes/wrapper.h is interface file for bindgen to generate Rust struct from C struct
In src/cuda/includes/kernel/, multiple kernel functions are defined
In field.rs, conversion between FieldBinding struct and halo2curves' field struct is implemented
In lib.rs, expose APIs to test CUDA kernels via GPUApiWrapper struct