Groth16 prover

This directory contains a reference CPU and GPU implementation of the Groth16 prover using libsnark.

There are two provers implemented in this repository. The first is libsnark/main.cpp, which has the CPU libsnark reference prover that we compare benchmarks against. The second is libsnark/opencl_main.cpp.

The GPU prover implements a custom DIT/DIF FFT and Pippenger Multiexponentiation algorithms in OpenCL. The Kernel is currently unoptimized and this is a WIP.

NOTE: Due to an AMD compiler bug we are unable to use this kernel on AMD hardware. We will continue to explore our options with Kronos Group to open this implementation up to both hardware vendors.

Dependencies

The code should compile and run on Ubuntu 18.04 with the following dependencies installed:

sudo apt-get install -y build-essential \
    cmake \
    git \
    libomp-dev \
    libgmp3-dev \
    libprocps-dev \
    python-markdown \
    libboost-all-dev \
    libssl-dev \
    pkg-config \
    nvidia-cuda-toolkit

Building on MacOS is not recommended as CUDA support is harder to use. (Apple mostly ships with AMD.)

Build

./build.sh

Generate parameters and inputs

./generate_parameters

When iterating on your implementation for correctness (and not performance) smaller constraint systems are fine. In that case, ./generate_parameters fast will give you smaller parameters that don't take as long to generate or prove with.

Run

./main MNT4753 compute MNT4753-parameters MNT4753-input MNT4753-output
./main MNT6753 compute MNT6753-parameters MNT6753-input MNT6753-output
./opencl_main MNT4753 compute MNT4753-parameters MNT4753-input MNT4753-output

Check results

sha256sum MNT4753-output MNT6753-output MNT4753-output_cuda MNT6753-output_cuda

‡ Due to some bug with nvcc in version 10.1, including any of the mnt4/mnt6 libff headers (a library libsnark uses for finite field arithmetic) leads to a compilation failure. The wrapper library uses something like the pImpl idiom to hide the need for those headers from the caller.