This repository implements a GPU accelerator written in futhark for Neptune's Triton-VM. The GPU kernels are accessible via rust bindings that can be directly incorporated into the triton-vm prover.

The functionalities implemented include:

• Low degree extension for the MasterBaseTable
• Low degree extension for the MasterExtTable
• Merkle building
• Tip5 hash function

NOTE: I left this project somewhat abruptly due to an opportunity that came my way, so it’s in a bit of a hacky, but functional, state at the moment. Apologies in advance, the setup is a bit involved and would definitely benefit from some refactoring.

First run ./setup.sh in the root dir. This will do two things. The first is that it will clone a fork of the triton-vm with some slight modifications necessary for testing (it also runs some code gen on the triton-vm repo). Second, It will run the rust-generator crate, which is a member of the ruthark workspace. This will perform rust code generation that will create rust bindings for the futhark GPU kernels located in the fut-src directory. The result will be a crate called gpu-accelerator that contains those bindings.

Also note that by default the accelerator backend in rust-generator/codegen.rs is set to OpenCL, but you can replace this with Cuda or C depending on your hardware.

fn main() {

    genfut(Options {

        name: GENERATED_RUST_MODULE_NAME.to_string(),

        file: std::path::PathBuf::from(FUTHARK_SOURCE_FILE),

        author: "Name <[email protected]>".to_string(),

        version: "0.1.0".to_string(),

        license: "NONE".to_string(),

        description: "Futhark accelerator for Neptune".to_string(),

        backend: Backend::OpenCL, //  <------ or Backend::Cuda or Backend::C

    })

}

Once you have the gpu-accelerators crate generated, you need to modify the Cargo.toml file by changing this line:

members = ["rust-generator"]

To this:

members  = ["triton-vm", "gpu-accelerator", "accelerators", "rust-generator"]

Then you should be able to access the GPU kernels from the accelerators crate, which implements an API that can be incorporated into triton-vm.

Within gpu-accelerators there is a struct called GpuParallel, which implements methods that can directly replace functions in the triton-vm prover. You should be able to run the following functionalities on the GPU:

• low degree extension for the master base table

• low degree extension for the master extension table

• Merkle building

• Tip5

If all goes right, all tests should pass when you run cargo test in the accelerators directory.

Submit an issue or contact me on telegram @huowli