Implementation of HyperPianist: Pianist with Linear-Time Prover and Logarithmic Communication Cost (SP2025). This is an optimized, fully distributed version of HyperPlonk.
This is the PIOP version. For the layered circuit version, check out the other branch.
The entire codebase is written with haste (to meet conference deadlines) and performance in mind. As such, there is A LOT of code duplication and ad-hoc optimizations. It probably has 3x as many lines as it should. Please do not look for readability or security while going through our code.
Requires Rust Nightly. For optimal performance, please build as --release, with
RUSTFLAGS='-C target-cpu=native -C target-feature=+bmi2,+adx'
Most of our code has distributed and non-distributed versions. There are a number of non-distributed tests that can be run as usual. For distributed tests (in the dTests folder of each project), you may run them locally with the run.sh script. (This will spawn 4 sub-provers on the same machine.) Example (in subroutines):
./dTests/run.sh dSumcheck
The main benchmark is hyperpianist/dTests/bench.rs.
You need to have a list of IPs of sub-provers, on each sub-prover.
Each sub-prover should have the same list. Please make sure to list LAN IPs in the files.
Then, using a script or whatever else, arrange for all sub-provers to run, simultaneously
<executable> <path to IP list> <index of self> <nv> [--dory] [--jellyfish]
<index of self> must correspond to the position of the sub-prover in the IP list. <nv> is number of variables to test (e.g. 22 corresponds to a circuit of 2^22 constraints).
Flags --dory and --jellyfish change the PCS and gate type, respectively. (When not provided, defaults to Mkzg, and vanilla gates.)