Awesome

bellperson

This is a fork of the great bellman library.

bellman is a crate for building zk-SNARK circuits. It provides circuit traits and primitive structures, as well as basic gadget implementations such as booleans and number abstractions.

Backend

There is currently one backend available for the implementation of Bls12 381:

• blstrs - optimized with hand tuned assembly, using blst

GPU

This fork contains GPU parallel acceleration to the FFT and Multiexponentation algorithms in the groth16 prover codebase under the compilation features cuda and opencl.

Requirements

• NVIDIA (Turing or newer) or
• AMD GPU Graphics Driver (OpenCL)

( For AMD devices we recommend ROCm )

Environment variables

The gpu extension contains some env vars that may be set externally to this library.

• BELLMAN_NO_GPU

  Will disable the GPU feature from the library and force usage of the CPU.

  // Example
  env::set_var("BELLMAN_NO_GPU", "1");
  

• BELLMAN_VERIFIER

  Chooses the device in which the batched verifier is going to run. Can be cpu, gpu or auto.

  Example
  env::set_var("BELLMAN_VERIFIER", "gpu");
  

• RUST_GPU_TOOLS_CUSTOM_GPU

  Will allow for adding a GPU not in the tested list. This requires researching the name of the GPU device and the number of cores in the format ["name:cores"].

  // Example
  env::set_var("RUST_GPU_TOOLS_CUSTOM_GPU", "GeForce RTX 2080 Ti:4352, GeForce GTX 1060:1280");
  

• BELLMAN_CPU_UTILIZATION

  Can be set in the interval [0,1] to designate a proportion of the multiexponenation calculation to be moved to cpu in parallel to the GPU to keep all hardware occupied.

  // Example
  env::set_var("BELLMAN_CPU_UTILIZATION", "0.5");
  

• RAYON_NUM_THREADS

  Restricts the number of threads used in the library to roughly that number (best effort). In the past this was done using BELLMAN_NUM_CPUS which is now deprecated. The default is set to the number of logical cores reported on the machine.

  // Example
  env::set_var("RAYON_NUM_THREADS", "6");
  

• EC_GPU_NUM_THREADS

  Restricts the number of threads used by the FFT and multiexponentiation calculations. In the past this setting was shared with RAYON_NUM_THREADS, now they are separate settings that can be controlled independently. The default is set to the number of logical cores reported on the machine.

  // Example
  env::set_var("EC_GPU_NUM_THREADS", "6");
  

• BELLMAN_GPU_FRAMEWORK

  Bellman can be compiled with both, OpenCL and CUDA support. When both are available, BELLMAN_GPU_FRAMEWORK can be used to set it to a specific one, either cuda or opencl.

  // Example
  env::set_var("BELLMAN_GPU_FRAMEWORK", "opencl");
  

• BELLMAN_CUDA_NVCC_ARGS

  By default the CUDA kernel is compiled for several architectures, which may take a long time. BELLMAN_CUDA_NVCC_ARGS can be used to override those arguments. The input and output file will still be automatically set.

  // Example for compiling the kernel for only the Turing architecture
  env::set_var("BELLMAN_CUDA_NVCC_ARGS", "--fatbin --gpu-architecture=sm_75 --generate-code=arch=compute_75,code=sm_75");
  

• BELLPERSON_GPUS_PER_LOCK

  Restricts the number of devices used by the FFT and multiexponentiation calculations.

  • If it's not set, a single lock will be created, and each calculation uses all devices
  • If BELLPERSON_GPUS_PER_LOCK = 0, no lock will be created, each calculation uses all devices, and each device can run multiple calculations. WARNING: this option can break things easily. Each kernel expects that it's run without anything else running on the GPU at the same time. If two kernels run at the same time, they might interfere with each other and lead to crashes or wrong results.
  • If BELLPERSON_GPUS_PER_LOCK > 0, create a lock for each device, each calculation uses BELLPERSON_GPUS_PER_LOCK (up to device number) devices

  // Example
  env::set_var("BELLPERSON_GPUS_PER_LOCK", "0");
  env::set_var("BELLPERSON_GPUS_PER_LOCK", "1");
  

Supported / Tested Cards

Depending on the size of the proof being passed to the gpu for work, certain cards will not be able to allocate enough memory to either the FFT or Multiexp kernel. Below are a list of devices that work for small sets. In the future we will add the cuttoff point at which a given card will not be able to allocate enough memory to utilize the GPU.

Running Tests

RUSTFLAGS="-C target-cpu=native" cargo test --release --all

To run using CUDA and OpenCL, you can use:

RUSTFLAGS="-C target-cpu=native" cargo test --release --all --features cuda,opencl

To run the multiexp_consistency test you can use:

RUST_LOG=info cargo test --features cuda,opencl -- --exact multiexp::gpu_multiexp_consistency --nocapture

Considerations

Bellperson uses rust-gpu-tools as its CUDA/OpenCL backend, therefore you may see a directory named ~/.rust-gpu-tools in your home folder, which contains the compiled binaries of OpenCL kernels used in this repository.

Experimental

The instance aggregation provided by groth16::aggregate::prove::aggregate_proofs_and_instances() has not yet been audited so should be used with caution. It is not recommended to use instance aggregation in production until it has been audited.

License

Licensed under either of

• Apache License, Version 2.0, |LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.