An open-source GPU library for accelerating Zero-Knowledge-Proof. Panda is an open GPU computing framework that absorbs excellent optimization methods, continuously improves, and optimizes. We welcome more people to participate in the development together.

• Supports multiple curves (BN254/BLS12-377 and more).
• Supports coordinate conversion, output results can be Jacobian coordinates or Projective coordinates.
• Supports msm and ntt module acceleration.
• Supports Rust API calls for msm and ntt acceleration.
• Supports cached bases and scalars.
• Good compatibility and easy to extend.

• src
  • cuda: All CUDA source code and test code.
  • gpu_ffi：FFI interface code.
  • gpu_manager：Panda GPU manager code, mainly used to manage GPUs and external interfaces.
• test

Responsible for initializing hardware, creating memory pools, initializing msm modules, preloading cached bases or scalars, etc.For example: init_hardware, init_msm, etc.

Configure parameters, get parameters and status, etc.For example: get_device_number, device_info, etc.

Call different execution functions to accelerate msm and ntt modules.For example: panda_msm_bn254_gpu

• If you want to use Rust API, you can refer to the ./tests/test.rs sample code.
• Currently, the msm CUDA code calculation process uses Jacobian coordinates, so if you expect the output coordinates to be Projective coordinates, you can set the return type conversion.
• The CUDA code framework supports multiple curves, and currently encapsulates the Rust API for BN254, and it will be easy to encapsulate the Rust API for BLS12-377 later.
• We also provide host code for debugging purposes.

• Random GPU test:

cargo test test_msm_bn254_correctness_device --release -- --nocapture

• cd ./src/cuda && ./build.sh && cd build && ./test/panda_test gpu
• We temporarily provide a set of local data.