GPU-accelerated ECDSA library for the SM2 curve.

• NVIDIA CUDA Toolkit
• GMP (GNU Multiple Precision Library)

1. Build with CMake

mkdir build && cd build
cmake -DCMAKE_CUDA_ARCHITECTURES=70 ..
make

2. Build manually

mkdir -p bin
nvcc -O3 -gencode arch=compute_70,code=sm_70 -o bin/gsv.o -c gsv.cu
g++ -O3 -I/usr/local/cuda-11/include -o bin/main.o -c main.cpp
g++ -O3 -o bin/main bin/gsv.o bin/main.o -L/usr/local/cuda-11/lib64 -lcuda -lcudart -lgmp

docker build -t rapidec:v1 .
docker run -it --gpus all rapidec:v1

#include "RapidSV/gsv_wrapper.h"
void GSV_init(int device_id = 0);
void GSV_verify(int count, sig_t *sig, int *results);
void GSV_close();

GSV_init() initializes the GPU device, allocates the GPU memory pool, and copies the precomputed table to constant memory.

GSV_verify() does the signature verification.

• The first parameter is the number of signatures.
• The second parameter is the array of signatures. Each signature is a sig_t struct that contains 5 big integers, namely signature (r, s), message digest e, and public key (key_x, key_y). Each big integer is represented by an array of unsigned 32-bit integers.
• The third parameter is the array of verification results. Correct verification returns 0, otherwise 1.

GSV_close() frees GPU memory.

Configurable flags in gsv_wrapper.h:

• GSV_TPI: Threads per instance, a parameter of the CGBN library. Can be set to 4, 8, 16, or 32. Default value is 16.
• GSV_256BIT: Enable this flag to use 256-bit integers for calculation rather than 512-bit integers. Will save space but a bit slower.
• GSV_KNOWN_PKEY: When the public key of a batch of signatures are the same, use a precomputed table to speed up verification. Need to generate different sm2_pkey_512.table for different keys.

This project used the CGBN library.