| dc.description.abstract | Recent advancements in cloud computing, data privacy, and cryptography have sparked a growing interest in Verifiable Computation (VC) in both industry and academia. In particular, zero-knowledge proof (ZKP) algorithms are gaining rapid traction due to their strong privacy guarantees. However, they are notoriously computationally intensive, making performance a critical concern. Given the inherent data parallelism and heavy use of vector operations in ZKP computations, multicore CPUs and GPUs offer a promising acceleration path. Unfortunately, accelerated programming for ZKP remains challenging: ZKP algorithms evolve rapidly, their structures grow increasingly complex, and writing high-performance ZKP code is tedious, error-prone, non-portable, and unfriendly to algorithm developers. We present an end-to-end compiler framework, Zera, that lowers ZKP algorithms to parallel hardware for efficient acceleration, with minimal programmer effort. By effectively leveraging ZKP algorithm patterns and trends, we are able to automate the key performance optimizations, with a succinct linguistic extension and a set of practical compiler customizations. Consequently, with just 92 lines of trivial high-level annotation added to the original 7,000 lines of C++ code, our single-source code solution delivers 33.9× and 24.0× speedup on GPU over a highly optimized serial C++ implementation on CPU and an existing multithreaded Rust baseline on CPU, respectively. Compared to our hand-optimized GPU/CUDA implementation requiring an extra 2,000 lines of low-level code (roughly 60 programmer hours), our compiler-generated GPU implementation is only 58% slower (1.58× slowdown) on large inputs, demonstrating a compelling trade-off between performance and productivity. | |
