Automatic Static Analysis-Guided Optimization of CUDA Kernels

We propose a framework for using static resource analysis to guide the automatic optimization of general-purpose GPU (GPGPU) kernels written in CUDA, NVIDIA's framework for GPGPU programming. In our proposed framework, optimizations are applied to the kernel and candidate kernels are evaluated for performance by running a static analysis that predicts the execution cost of GPU kernels. The use of static analysis, in contrast to many existing frameworks for performance tuning GPU kernels, lends itself to high-level, hardware-independent optimizations that can be of particular benefit to novice programmers unfamiliar with CUDA's performance pitfalls. As a proof of concept, we have implemented two example optimizations and a simple search strategy in a tool called COpPER (CUDA Optimization through Programmatic Estimation of Resources), which makes use of a static resource analysis tool for CUDA from prior work. The prototype tool automatically improves the performance of sample kernels by 2-4% in initial experiments, and demonstrates the feasibility of using static analysis as part of automated performance tuning for GPU kernels.