A Similarity Measure for GPU Kernel Subgraph Matching

TL;DR

This paper introduces CUDAflow, a tool that analyzes CUDA kernel control flow graphs to measure similarity and extract insights, aiding optimization and understanding of GPU kernel behavior.

Contribution

CUDAflow provides a novel static and dynamic analysis method for CUDA binaries, enabling subgraph matching of CFGs to characterize kernel resource requirements.

Findings

Reveals new thread divergence patterns in GPU kernels

Demonstrates effectiveness on SHOC and Rodinia benchmarks

Aids in autotuning and compiler optimization

Abstract

Accelerator architectures specialize in executing SIMD (single instruction, multiple data) in lockstep. Because the majority of CUDA applications are parallelized loops, control flow information can provide an in-depth characterization of a kernel. CUDAflow is a tool that statically separates CUDA binaries into basic block regions and dynamically measures instruction and basic block frequencies. CUDAflow captures this information in a control flow graph (CFG) and performs subgraph matching across various kernel's CFGs to gain insights to an application's resource requirements, based on the shape and traversal of the graph, instruction operations executed and registers allocated, among other information. The utility of CUDAflow is demonstrated with SHOC and Rodinia application case studies on a variety of GPU architectures, revealing novel thread divergence characteristics that facilitates end users, autotuners and compilers in generating high performing code.