Evaluating CUDA-Aware Approximate Computing Techniques

Abstract

Approximate computing techniques offer performance improvements by performing inexact computations. Moreover, CUDA programs written to be executed on GPU devices employ specific features to utilize the parallel computation units of heterogeneous GPU architectures. While generic software-level approximate computing techniques have been applied to heterogeneous CUDA programs, CUDA-specific approaches may introduce promising performance improvements by not corrupting the target computations. In this work, we propose software approximation techniques for CUDA programs: kernel-aware loop perforation, partition-level synchronization, block-level atomic operations, and warp divergence elimination. We perform source code transformations on target benchmark programs by applying our techniques. We evaluate performance improvements by trading off accuracy in our target computations. Our experimental results reveal that CUDA-aware approximation techniques offer significant performance improvements at the expense of acceptable accuracy loss. © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).