Dynamic SIMD Re-convergence with Paired-Path Comparison

Abstract

SIMD divergence is one of the critical causes that decrease the parallel computing efficiency in contemporary GPGPU (General Purpose Graphic Processor Unit) architecture In this thesis we evaluate a cycle accurate GPU simulator platform based on HSAIL under OpenCL framework by offloading the kernel programs into simulator A wavefront (“wavefront” and “warp” in AMD and NVIDIA terminology respectively) is the gathering of multiple threads that execute the same instruction in SIMD fashion When a wavefront or a warp executes a conditional branch instruction threads in the warp may go to distinct PCs if the threads have different branch targets and it’s called SIMD control divergence Re-convergence mechanisms are applied to help divergent wavefront to execute instructions properly We develop a new dynamic stack-based re-convergence scheme that can be implemented with or without finalizer generated re-convergence instructions Using the scheme we propose the divergent warp re-converges dynamically and get a 13 36% activity factor improvement on average from opportunistic early re-convergence in the unstructured control flow and the performance is better in the way that warp re-convergence without finalier generated hint instructions

Date of Award	2015 Aug 14
Original language	English
Supervisor	Chung-Ho Chen (Supervisor)