Decoupled Vector Runahead for Prefetching Nested Memory-Access Chains
Ajeya Naithani, Jaime Roelandts, Sam Ainsworth, Timothy M. Jones, Lieven Eeckhout · IEEE Micro · 2024
Decoupled Vector Runahead (DVR) exploits massive amounts of memory-level parallelism to improve the performance of applications featuring indirect memory accesses by dynamically inferring loop bounds at run-time, recognizing striding loads, and speculatively vectorizing subsequent instructions that are part of an indirect chain. DVR runs as an on-demand, speculative, in-order, lightweight hardware subthread alongside the main thread within the core. DVR incurs a minimal hardware overhead while delivering a substantial performance boost.