Vector Runahead for Indirect Memory Accesses

Abstract

Vector runahead delivers extremely high memory-level parallelism even for the chains of dependent memory accesses with complex intermediate address computation, which conventional runahead techniques fundamentally cannot handle and, therefore, have ignored. It does this by rearchitecting runahead to use speculative data-level parallelism, rather than work skipping, as its primary form of extracting more memory-level parallelism in runahead mode than a true execution can, which we hope will bring about an entirely new dimension for high-performance processors.