Multiversioned decoupled access-execute: the key to energy-efficient compilation of general-purpose programs

Abstract

Computer architecture design faces an era of great challenges in an attempt to simultaneously improve performance and energy efficiency. Previous hardware techniques for energy management become severely limited, and thus, compilers play an essential role in matching the software to the more restricted hardware capabilities. One promising approach is software decoupled access-execute (DAE), in which the compiler transforms the code into coarsegrain phases that are well-matched to the Dynamic Voltage and Frequency Scaling (DVFS) capabilities of the hardware. While this method is proved efficient for statically analyzable codes, generalpurpose applications pose significant challenges due to pointer aliasing, complex control flow and unknown runtime events. We propose a universal compile-time method to decouple generalpurpose applications, using simple but efficient heuristics. Our solutions overcome the challenges of complex code and show that automatic decoupled execution significantly reduces the energy expenditure of irregular or memory-bound applications and even yields slight performance boosts. Overall, our technique achieves over 20% on average energy-delay-product (EDP) improvements (energy over 15% and performance over 5%) across 14 benchmarks from SPEC CPU 2006 and Parboil benchmark suites, with peak EDP improvements surpassing 70%.

Clairvoyance: Look-ahead compile-time scheduling

“…Clairvoyance builds upon techniques such as software pipelining [9,10], program slicing [11], and decoupled access-execute [12][13][14] and generates code that exhibits improved memory-level parallelism (MLP) and instructionlevel parallelism (ILP). For this, Clairvoyance prioritizes the execution of critical instructions, namely loads, and identifies independent instructions that can be interleaved between loads and their uses.…”

“…Solution: We develop a metric, called indirection count, based on the number of memory accesses required to compute the memory address (indirections) [14] and the number of memory accesses required to reach the load. For example,…”

“…We compare our techniques to Software Decoupled AccessExecute (DAE) [13,14] and the LLVM standard instruction schedulers list-ilp (prioritizes ILP), list-burr (prioritizes register pressure) and list-hybrid (balances ILP and register pressure). DAE reduces the energy consumption by creating a duplicated loop that prefetches data ahead of time, while running at low frequency and maintaining its original performance.…”

“…Clairvoyance attempts to fully hide memory latency with independent instructions (ILP) and to cluster memory operations together and increase MLP by decoupling the loop. Software Decoupled Access-Execute (DAE) [13,14] targets reducing energy expenditure using DVFS, while maintaining performance, whereas Clairvoyance focuses on increasing performance. DAE generates Access-Execute phases that merely prefetch data and duplicate a significant part of the original loop (control instructions and address computation).…”

“…Contrary to Clairvoyance, CFD requires hardware support to ensure low-overhead communication between the decoupled phases. A software only version, Data-flow Decoupling (DFD), relies on prefetch instructions and ensures communication between phases by means of caches, akin to DAE [13,14], using code duplication. As the CFD solution is not entirely automatic and requires manual intervention, Clairvoyance provides the missing compiler support and is readily applicable to decouple the CFG and hoist branch-evaluation, in lieu of long latency loads.…”