Libra: Achieving Efficient Instruction- and Data- Parallel Execution for Mobile Applications.

Abstract

Mobile computing as exemplified by the smart phone has become an integral part of our daily lives. The next generation of these devices will be driven by providing richer user experiences and compelling capabilities: higher definition multimedia, 3D graphics, augmented reality, and voice interfaces. To meet these goals, the core computing capabilities of the smart phone must be scaled. But, the energy budgets are increasing at a much lower rate, thus fundamental improvements in computing efficiency must be garnered. To meet this challenge, computer architects employ hardware accelerators in the form of SIMD and VLIW. Single-instruction multiple-data (SIMD) accelerators provide high degrees of scalability for applications rich in data-level parallelism (DLP). Very long instruction word (VLIW) accelerators provide moderate scalability for applications with high degrees of instruction-level parallelism (ILP). Unfortunately, applications are not so nicely partitioned into two groups: many applications have some DLP, but also contain significant fractions of code with low trip count loops, complex control/data dependences, or non-uniform execution behavior for which no DLP exists. Therefore, a more adaptive accelerator is required to be able to deploy resources as needed: exploit DLP on SIMD when it’s available, but fall back to ILP on the same hardware when necessary. In this thesis, we first focus on various compiler solutions that solve inefficiency problem in both VLIW and SIMD accelerators. For SIMD accelerators, a new vectorization pass, called SIMD Defragmenter, is introduced to uncover hidden DLP using subgraph identification in SIMD accelerators. CGRA express effectively accelerates sequential code regions using a bypass network in VLIW accelerators, and Resource Recycling leverages stream-graph modulo scheduling technique for scheduling of multiple code regions in multi-core accelerators. Second, we propose the new scalable multicore accelerator referred to as Libra for mobile systems, which can support execution of code regions having both DLP and ILP, as well as hybrid combinations of the two. We believe that as industry requires higher performance, the proposed flexible accelerator and compiler support will put more resources to work in order to meet the performance and power efficiency requirements.