Libra Tailoring SIMD Execution using Heterogeneous Hardware and

Libra: Tailoring SIMD Execution using Heterogeneous Hardware and Dynamic Configurability Yongjun Park 1, Jason Jong Kyu Park 1 , Hyunchul Park 2, 1 and Scott Mahlke December 3, 2012 1 University of Michigan, Ann Arbor 2 Programmin Systems Lab, Intel Labs, Santa Clara, CA 1 University of Michigan Electrical Engineering and Computer Science

Current mobile Solutions & ILP-based Challenges DLPbased ULP Ge. Force Adreno 320 ARM Mali-400 MP 4 1. 6 GHz ARM Cortex-A 9 1. 7 GHz Krait 1. 6 GHz ARM Cortex-A 9 Good for ILP Good for DLP legacy workloads media processing web browsing scientific computing wireless communication Image processing Mixture of ILP/DLP Goal: Design of a flexible & energy efficient unified accelerator 2 University of Michigan Electrical Engineering and Computer Science

Libra: Loop-adaptive SIMD Accelerator Application Low/No-DLP loops Ex. Op-intensive loops • • Heterogeneous Traditional SIMD High-DLP loops 0 1 2 3 4 5 6 7 0 1 2 3 Int Int Expensive unit Expensive unit Heterogeneous lane structure: less power/area Dynamic configurability: change ILP/DLP capability 3 University of Michigan Electrical Engineering and Computer Science

Homogeneous SIMD vs. Heterogeneous Libra • Control power overheads can be minimized by Functional Units power savings. 4 University of Michigan Electrical Engineering and Computer Science