ABSTRACTIONS FOR APPROXIMATE AND ENERGY-EFFICIENT COMPUTING ON MODERN SIMD/SIMT ARCHITECTURES

THE INCREASING DEMAND FOR PARALLELISM IN MODERN APPLICATIONS HAS EXPOSED THE LIMITATIONS OF HOMOGENEOUS SYSTEMS, DRIVING THE ADOPTION OF HETEROGENEOUS COMPUTING ARCHITECTURES. IN RESPONSE, A DIVERSE ECOSYSTEM OF PROGRAMMING MODELS HAS EMERGED, FROM LOW-LEVEL VENDOR-SPECIFIC APIS TO HIGH-LEVEL PORTABLE ABSTRACTIONS, EACH AIMING TO BALANCE DEVELOPER PRODUCTIVITY AND EFFICIENT HARDWARE UTILIZATION ACROSS DIFFERENT ARCHITECTURES. WHILE PRIOR RESEARCH HAS EVALUATED LOW-LEVEL PROGRAMMING MODELS, THERE IS A LACK OF DETAILED ANALYSIS OF HIGH-LEVEL PROGRAMMING MODELS AND THEIR ABILITY TO ACHIEVE COMPETITIVE PERFORMANCE, AS WELL AS A LACK OF DOMAIN-SPECIFIC ABSTRACTIONS THAT EXPOSE APPROXIMATE AND ENERGY-EFFICIENT COMPUTING TECHNIQUES TO DEVELOPERS. THIS THESIS ADDRESSES THESE GAPS ACROSS THE PROGRAMMING-MODEL LANDSCAPE THROUGH THREE MAIN CONTRIBUTIONS: THE ANALYSIS OF HIGH- AND LOW-LEVEL ABSTRACTIONS FOR SIMT ARCHITECTURES, THE DESIGN OF DOMAIN-SPECIFIC ABSTRACTIONS AND NOVEL TECHNIQUES FOR APPROXIMATE AND ENERGY-EFFICIENT COMPUTING, AND THE EXPLORATION OF COMPILER APPROACHES FOR AUTOMATICALLY GENERATING VECTORIZED CODE ON SIMD ARCHITECTURES. FIRST, WE PROVIDE A COMPREHENSIVE EVALUATION OF PROGRAMMING MODELS, ASSESSING HOW HIGH-LEVEL CONSTRUCTS MAP EFFICIENTLY TO GPUS AND APPROACH THE PERFORMANCE OF NATIVE LOW-LEVEL APIS. SECOND, WE EXTEND HIGH-LEVEL PROGRAMMING MODELS WITH DOMAIN-SPECIFIC ABSTRACTIONS AND TECHNIQUES FOR APPROXIMATE AND ENERGY-EFFICIENT COMPUTING, ENABLING DEVELOPERS TO EXPLOIT THESE APPROACHES WITHOUT SPECIALIZED HARDWARE KNOWLEDGE. LASTLY, WE INVESTIGATE THE AUTOVECTORIZATION CAPABILITIES OF MODERN COMPILERS FOR RISC-V VECTOR ARCHITECTURES, PROVIDING INSIGHTS INTO COMPILER EFFECTIVENESS, VECTORIZATION COVERAGE, AND PERFORMANCE OPTIMIZATION OPPORTUNITIES. OVERALL, THIS THESIS CONTRIBUTES TO IMPROVE THE PERFORMANCE OF HIGH-LEVEL PROGRAMMING ABSTRACTIONS FOR SIMD AND SIMT ARCHITECTURES, AS WELL AS THE DESIGN OF NEW PROGRAMMING-MODEL ABSTRACTIONS AND TECHNIQUES FOR APPROXIMATE AND ENERGY-EFFICIENT COMPUTING ON HETEROGENEOUS ARCHITECTURES.