Enjeux de conception des architectures GPGPU : unités arithmétiques spécialisées et exploitation de la régularité
Caroline Collange · HAL (Le Centre pour la Communication Scientifique Directe) · 2010
Current Graphics Processing Units (GPUs) are high-performance, low-cost parallel processors. This makes them attractive even for non-graphics calculations, opening the field of General-Purpose computation on GPUs (GPGPU). In this thesis, we develop software techniques to take advantage of the fixed-function units that GPU provide for scientific computations, and consider hardware modifications to execute general-purpose applications more efficiently. More specifically, we identify parallel regularity as an opportunity for improving the efficiency of parallel architectures. We expose its potential through the simulation of an actual GPU architecture. Subsequently, we consider two alternatives to take advantage of regularity. First, we design a dynamic hardware scheme which improves the energy efficiency of datapaths and register files. Then, we present a static compiler analysis to reduce the complexity associated with instruction control on GPUs.