high performance computing on graphics processing units: hgpu.org

An important class of compute accelerators are graphics processing units (GPUs). Popular programming models for non-graphics computation on GPUs, such as CUDA and OpenCL, provide an abstraction of many parallel scalar threads. Contemporary GPU hardware groups 32 to 64 scalar threads as a single warp or wavefront and executes this group of scalar threads in lockstep. The inherent mismatch between scalar programming model and vector hardware creates a challenge when developing applications that employ synchronization on the GPU. This challenge arises from the use of a hardware stack to manage control flow divergence among scalar threads. This thesis explains the porting of the Apriori benchmark to a GPU which led to the research on synchronization in SIMT hardware. It then proposes instruction set and hardware changes that simplify the implementation of mutual exclusion when porting multiple-instruction, multiple data (MIMD) programs with synchronization to accelerators employing single-instruction, multiple thread (SIMT) hardware. These instructions when compared with more complex software only solutions, achieve similar performance. This thesis also implements and evaluates queue based mutual exclusion on SIMT hardware.