Towards enhancing parallelism by exploiting memory hierarchy in GPUs

Abstract

newline Graphics Processing Units (GPUs) are highly parallel, multithreaded processors consisting of thousands of cores designed for handling parallelism in the form of Single-Instruction Multiple-Data (SIMD) as well as Single- Program Multiple-Data (SPMD). These GPUs, in the recent years, have evolved in terms of architectural features, supporting programming models, to solve many complex computational problems from high performance computing, which are highly compute intensive and embarrassingly parallel. Since GPUs are efficient enough to perform non-specialized calculations they newlineare also used as General Purpose Graphical Processing Units (GPGPUs). Though the architectural specializations have enhanced many applications, there are classes of HPC problems that may not be as amenable to this special class of GPU hardware as a natural choice and it becomes challenging to solve them. This is due to dependences among computations or insufficient degree of parallelism in the applications. This parallelism has to be unveiled with explicit modifications in the algorithms or implicit transformations using compiler optimizations. newline newline