Research Projects

The use of heterogeneous systems, which include multiple processors like CPUs, GPUs, and FPGAs, has become increasingly popular in recent years. Collaborative computation between CPUs and GPUs is crucial for high program performance, and a CPU- or GPU-only optimization technique cannot fully exploit this potential. This work proposes new global compiler optimizations that allow CPUs and GPUs to work cooperatively to achieve high program performance in heterogeneous systems. The research addresses the underutilization of processors in heterogeneous systems due to non-overlapping computations. The authors propose a new optimization technique called hetero-sync motion, which can relocate barrier instructions to new locations, resulting in better CPU utilization in heterogeneous systems. They also propose GSOHC, a compiler analysis and optimization framework that automatically finds opportunities for hetero-sync motion in the input program and performs code transformation to apply the optimization. The research also addresses limitations with existing compiler optimizations in heterogeneous programs, proposing efficient inter-procedural data flow analyses for CPU-GPU heterogeneous computing that take into account data flow across CPU and GPU codes. If the objectives are achieved, this research leads to simple but effective static analyses for various global optimizations in CPU-GPU heterogeneous programs, achieving good performance gains with low compiler overheads, making them practical for production compilers.