Incremental Auto-Tuning for Hybrid Parallelization Using OpenCL

Some recent processors incorporate both a GPU and multiple processing cores, and hybrid parallelism, in which parallel processing based on GPGPU on the GPU and parallel processing based on multi-threading on the CPU are performed simultaneously, is now available. ratio of CPU and GPU differs depending on the application, and it is difficult to determine the optimal load distribution in advance. The authors previously proposed an on-the-fly auto-tuning method to determine the optimal load distribution at runtime, but it was necessary to determine in advance the ratio of the computational range to be used for the fallback execution (AutoRatio). We proposed an on-the-fly auto-tuning method that performs incremental preruns at runtime without determining the AutoRatio, and confirmed the effectiveness of the method for four different applications.