Compilation Optimization of DCU-oriented OpenMP Thread Scheduling

Abstract OpenMP is one of the mainstream parallel programming models in recent years. After version 4.0, introduced a new target instruction to increase functionality heterogeneous programming, called Offload. For domestic platform DCU, thread scheduling algorithm under computing has low performance default mode, which does not take best advantage GPU and wasted resources. To address this problem, paper performs improvement at compiler level, analyzes available resources system by combining DCU hardware facilities, then further parses program based on its array information get iteration number, reallocates number threads for different execution modes OpenMP, optimizes group factor adjust thread. This uses SPEC ACCEL test set optimize DCU. In paper, we use Polybench standard redistribution groups two using optimization algorithm. The average speedup ratio was improved 40%.