| Coarse-grained reconfigurable Architecture(CGRA)has been recognized as a kind of domain-specific processor architecture featured by high efficiency and modest flexibility at the same time.Because of this,CGRA in some way fills the gap between ASIC and general purpose processor.However,due to the data-driven nature,CGRA is low-efficient when dealing with control-flow which widely exits in various kinds of application.Although State-based full predication(SFP)is a kind of prevailing solution,it still suffers from performance degradation resulted from redundant operations aiming at state register modification.TFP(Tag-based Full Predication)strategy,which expands the tag-based waking mechanism introduced in PSFP(Pseudo branch SFP)to operation nullification,is introduced in this thesis.TFP strategy nullifies operations based on the inconsistence between the tag register and the tag field in configuration information,which differs from state-based nullification mechanism introduced by SFP strategies.Furthermore,TFP exploits parallelism between tag register modification and any of the common operations,in which way redundant operations in SFP can be eliminated.Power optimization is also considered in this thesis by applying clock-gating technique to reducing dynamic power consumption.Hardware support for TFP is present,and the verification is based on FPGA hardware platform and Vivado environment.Traditional PSFP and counter-based SFP(CSFP)is also realized in form of RTL model for comparison.Target applications are loops containing ITE and NITE structures,extracted from SPEC CPU2006 and Mibench benchmark suite.Functional correctness is verified by comparing the C program tracing results and RTL simulation results.On top of this,performance and power estimation are conducted among three strategies.Experimental results show that compared with traditional PSFP and CSFP strategies,TFP achieves over 30% performance improvement at the power cost lower than 10%.The elimination of redundant operations brings about obvious improvement to performance. |
