Researched how the multi-pipeline processor accelerates the running of thread ,found that when the branch predictor facing the random branch instruction, the hit rate will become very low, so bring out a new method that using the free pipeline to accelerate the running of branch instruction. If the right prediction from branch predictor is less than 70% and there is a free pipeline, then using two pipelines to run the two sides of a branch instruction at the same time. In order to test the new method, the HLA (High Level architecture) architecture-based simulation system is established, the results show that the new method can really reduce the time when processing the random branch instructions.
StencilFlow: Mapping Large Stencil Programs to Distributed Spatial Computing Systems
