{"title":"数据流图中的节点预取预测","authors":"N.G. Petersen, M. R. Wójcik","doi":"10.1109/SIPS.2004.1363068","DOIUrl":null,"url":null,"abstract":"Dataflow languages provide a high-level description that can expose inherent parallelism in many applications. This high level description can be applied to automatically create efficient code and schedules based on patterns in the dataflow graphs and knowledge of the target architecture. When targeting a dataflow graph to custom hardware, it is sometimes advantageous to share nodes with similar functionality to save silicon. Any state information associated with the caller of the shared node must be stored and subsequently loaded upon firing. If prediction logic can predict which caller of a shared node is next, the associated state information can be prefetched while other nodes of the graph are executing. While some applications can be entirely scheduled at compile time, many multi- channel measurement and control applications require some degree of dynamic scheduling. This paper's key contribution is a lightweight call prediction unit with 100% prediction accuracy given a runtime-determined periodic calling schedule. While applications are varied, we show a 33% speedup in a filtering application possible in wireless ad hoc networks.","PeriodicalId":384858,"journal":{"name":"IEEE Workshop onSignal Processing Systems, 2004. SIPS 2004.","volume":"27 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2004-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Node prefetch prediction in dataflow graphs\",\"authors\":\"N.G. Petersen, M. R. Wójcik\",\"doi\":\"10.1109/SIPS.2004.1363068\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Dataflow languages provide a high-level description that can expose inherent parallelism in many applications. This high level description can be applied to automatically create efficient code and schedules based on patterns in the dataflow graphs and knowledge of the target architecture. When targeting a dataflow graph to custom hardware, it is sometimes advantageous to share nodes with similar functionality to save silicon. Any state information associated with the caller of the shared node must be stored and subsequently loaded upon firing. If prediction logic can predict which caller of a shared node is next, the associated state information can be prefetched while other nodes of the graph are executing. While some applications can be entirely scheduled at compile time, many multi- channel measurement and control applications require some degree of dynamic scheduling. This paper's key contribution is a lightweight call prediction unit with 100% prediction accuracy given a runtime-determined periodic calling schedule. While applications are varied, we show a 33% speedup in a filtering application possible in wireless ad hoc networks.\",\"PeriodicalId\":384858,\"journal\":{\"name\":\"IEEE Workshop onSignal Processing Systems, 2004. SIPS 2004.\",\"volume\":\"27 1\",\"pages\":\"0\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2004-12-06\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"IEEE Workshop onSignal Processing Systems, 2004. SIPS 2004.\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/SIPS.2004.1363068\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Workshop onSignal Processing Systems, 2004. SIPS 2004.","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/SIPS.2004.1363068","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Dataflow languages provide a high-level description that can expose inherent parallelism in many applications. This high level description can be applied to automatically create efficient code and schedules based on patterns in the dataflow graphs and knowledge of the target architecture. When targeting a dataflow graph to custom hardware, it is sometimes advantageous to share nodes with similar functionality to save silicon. Any state information associated with the caller of the shared node must be stored and subsequently loaded upon firing. If prediction logic can predict which caller of a shared node is next, the associated state information can be prefetched while other nodes of the graph are executing. While some applications can be entirely scheduled at compile time, many multi- channel measurement and control applications require some degree of dynamic scheduling. This paper's key contribution is a lightweight call prediction unit with 100% prediction accuracy given a runtime-determined periodic calling schedule. While applications are varied, we show a 33% speedup in a filtering application possible in wireless ad hoc networks.