Skip to content

4.2 CUDA Graphs ​

待翻译

本节已建立目录框架,正文尚未翻译。

4.2.1 Graph Structure ​

4.2.1.1 Node Types ​

4.2.1.2 Edge Data ​

4.2.2 Building and Running Graphs ​

4.2.2.1 Graph Creation ​

4.2.2.1.1 Graph APIs ​

4.2.2.1.2 Stream Capture ​

4.2.2.1.2.1 Cross-stream Dependencies and Events ​

4.2.2.1.2.2 Prohibited and Unhandled Operations ​

4.2.2.1.2.3 Invalidation ​

4.2.2.1.2.4 Capture Introspection ​

4.2.2.1.3 Putting It All Together ​

4.2.2.2 Graph Instantiation ​

4.2.2.3 Graph Execution ​

4.2.3 Updating Instantiated Graphs ​

4.2.3.1 Whole Graph Update ​

4.2.3.2 Individual Node Update ​

4.2.3.3 Individual Node Enable ​

4.2.3.4 Graph Update Limitations ​

4.2.4 Conditional Graph Nodes ​

4.2.4.1 Conditional Handles ​

4.2.4.2 Conditional Node Body Graph Requirements ​

4.2.4.3 Conditional IF Nodes ​

4.2.4.4 Conditional WHILE Nodes ​

4.2.4.5 Conditional SWITCH Nodes ​

4.2.5 Graph Memory Nodes ​

4.2.5.1 Introduction ​

4.2.5.2 API Fundamentals ​

4.2.5.2.1 Graph Node APIs ​

4.2.5.2.2 Stream Capture ​

4.2.5.2.3 Accessing and Freeing Graph Memory Outside of the Allocating Graph ​

4.2.5.2.4 cudaGraphInstantiateFlagAutoFreeOnLaunch ​

4.2.5.2.5 Memory Nodes in Child Graphs ​

4.2.5.3 Optimized Memory Reuse ​

4.2.5.3.1 Address Reuse within a Graph ​

4.2.5.3.2 Physical Memory Management and Sharing ​

4.2.5.4 Performance Considerations ​

4.2.5.4.1 First Launch / cudaGraphUpload ​

4.2.5.5 Physical Memory Footprint ​

4.2.5.6 Peer Access ​

4.2.5.6.1 Peer Access with Graph Node APIs ​

4.2.5.6.2 Peer Access with Stream Capture ​

4.2.6 Device Graph Launch ​

4.2.6.1 Device Graph Creation ​

4.2.6.1.1 Device Graph Requirements ​

4.2.6.1.2 Device Graph Upload ​

4.2.6.1.3 Device Graph Update ​

4.2.6.2 Device Launch ​

4.2.6.2.1 Fire and Forget Launch ​

4.2.6.2.1.1 Graph Execution Environments ​

4.2.6.2.2 Tail Launch ​

4.2.6.2.2.1 Tail Self-launch ​

4.2.6.2.3 Sibling Launch ​

4.2.7 Using Graph APIs ​

4.2.8 CUDA User Objects ​