Module cuda_graph 

CUDA Graph capture backend.

Converts an ExecutionPlan into an oxicuda_driver::graph::Graph, which can be instantiated and launched with minimal CPU overhead via the CUDA driver’s graph API.

Mapping

PlanStep	driver::graph::GraphNode
KernelLaunch	GraphNode::KernelLaunch
Memcpy	GraphNode::Memcpy
Memset	GraphNode::Memset
Barrier	GraphNode::Empty
EventRecord/Wait	GraphNode::Empty (sync barrier)
HostCallback	GraphNode::Empty (placeholder)

The dependency edges from the original plan are preserved: each step depends on all steps that produced an event it is waiting for.

Limitations

• Stream-parallel execution is not yet mapped to separate CUDA stream arguments — all nodes are enqueued on the same graph-internal stream. Proper multi-stream CUDA graph support requires the cuGraphAddNode API with cuGraphNodeParams, which is future work.
• EventRecord/EventWait steps collapse into Empty nodes with a dependency edge, which enforces ordering without the overhead of true event signalling within a CUDA graph.

Functions

capture

Captures an ExecutionPlan into a driver::graph::Graph.