Module graph 

CUDA Graph API for recording and replaying sequences of GPU operations.

CUDA Graphs allow capturing a sequence of operations (kernel launches, memory copies, memsets) into a graph data structure that can be instantiated and launched repeatedly with minimal CPU overhead.

Architecture

This module exposes a Rust-side graph representation that records operations as nodes with explicit dependency edges. Graph::instantiate translates that representation into the native CUDA Graph API (cuGraphCreate + cuGraphInstantiate) whenever a CUDA driver is available, and GraphExec::launch issues a real cuGraphLaunch. On macOS (or any host without a driver) the graph is still built and validated CPU-side, and launching reports CudaError::NotInitialized.

Node lowering (important)

A GraphNode stores only an operation specification — a kernel name, copy direction/size, or memset size/value — and carries no resolved CUfunction or device pointers. Kernel / memcpy / memset nodes therefore cannot yet be lowered to their real cuGraphAddKernelNode / cuGraphAddMemcpyNode / cuGraphAddMemsetNode form: they are added as cuGraphAddEmptyNode barriers. The instantiated graph faithfully reproduces the node count and dependency topology, but a cuGraphLaunch performs none of those operations’ work (instantiate logs a tracing::warn! when any such node is lowered). Only genuine GraphNode::Empty barriers are represented exactly.

Example

let mut graph = Graph::new();

let n0 = graph.add_memcpy_node(MemcpyDirection::HostToDevice, 4096);
let n1 = graph.add_kernel_node(
    "vector_add",
    (4, 1, 1),
    (256, 1, 1),
    0,
);
let n2 = graph.add_memcpy_node(MemcpyDirection::DeviceToHost, 4096);

graph.add_dependency(n0, n1).ok();
graph.add_dependency(n1, n2).ok();

assert_eq!(graph.node_count(), 3);
assert_eq!(graph.dependency_count(), 2);

Structs

Graph

A CUDA graph representing a DAG of GPU operations.

GraphExec

An instantiated, executable graph.

StreamCapture

Records GPU operations submitted to a stream into a Graph.

StreamGraphCapture

Driver-backed CUDA stream capture.

Enums

GraphNode

A single operation node within a Graph.

MemcpyDirection

Direction of a memory copy operation within a graph node.