Struct GraphExec 

pub struct GraphExec { /* private fields */ }

An instantiated, executable graph.

Created by Graph::instantiate, a GraphExec holds a snapshot of the graph and a pre-computed execution order.

Driver backing

When a CUDA driver is available, instantiate builds a genuine CUgraph (cuGraphCreate + one cuGraphAdd*Node per in-memory node, with the dependency DAG wired through real CUgraphNode edges) and finalises it into a CUgraphExec via cuGraphInstantiate. In that case launch issues a real cuGraphLaunch.

The in-memory GraphNode representation stores only an operation specification (kernel name, copy direction/size, memset size/value) — it carries no resolved CUfunction or device pointers. Every node is therefore translated to a real cuGraphAddEmptyNode: the resulting driver graph reproduces the node count and dependency topology exactly and executes on the GPU as a DAG of synchronisation barriers. The per-node dispatch in Graph::build_driver_graph is structured so that kernel / memcpy / memset nodes that gain concrete device operands can be promoted to cuGraphAddKernelNode / cuGraphAddMemcpyNode / cuGraphAddMemsetNode without further restructuring.

On macOS (or any host without a CUDA driver), no driver handles are created; the graph is still validated (topological sort) and launch returns CudaError::NotInitialized.

Implementations

impl GraphExec

pub fn launch(&self, stream: &Stream) -> CudaResult<()>

Launches the executable graph on the given stream.

When this GraphExec is backed by a real CUgraphExec, this issues cuGraphLaunch(hGraphExec, hStream), submitting the entire graph to the stream with minimal CPU overhead. Otherwise it surfaces the driver-load error.

Errors

• CudaError::NotInitialized if the CUDA driver is not available (e.g. on macOS, or a host without an NVIDIA GPU).
• Any CudaError mapped from cuGraphLaunch.

pub fn graph(&self) -> &Graph

Returns a reference to the underlying graph.

pub fn execution_order(&self) -> &[usize]

Returns the pre-computed execution order (topological sort).

pub fn node_count(&self) -> usize

Returns the total number of nodes that would be executed.

pub fn is_driver_backed(&self) -> bool

Returns true if this GraphExec is backed by a real, live CUgraphExec driver handle (as opposed to a CPU-side-only graph).

Trait Implementations

impl Debug for GraphExec

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Drop for GraphExec

fn drop(&mut self)

Executes the destructor for this type.

fn pin_drop(self: Pin<&mut Self>)

🔬This is a nightly-only experimental API. (pin_ergonomics)

Execute the destructor for this type, but different to Drop::drop, it requires self to be pinned.