Struct CapturedCudaGraph

Source

pub struct CapturedCudaGraph { /* private fields */ }

Expand description

Instantiated CUDA Graph with owned graph + exec handles.

Implementations§

Source §

impl CapturedCudaGraph

Source

pub fn bind_resident_lifecycle(self, runtime: &XlogDeviceRuntime) -> Self

Tie resident lifecycle accounting to this real graph/exec owner.

Binding is idempotent. The lease is created only after graph instantiation has succeeded and is dropped after this type’s Drop implementation destroys the executable and parent graph handles.

Source

pub fn conditional_while_on_stream<F>( stream: &CudaStream, initial_value: u32, assign_default_on_launch: bool, populate_body: F, ) -> Result<Self, CudaConditionalGraphUnavailable>
where F: FnOnce(ConditionalCudaGraphBody) -> Result<(), CudaConditionalGraphUnavailable>,

Create, populate, and instantiate a parent graph containing exactly one root conditional-WHILE node.

The handle and body passed to populate_body remain valid until this CapturedCudaGraph is dropped. CUDA permits only one live executable instantiation of a graph containing a conditional node.

Source

pub unsafe fn instantiate_owned_graph( graph: CUgraph, context: Arc<CudaContext>, ) -> Result<Self, CudaConditionalGraphUnavailable>

Instantiate and assume sole ownership of graph.

The graph is destroyed on instantiation failure. On success, this value destroys the executable first and the source graph second.

§Safety

graph must be a valid, unowned graph in context, and no other owner may destroy or instantiate it while this value exists.

Source

pub fn capture_on_stream<F>(stream: &CudaStream, record: F) -> Result<Self>
where F: FnOnce() -> Result<()>,

Capture work submitted by record on stream, instantiate it, and take ownership of the resulting graph handles.

Source

pub fn launch(&self, stream: &CudaStream) -> Result<()>

Replay the instantiated graph on stream.

Source

pub fn node_count(&self) -> Result<usize>

Number of nodes in the captured graph. Used by bounded CSM CUDA Graph cache-key and node-inventory certs to prove topology stability.

Source

pub fn nodes(&self) -> Result<Vec<CudaGraphNode>>

Return graph nodes in CUDA’s enumeration order with their node type.

CUDA does not define this list as dependency or execution order. Use Self::linear_chain_node_kinds when a linear topology is required.

Source

pub fn linear_chain_node_kinds(&self) -> Result<Vec<CudaGraphNodeKind>>

Return actual node kinds in root-to-leaf dependency order.

This fails unless the graph is one connected linear chain with exactly one root, one leaf, one immediate dependency per non-root node, and no branches, duplicate dependencies, foreign dependencies, or cycles.

Source

pub fn kernel_node_params( &self, node: CudaGraphNode, ) -> Result<CUDA_KERNEL_NODE_PARAMS>

Read CUDA’s raw kernel-node params for inventory/update code.

The returned kernelParams pointer is CUDA-owned capture metadata. Treat it as read-only unless constructing a fresh params object for Self::set_kernel_node_params.

Source

pub unsafe fn set_kernel_node_params( &self, node: CudaGraphNode, params: &CUDA_KERNEL_NODE_PARAMS, ) -> Result<()>

Update a kernel node in the instantiated graph.

§Safety

CUDA requires the replacement params to be topology-compatible with the captured node. The caller must keep every pointed-to kernel argument alive until CUDA has consumed the update and launched work that uses it.

Source