Struct StreamCapture 

pub struct StreamCapture { /* private fields */ }

Records GPU operations submitted to a stream into a Graph.

Stream capture intercepts operations that would normally be submitted to a CUDA stream and instead records them as graph nodes. The captured operations can then be replayed efficiently via GraphExec.

Usage

let mut capture = StreamCapture::begin(&stream)?;

capture.record_kernel("my_kernel", (4, 1, 1), (256, 1, 1), 0);
capture.record_memcpy(MemcpyDirection::DeviceToHost, 1024);

let graph = capture.end()?;
assert_eq!(graph.node_count(), 2);

Implementations

impl StreamCapture

pub fn begin(_stream: &Stream) -> CudaResult<Self>

Begins capturing operations on the given stream.

On a real CUDA system, this would call cuStreamBeginCapture(stream, CU_STREAM_CAPTURE_MODE_GLOBAL).

Errors

Returns CudaError::NotInitialized if the CUDA driver is not available.

pub fn record_kernel( &mut self, function_name: &str, grid: (u32, u32, u32), block: (u32, u32, u32), shared_mem: u32, )

Records a kernel launch operation in the capture.

Parameters

• function_name - Name of the kernel function.
• grid - Grid dimensions (x, y, z).
• block - Block dimensions (x, y, z).
• shared_mem - Dynamic shared memory in bytes.

pub fn record_memcpy(&mut self, direction: MemcpyDirection, size: usize)

Records a memory copy operation in the capture.

Parameters

• direction - Direction of the memory copy.
• size - Size of the transfer in bytes.

pub fn record_memset(&mut self, size: usize, value: u8)

Records a memset operation in the capture.

Parameters

• size - Number of bytes to set.
• value - Byte value to fill with.

pub fn recorded_count(&self) -> usize

Returns the number of operations recorded so far.

pub fn is_active(&self) -> bool

Returns whether the capture is still active.

pub fn end(self) -> CudaResult<Graph>

Ends the capture and returns the resulting Graph.

On a real CUDA system, this would call cuStreamEndCapture and return the captured graph handle.

The captured nodes are connected in a linear chain (each node depends on the previous one) to preserve the order in which operations were recorded.

Errors

Returns CudaError::StreamCaptureUnmatched if the capture was already ended.