Struct DeviceScalar 

pub struct DeviceScalar<T: DeviceRepr + ValidAsZeroBits + Copy> { /* private fields */ }

A single scalar value stored in GPU device memory.

Used for CUDA graph capture: the graph records the device pointer (fixed address), and the caller updates the value via update before each CapturedGraph::launch. The update is a 4-or-8 byte cuMemcpyHtoDAsync — effectively zero cost.

Implementations

impl<T: DeviceRepr + ValidAsZeroBits + Copy> DeviceScalar<T>

pub fn new(stream: &Arc<CudaStream>, initial: T) -> GpuResult<Self>

Allocate a device scalar with the given initial value.

pub fn update(&mut self, value: T) -> GpuResult<()>

Update the device value. This is an async H→D memcpy of size_of::<T>() bytes. Must be called on the same stream as the graph to ensure ordering.

pub fn inner(&self) -> &CudaSlice<T>

Borrow the underlying CudaSlice for use as a kernel parameter. The graph captures this pointer address; updating the value later changes what the kernel reads without re-capturing.