Enum KernelArg 

pub enum KernelArg<'a> {
    Scalar(Vec<u8>),
    DevicePtr(*mut c_void, PhantomData<&'a ()>),
}

A typed argument for a raw CUDA kernel launch.

Either a value scalar (copied into a stable buffer) or a device pointer obtained from a validated TensorRef/TensorMut or DeviceBytes. Constructing an argument is safe; the ABI contract is enforced by the caller at the unsafe launch site.

Variants

Scalar(Vec<u8>)

A scalar value encoded as an exact-size byte sequence.

DevicePtr(*mut c_void, PhantomData<&'a ()>)

A device memory address argument (in/out pointer).

Implementations

impl<'a> KernelArg<'a>

pub fn scalar(bytes: &[u8]) -> Self

Build a scalar argument from an exact-size value encoding.

pub fn u32(value: u32) -> Self

Build a u32 scalar parameter.

pub fn i32(value: i32) -> Self

Build an i32 scalar parameter.

pub fn f32(value: f32) -> Self

Build a f32 scalar parameter.

pub fn u64(value: u64) -> Self

Build a u64 scalar parameter.

pub fn i64(value: i64) -> Self

Build an i64 scalar parameter.

pub fn f64(value: f64) -> Self

Build a f64 scalar parameter.

pub fn tensor<T>(reference: &TensorRef<'a, T>) -> Self

Build a device-pointer parameter from a read-only tensor reference.

The kernel may only read the tensor unless the launch safety contract states otherwise; for a write target use KernelArg::tensor_mut or KernelArg::output.

pub fn tensor_mut<T>(reference: &TensorMut<'a, T>) -> Self

Build a device-pointer parameter from a mutable tensor reference.

pub fn output<T>(reference: &TensorMut<'a, T>) -> Self

Build a device-pointer parameter from a freshly allocated output tensor.

pub fn workspace(bytes: &DeviceBytes<'a>) -> Self

Build a device-pointer parameter from a workspace allocation.

Trait Implementations

impl<'a> Clone for KernelArg<'a>

fn clone(&self) -> KernelArg<'a>

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.