Struct ScalarKernel

pub struct ScalarKernel { /* private fields */ }

An instance of a scalar/elementwise kernel. Can be build for any operation, rank, strides, and sizes. The first axis is runtime-dynamic without recompiling the kernel.

Implementations

impl ScalarKernel

pub fn new( device: CudaDevice, operation: &str, inner_shape: Vec<usize>, operand_types: Vec<String>, operand_strides: Vec<Vec<isize>>, ) -> Self

Compile an instance of a new scalar kernel.

operation has the format *x0 = *x1 + *x2;.

pub fn new_for_shapes( device: CudaDevice, operation: &str, shapes: &[StridedShape], operand_types: Vec<String>, ) -> Self

Wrapper around Self::new that’s a bit easier to use if you know the full shape of the operands up front.

pub unsafe fn run(&self, stream: &CudaStream, tensors: &[DeviceTensor])

pub unsafe fn run_custom( &self, stream: &CudaStream, tensors: &[DeviceTensor], items_per_thread: u32, threads_per_block: u32, )

Trait Implementations

impl Debug for ScalarKernel

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

Formats the value using the given formatter.