Struct CudaRuntime 

pub struct CudaRuntime;

Trait Implementations

impl Debug for CudaRuntime

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

Formats the value using the given formatter.

impl Runtime for CudaRuntime

type Compiler = CppCompiler<CudaDialect<CudaWmmaCompiler>>

The compiler used to compile the inner representation into tokens.

type Server = CudaServer

The compute server used to run kernels and perform autotuning.

type Device = CudaDevice

The device used to retrieve the compute client.

fn client(device: &Self::Device) -> ComputeClient<Self::Server>

Retrieve the compute client from the runtime device.

fn name(_client: &ComputeClient<Self::Server>) -> &'static str

The runtime name on the given device.

fn require_array_lengths() -> bool

Return true if global input array lengths should be added to kernel info.

fn supported_line_sizes() -> &'static [u8]

Returns the supported line sizes for the current runtime’s compiler.

fn max_cube_count() -> (u32, u32, u32)

Returns the maximum cube count on each dimension that can be launched.

fn can_read_tensor(shape: &[usize], strides: &[usize]) -> bool

fn target_properties() -> TargetProperties

Returns the properties of the target hardware architecture.

fn max_global_line_size() -> u8

The maximum line size that can be used for global buffer bindings.

fn io_optimized_line_sizes( elem: &StorageType, ) -> impl Iterator<Item = u8> + Clone

Returns all line sizes that are useful to perform optimal IO operation on the given element.

fn io_optimized_line_sizes_unchecked( size: usize, ) -> impl Iterator<Item = u8> + Clone

Returns all line sizes that are useful to perform optimal IO operation on the given element. Ignores native support, and allows all line sizes. This means the returned size may be unrolled, and may not support dynamic indexing.