Struct CudaBackend 

pub struct CudaBackend;

Re-export CUDA backend (requires cuda feature + NVIDIA CUDA Toolkit). The CUDA GPU backend. This is a zero-sized marker type.

Trait Implementations

impl Backend for CudaBackend

type Device = CudaDevice

The device type for this backend.

type Storage = CudaStorage

The storage type for this backend.

fn zeros( shape: &Shape, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Allocate storage filled with zeros.

fn ones( shape: &Shape, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Allocate storage filled with ones.

fn full( shape: &Shape, val: f64, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Allocate storage filled with a constant value.

fn from_f64_slice( data: &[f64], dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Create storage from a flat f64 slice, converting to the target dtype.

fn rand_uniform( shape: &Shape, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Create storage with random uniform values in [0, 1).

fn rand_normal( shape: &Shape, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Create storage with random normal values (mean=0, std=1).

fn binary_op( op: BinaryOp, lhs: &CudaStorage, lhs_layout: &Layout, rhs: &CudaStorage, rhs_layout: &Layout, ) -> Result<CudaStorage, Error>

Apply a binary op element-wise: result[i] = op(lhs[i], rhs[i]). The layouts handle broadcasting and non-contiguous access.

fn unary_op( op: UnaryOp, input: &CudaStorage, layout: &Layout, ) -> Result<CudaStorage, Error>

Apply a unary op element-wise: result[i] = op(input[i]).

fn reduce_op( op: ReduceOp, input: &CudaStorage, layout: &Layout, dims: &[usize], _keep_dim: bool, ) -> Result<CudaStorage, Error>

Reduce along specific dimensions. If dims is empty, reduce over all elements.

fn matmul( lhs: &CudaStorage, lhs_layout: &Layout, rhs: &CudaStorage, rhs_layout: &Layout, ) -> Result<CudaStorage, Error>

General matrix multiply: C = A @ B. Supports batched matmul for tensors with rank > 2.

fn to_contiguous( input: &CudaStorage, layout: &Layout, ) -> Result<CudaStorage, Error>

Make a contiguous copy of the storage following the given layout. If the layout is already contiguous, this may just clone the storage.

fn to_f64_vec(input: &CudaStorage, layout: &Layout) -> Result<Vec<f64>, Error>

Copy data from this storage to a Vec on the host (for inspection).

fn cmp_op( op: CmpOp, lhs: &CudaStorage, lhs_layout: &Layout, rhs: &CudaStorage, rhs_layout: &Layout, ) -> Result<CudaStorage, Error>

Element-wise comparison, returns a u8 storage (0 or 1).

fn affine( input: &CudaStorage, layout: &Layout, mul: f64, add: f64, ) -> Result<CudaStorage, Error>

Affine transform: result = input * mul + add. Used for normalization and other fused operations.

fn index_select( input: &CudaStorage, input_layout: &Layout, indices: &CudaStorage, indices_layout: &Layout, dim: usize, ) -> Result<CudaStorage, Error>

Gather elements along a dimension using index tensor.

fn powf( input: &CudaStorage, layout: &Layout, exponent: f64, ) -> Result<CudaStorage, Error>

Element-wise power: result[i] = input[i] ^ exponent.

fn clamp( input: &CudaStorage, layout: &Layout, min: f64, max: f64, ) -> Result<CudaStorage, Error>

Element-wise clamp: result[i] = clamp(input[i], min, max).

fn where_cond( mask: &CudaStorage, mask_layout: &Layout, on_true: &CudaStorage, on_true_layout: &Layout, on_false: &CudaStorage, on_false_layout: &Layout, ) -> Result<CudaStorage, Error>

Element-wise conditional: result[i] = if mask[i] != 0 { on_true[i] } else { on_false[i] }.

fn gather( input: &CudaStorage, input_layout: &Layout, index: &CudaStorage, index_layout: &Layout, dim: usize, ) -> Result<CudaStorage, Error>

Gather elements along dim using index tensor.

fn cat( inputs: &[(&CudaStorage, &Layout)], out_shape: &Shape, dim: usize, ) -> Result<CudaStorage, Error>

Concatenate multiple storages along dim into a single contiguous storage. Each entry is (storage, layout) so non-contiguous inputs are handled correctly. out_shape is the pre-validated output shape.

fn cast( input: &CudaStorage, layout: &Layout, dtype: DType, device: &CudaDevice, ) -> Result<CudaStorage, Error>

Cast storage to a different dtype on-device (no host round-trip).

impl Clone for CudaBackend

fn clone(&self) -> CudaBackend

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CudaBackend

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

Formats the value using the given formatter.