Struct GpuNdarray 

pub struct GpuNdarray<T: GpuScalar> { /* private fields */ }

A GPU-backed n-dimensional array backed by a real wgpu Buffer.

Created via GpuNdarray::from_ndarray_data or GpuNdarray::from_data. Converted back with GpuNdarray::to_ndarray.

Implementations

impl<T: GpuScalar> GpuNdarray<T>

pub fn buffer_arc(&self) -> &Arc<Buffer>

Expose the underlying Arc<wgpu::Buffer> for zero-copy checks in tests.

impl GpuNdarray<f32>

pub fn from_ndarray_data( data: &[f32], shape: Vec<usize>, context: Arc<WebGPUContext>, ) -> Result<Self, GpuError>

Upload a CPU NdarrayWrapper<f32, _> to the GPU.

Returns Err(GpuError) if no adapter is available or upload fails.

pub fn to_vec(&self) -> Result<Vec<f32>, GpuError>

Download the GPU buffer to a flat Vec<f32>.

pub fn to_ndarray(&self) -> Result<ArrayD<f32>, GpuError>

Download to a dynamic ndarray.

pub fn from_data(data: &[f32], shape: Vec<usize>) -> Result<Self, GpuError>

Construct from a WebGPUContext, uploading data.

Uses the process-wide singleton WebGPUContext so that all arrays created via this function share the same wgpu::Device and can be combined freely in kernel dispatches.

pub fn add(&self, other: &GpuNdarray<f32>) -> Result<GpuNdarray<f32>, GpuError>

Elementwise add: self + other. Both must have the same shape.

pub fn subtract( &self, other: &GpuNdarray<f32>, ) -> Result<GpuNdarray<f32>, GpuError>

Elementwise subtract: self - other. Both must have the same shape.

pub fn multiply( &self, other: &GpuNdarray<f32>, ) -> Result<GpuNdarray<f32>, GpuError>

Elementwise multiply: self * other. Both must have the same shape.

pub fn multiply_by_scalar_f32( &self, scalar: f32, ) -> Result<GpuNdarray<f32>, GpuError>

Multiply every element by a scalar: self * scalar.

pub fn sum_all(&self) -> Result<f32, GpuError>

Compute the sum of all elements (dot product via a.multiply(b)?.sum_all()).

pub fn dot_gpu(&self, other: &GpuNdarray<f32>) -> Result<f32, GpuError>

Compute the dot product self · other = sum(self * other).

pub fn matmul( &self, other: &GpuNdarray<f32>, ) -> Result<GpuNdarray<f32>, GpuError>

Tiled 16x16 matrix multiplication: self x other.

Both arrays must be 2-D. For a matrix-vector product, upload the vector as a [n, 1] column; the result will have shape [m, 1].

Trait Implementations

impl ArrayProtocol for GpuNdarray<f32>

fn array_function( &self, func: &ArrayFunction, _types: &[TypeId], args: &[Box<dyn Any>], kwargs: &HashMap<String, Box<dyn Any>>, ) -> Result<Box<dyn Any>, NotImplemented>

Implementation of the array protocol.

fn as_any(&self) -> &dyn Any

Get the array as Any for downcasting

fn shape(&self) -> &[usize]

Get the shape of the array (default implementation returns empty slice)

fn dtype(&self) -> TypeId

Get the data type of the array (default implementation returns f64)

fn box_clone(&self) -> Box<dyn ArrayProtocol>

Clone this array protocol object.

impl<T: GpuScalar> Clone for GpuNdarray<T>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: GpuScalar> Debug for GpuNdarray<T>

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

Formats the value using the given formatter.

impl GPUArray for GpuNdarray<f32>

fn to_gpu(&self) -> CoreResult<Box<dyn GPUArray>>

Move the array to GPU.

fn to_cpu(&self) -> CoreResult<Box<dyn ArrayProtocol>>

Move the array from GPU to CPU.

fn is_on_gpu(&self) -> bool

Check if the array is on GPU.

fn device_info(&self) -> HashMap<String, String>

Get information about the GPU device that holds this array.