Module gpu_ndarray 

Real wgpu-backed GpuNdarray<T> that implements ArrayProtocol.

Enabled only with the array_protocol_wgpu feature, which implies wgpu_backend.

Supported operations (GPU dispatch)

• add, subtract, multiply — elementwise binary, workgroup (256,1,1), uses arrayLength
• multiply_by_scalar_f32 — elementwise scalar multiply, workgroup (256,1,1)
• matmul — naive (one thread per output element), workgroup (16,16,1)
• sum(axis=None) — two-pass reduce, workgroup (256,1,1)
• transpose (2-D) — 16×16 bank-conflict-padded tile, workgroup (16,16,1) (32×32 exceeds Metal’s 256-invocation-per-workgroup limit)
• concatenate(axis=0) — via copy_buffer_to_buffer, no shader
• concatenate(axis>0) — WGSL gather kernel (CONCAT_AXISN_WGSL), storage-buf strides
• sum(axis=Some(ax)) — WGSL per-output-element axis reduction (REDUCE_SUM_AXIS_WGSL)
• reshape — zero-copy (clone Arc<Buffer>, new shape/strides)

CPU-fallback operations

• svd — falls back to CPU NdarrayWrapper
• inverse — falls back to CPU NdarrayWrapper
• multiply_by_scalar_f64, divide_by_scalar_f64 — convert to f32, then GPU
• GPU kernel errors on axis ops — fallback to CPU (graceful degradation)

GPU threshold

Arrays with fewer than 4096 elements skip GPU dispatch entirely and fall back to CPU.

Structs

GpuNdarray

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

Traits

GpuScalar

Marker trait for element types that wgpu-29 supports natively (f32 only; f64 is not supported in WGSL without extensions).

Functions

global_context

Returns the shared WebGPUContext, or None if no adapter is available.

is_gpu_available

Returns true if a wgpu adapter was found when first called; cached afterwards.