gpu_tensor library

GPU Tensor library.

This library provides a simple API for performing tensor operations on the GPU.

Classes

DefaultMinigpu

A helper that creates (or reuses) a default GPU context.

Fft1dPlan

A reusable execution plan for repeated same-size 1D real-input FFTs (the audio-frame hot path).

Tensor<T extends TypedData>

A generic tensor that works with a specific TypedData type.

TensorShaderCache

Process-wide cache of compiled compute shaders, keyed per-Minigpu instance by the FULL WGSL source.

Enums

BufferDataType

Enum representing supported buffer data types.

Extensions

CachedShaderAcquire on Minigpu

GpuActivation on Tensor<T>

GpuFft on Tensor<TypedData>

LinearDispatch on ComputeShader

Dispatches a 1D logical workload of threads invocations, folding into (x, y) workgroup counts when the workgroup count exceeds WebGPU's 65535 per-dimension limit (tensors > ~16.7M elements at workgroup_size 256).

TensorData on Tensor<T>

TensorLinearOperator on Tensor<TypedData>

TensorOperator on Tensor<T>

TensorPoolingMax on Tensor<TypedData>

TensorPoolingMin on Tensor<TypedData>

TensorPrintHelpers on Tensor<T>

TesorHelper on Tensor<T>

Constants

linearIndexWGSL → const String

linearMainSignature → const String

Canonical main-signature + linear-index preamble for 1D kernels that are dispatched via LinearDispatch.dispatchLinear. Uses num_workgroups so the source stays independent of the dispatch size (cache friendly).

Functions

getByteConversionCode(BufferDataType dataType) → String

getCastExpression(String value, BufferDataType dataType) → String

getTypedDataElement(TypedData data, int index) → num

Returns the element (as num) at index for a TypedData, if supported.

getTypedDataSublist<T extends TypedData>(T data, int start, int end) → T

Returns a sublist of data from start (inclusive) to end (exclusive), preserving the underlying type.

getZeroValue(BufferDataType dataType) → String

prepareShader(String template, BufferDataType dataType, Map<String, dynamic> values) → String