Crate dspk_ops 

Signal processing for multi-channel recordings on the device (any GPU through wgpu, CUDA or HIP, or the CPU), written as CubeCL kernels: filters, spatial operators, statistics, linear algebra, peak finding, threshold crossings and resampling.

Where things are

Need	Module	Start with
Chain preprocessing steps	pipeline	Pipeline, PipelineStage
Filters (IIR, FIR, median, templates)	filter	filter::FilterSpec, filter::DeviceFilter
Mix channels (CAR, whitening, Laplacian)	spatial	SpatialWhitening, SurfaceLaplacian
Noise, percentiles, element-wise math	math	math::estimate_noise_std
Matrix products, PCA, ICA	linalg	PcaModel, linalg::matmul
Peaks (find_peaks)	peaks	peaks::find_peak_candidates
Threshold crossings as events (onset, offset, peak)	crossings	crossings::detect_crossings_on_device
Resampling	resampler	resampler::execute_resample_poly
Device buffers and shared kernels	core	core::buffer

Data is a [channels, samples] buffer, channel-major. Functions take a cubecl Client, from dspk_core::compute::ComputeTarget: ComputeTarget::from_env()?.client()? picks the runtime named by DSPK_RUNTIME, otherwise the first compiled in. See Pipeline for a complete example.

Re-exports

pub use error::OpsError;

pub use error::OpsResult;

pub use pipeline::Pipeline;

pub use pipeline::PipelineStage;

pub use linalg::FastIcaModel;

pub use linalg::IcaContrast;

pub use linalg::PcaModel;

pub use linalg::PpcaModel;

pub use spatial::SpatialWhitening;

pub use spatial::SurfaceLaplacian;

Modules

autotune

Tunes every autotuned operation on this device ahead of time.

core

Device-side building blocks shared by every dspk-ops algorithm.

crossings

Threshold crossings as events: where each channel goes past a level (above it, below it, or either) and where it comes back. One tool for any signal: muscle contractions on an EMG envelope (band-pass → rectify → low-pass → crossings), excursions of a raw trace past ±level, a temperature above a limit, a trigger line going high.

error

The one error type of dspk-ops.

filter

Filters for [channels, samples] signals on the device.

linalg

Linear algebra on the device: matrix products (matmul), covariance, symmetric eigendecomposition, Cholesky; and the models built on them: PcaModel, TopComponents (the leading principal components of many rows, streamed in batches), PpcaModel (probabilistic PCA), FastIcaModel (independent components), and DeviceProjection (applying a fitted projection to device buffers).

math

Element-wise operations and statistics.

peaks

Peak finding, as scipy.signal.find_peaks:

pipeline

resampler

Sample-rate conversion on the device, following scipy.signal:

spatial

Operators that mix channels: each output channel is a weighted sum of input channels.