Image-signal front
Data as a regular grid of samples — the DWT transform pair, convolution, reductions. The classic image-analysis primitives.
GPU Spatial Primitives
A small toolbox of composable, interpretable GPU primitives for analysing point clouds and shapes — cells, transcripts, boundaries — on TypeGPU / WebGPU.
Two fronts, one runtime
Section titled “Two fronts, one runtime”This toolbox grew out of a WebGPU wavelet codec and has two complementary sides:
Discrete-cell front
Data as irregular points and shapes — neighbour search, density, hotspots, fuzzy graphs. Spatial-biology-style analysis, but the maths is generic.
The bridge between them is rasterising a point cloud into a grid (a density splat), after which the image primitives apply unchanged. The shared idea running through both is windowing: a smooth, overlapping kernel in place of a hard threshold.
Design commitments
Section titled “Design commitments”Composable
Each primitive takes plain arrays in and returns plain arrays out. Chain them:
points → splat → window → z-score is a hotspot map.
Interpretable
Every output has a meaning you can state in a sentence — clustered vs dispersed, hot spot vs cold spot, connection strength — not a black box.
Validated
Every primitive has a CPU golden and a GPU test asserting bit-exact (integer) or bounded-error (float) agreement.
Translatable
Dependency-light raw WebGPU where it counts, so the render/compute kernels port to deck.gl / MDV / SpatialData.js later rather than locking in.
Together these aren’t just engineering taste — they’re what make the toolbox FAIR: a self-describing op catalogue you can find, an open WebGPU runtime and open formats you can access, plain-array contracts that interoperate, and declarative graphs that record their own provenance so results are reusable.