Choose a scientific operation

Each page starts from the science, then maps it to optimized kernels. The same page is the shared specification for CUDA, Python MPS, native Swift/Metal, WebGPU, and the CPU reference.

Operation	Read this page when you need to	Main outputs
Data model and coordinates	interpret axes, shapes, units, regions, binning, and provenance	one cross-language contract
Load, decode, and bin	move compressed detector data into an accelerator-resident representation	resident 4D-STEM data and load provenance
BF, DF, and ADF	reduce detector counts with circular, annular, or arbitrary masks	scan-shaped scalar images
CoM, DPC, and iDPC	compute detector first moments, align the vector field, and integrate phase	two-component vector fields and phase
Single-sideband ptychography	reconstruct complex object information from overlapping diffraction information	object amplitude, phase, loss, aberrations
Explicit scan regions	intentionally analyze a half-open real-space subset	labeled scan subset
Display and export kernels	compute ranges, histograms, colors, FFT views, or encoded frames	presentation-ready derived buffers

Coordinate rule used everywhere

Array order is always

\[ (\text{row},\text{column}) \equiv (r,c). \]

For 4D-STEM data,

\[ I[R_r,R_c,k_r,k_c], \]

where \(\mathbf R=(R_r,R_c)\) is the real-space probe/scan coordinate and \(\mathbf k=(k_r,k_c)\) is the detector coordinate. A backend-specific launch order may differ internally, but its public inputs, outputs, masks, metadata, and parity artifacts must preserve this order.

Optimization rule used everywhere

Optimize transfers and traversals, not the scientific definition. Keep large arrays accelerator-resident; fuse compatible operations; reuse buffers; avoid per-batch synchronization; and return only the small result that the caller requested. Any crop, bin, mask, precision change, or cached result remains an explicit, provenance-recorded choice.