API and source reference

The website explains the algorithms and execution model. Rust documentation and source provide exact signatures, defaults and trait contracts for your selected revision. Use both: a function name alone does not say whether its arguments are bounded windows or complete resident images.

Build documentation matching the application

Build the core API from the repository revision used by your application:

cargo doc --release --no-deps --open

For the filesystem image path, include the Zarr feature:

cargo doc --release --no-deps --features zarr --open

Feature-gated modules appear only when those features are enabled. Select the backend combination used by the relevant example rather than assuming a build with every optional model/device dependency is a useful reference. Source links on this website follow main; a pinned application can have a different API.

Follow one pipeline through the public layers

Question Entry point What the interface owns
How do I compose work? PlanBuilder, Chain Phase assembly and operation composition
How do I name an image? ImageId, Phase Supplied-input and produced-image identities
What can an image operation request? BlockOp, SourceInput Input/output types, reach, sources and workspace contracts
How do I emit evidence or reduce it? FragmentOp, BlockView, PhaseView Fragment dependencies, output ownership and barrier state
How is repeated state expressed? IterativeOp, SubstageOperand Running/fixed/source operands and per-substage dependencies
What geometry gets recorded? BlockGrid, BlockGeometry, Decomposition Cores, reads, trustworthy output and the binding layout
How does work run? execute_phases, Hints Execution under the plan with advisory runtime choices
Where do image bytes come from? Environment, ZarrEnvironment Declared block reads/writes and storage behavior

A practical reading order is assembly → operation contract → geometry → execution → storage. Reading only a kernel implementation can miss why the planner rejects a dtype or retains an extra image. Reading only the scheduler can miss that an operation’s whole-axis dependency already fixes how it may be partitioned.

Three kinds of algorithm entry point

A resident numerical helper is useful for small arrays and independent tests. It generally assumes the caller has already supplied all needed data. A planned wrapper declares how to obtain the required windows and commit owned output. A specialized solver can have a separate execution interface when a generic image-to-image phase would hide its real state or synchronization requirements.

For example, finite-PSF deconvolution has region source/state interfaces and a disk-backed solver. It is distinct from both the older fixed-count spatial operation and the generic iterative executor. The calibration and restoration catalog explains their numerical and memory differences.

Read contracts at their definitions

The public module map and operation module index are useful starting points. For a named type, follow its inline source link to the declaration and then read the relevant implementation methods. Trait defaults are significant: BlockOp::accepts defaults to f64, and default behavior is not an implicit promise to accept every dtype the storage layer can decode.

When adding an operation, identify its independent numerical kernel, planned wrapper, resource accounting and seam/whole-image test. Those pieces explain more than a constructor signature by itself. The correctness page describes the checks that make an operation safe to cut and schedule.