laue_index

The Python orchestration for LaueMatching, packaged as typed pipeline stages around the C/CUDA indexer (REFACTOR_PLAN). Mirrors the in-repo laue_torch conventions (curated __init__, single-responsibility modules, typed records) but stays independent of the paper-tied packages (laue_torch / laue_jax / jax_cpfem) — shared pure math is duplicated with # TODO(unify-after-publish).

Install

pip install 'laue-index[run]'                # everything the pipeline needs
pip install laue-index                       # library + the CPU indexer (numpy only)
LAUEMATCHING_CUDA=1 pip install laue-index   # + the CUDA binaries (needs nvcc)

laue-index fetch-db --dest ~/laue            # the 6.7 GB orientation database
export LAUEMATCHING_ORIENT_DB=~/laue/100MilOrients.bin
laue-index run process -c params.txt -i frame.h5 -n 8

The C indexer is compiled on your machine at install time (this is an sdist, not a wheel — a binary is tied to the toolkit and GPU architectures that built it). Without a C compiler and OpenMP the install still succeeds; only the binary is missing. The CUDA build is opt-in because a toolkit that cannot compile these sources fails at build time, which would take the working CPU binary down with it.

from laue_index import indexer
indexer.available()             # is the CPU indexer usable?
indexer.available("GPU")        # is LaueMatchingGPU usable?
indexer.binary_path("GPU")      # where it came from

Set LAUEMATCHING_BIN to a binary — or to a directory holding them — to use one you built elsewhere or downloaded from a release. It takes precedence over everything else, and the not-found error names every path it tried.

Modules

Relationship to scripts/

The orchestrators used to live in the repo's scripts/ and are now in pipeline/ here, so the package ships something that can actually run a frame. scripts/ keeps a one-line shim per entry point — python scripts/RunImage.py … and the shell pipeline behave exactly as before from a checkout. laue_stream_utils remains a thin re-export of this package's stages, so RunImage, laue_postprocess and laue_image_server are unchanged by any of it.

Testing

pytest (from the repo root) runs the unit suite — golden-anchored characterization tests pin behaviour through the refactor. The full end-to-end test (tests/test_char_e2e.py) is opt-in: set LAUE_E2E=1 with the orientation DB, the built C binary, and a prebuilt forward cache present; otherwise it skips, so CI is safe with SKIP_DOWNLOAD=1 (no 6.7 GB database needed for units).