The crystal refinement (XtalOptimizer, both the seven-block and the reduced
beam+orientation form, and XtalOptimizerRotationOnly) no longer builds a
ceres::Problem. XtalRefine holds the problem as data and solves it with
LMSolver, which follows Ceres' trust-region LM step for step - Jacobi scaling,
damping and radius updates, stopping rules, box projection, the projected
Armijo line search with cubic interpolation on bounded problems, the
SphereManifold for the spindle - but takes J^T J and J^T r directly instead of
a Jacobian. The residual is the same XtalResidual code, now Ceres-free and
evaluated on a forward-mode Dual (Dual.h); everything that depends on
parameters alone (detector-angle trig, per-frame back-rotation, reciprocal
basis, orientation rotation) is worked out once per evaluation, and the
observed and predicted halves carry 6 and 9 derivative lanes rather than 16.
The sums are cut into blocks that depend on the residual count alone, so the
answer does not depend on the thread count. Because the line-search trial point
is the candidate point, a bounded iteration costs one evaluation instead of
Ceres' three.
Validation (rc174 + this, -march=x86-64-v3):
- p.mtz md5 identical to the Ceres build on myob/cytc/thau x10sa, GPU and CPU
builds, and on the lyso8 stills reference.
- Solve corpus (every 16-parameter solve and every 10th per-image solve of the
three sets, 8.3k problems, inputs and Ceres results dumped from a run that
reproduced the md5s): usable/failed agree on all, iteration counts identical
on all, parameters agree to <2e-11 (in px / rad / 0.01 A units), costs to
1e-13.
- Same process, same threads: 7-9x faster per solve than Ceres.
- In-run (GPU, loaded box): xtal 16-parameter solves myob 22.2 -> 6.4 core-s,
cytc 88 -> 26 core-s; per-image solves 5.4 -> 1.2 core-s (myob); cytc first
pass indexing windows 1.9 -> 0.85 s, myob 1.1 -> 0.45 s; solver share of the
whole cytc run 17% -> 4% of CPU samples.
New tests compare the solver with Ceres on synthetic rotation problems
(full/weighted/reduced) and check thread-count independence.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5K8jvPPbmCrbqnWkddTuB
