- GPU_DATA_ANALYSIS: rotation scaling's GPU/CPU decision before the card is touched (sort and
scaling at 80% of the card's total) and the two-pass ingest; FFT indexer batches of 4096
directions and indexers released after a pass's last indexing; beam-centre FFT on the CPU above
half of the card; model validation and MR on one 80% budget (the scale fit's "eighth of the card"
rule removed); CUDA profile fit in the CPU engine's single-precision arithmetic; Metal engine not
used under --background-trim.
- RUGNUX_INSTALL: the scaling of a set too large for the card now goes to the CPU by itself (it no
longer stops the run), the up-front decisions listed, the rigid-body CPU fallback removed; Bragg
integration on the Apple GPU on macOS.
- CPU_DATA_ANALYSIS_IMAGE: beam-stop rings of constant 2theta and the tilted detector's solid angle
divided out (solid angle no longer described as constant along a ring); beam-centre FFT half-card rule.
- CPU_DATA_ANALYSIS_DECISIONS / RUGNUX_REPORT: MR's paired-residual tie rule; MR CPU transform FFT
threaded independently of the thread count; GPU_COUNT on macOS.
- JFJOCH_VIEWER: processing starts from a bare rugnux's defaults, Known unit cell unticked on open
and when a run is shown, MX jobs with rugnux's azint defaults, raw sweeps' runs shown from
_process.h5, Copy command omits automatic -N/--max-spots/--integration-radius; Metal on macOS.
- RELEASE_CONTENTS: macOS artefacts have no CUDA but integrate on the Apple GPU.
- RUGNUX, RUGNUX_OVERVIEW, RUGNUX_TUTORIAL, RUGNUX_ADVANCED: a stationary goniometer axis is stills
(and gets the stills --refine-geometry default); -N worker ceilings on a small card; --no-gpu also
hides the Apple GPU.
- tools/battery/README: the timing benchmark's cytochrome c set is no longer UNDECIDED.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PXc1zD1yxHqaR1DGdCTHL4
leonarski_fandClaude Opus 5.5
9293757853
