Repository navigation
STAR Suite v1.12.2 Release Notes
Date: 2026-10-27
This patch makes spatial R1 barcode decoding faster without changing its
results. It adds commit 2c7159e ("Reuse invariant work in spatial barcode
decoding") to v1.12.1.
The decoder previously repeated two calculations that do not change within a
loop:
- The N-base target scan prepared an edit-distance bit pattern for every
candidate target. It now prepares the observed span once per scan and
compares it with each target. Levenshtein distance is symmetric, and the
64-base truncation, length filter and literal treatment ofNare unchanged. - The raw-candidate alignment likelihood computed the base-quality error terms
in every dynamic-program cell. They depend only on the observed base and its
quality, so they are now computed once per row. Each cell keeps the original
operations and order, so likelihoods, including ties, are bit-identical.
Candidate enumeration, thresholds, ordering, policies and output manifests are
unchanged. No other module changes.
In development measurements on full Visium HD slides, wall time fell by 38.7%
on a colorectal slide and by 27.7% on an ovarian slide. Each figure comes from
one matched pair of runs, unmodified v1.12.1 against this change, on one host
with 20 threads and all four resolution policies written. They are not
benchmark results. Gains elsewhere depend on how much of a run is spent
decoding barcodes. (This paragraph was clarified on 2026-10-28, after the
release; the tagged copy states the same figures without the qualification.)
Validation:
- A new regression,
make -C core/legacy/source test_SpatialR1DecoderHotpaths,
is part ofmake test. It checks 151,072 likelihoods bit-for-bit against the
previous recurrence, 12,675 edit distances against an independent scalar
oracle, and 800 target-scan and cache cases. It also passes under UBSan. - All core test targets, the synthetic Solo smoke, all eight scRNA
deduplication modes and the spatial half-probe suite pass. The half-probe
suite's five reader, thread and spill variants agree in all 36 MEX files. - On the full colorectal and ovarian slides, all 36 MEX components of each
run match the frozen v1.9.5 reference checksums byte for byte, as the
matched v1.12.1 control does. All retained summary fields agree with the
control, apart from source revision, memory budget and timers.
STAR --version reports 1.12.2. Debian source packaging uses 1.12.2-1;
Ubuntu packages use 1.12.2-1~ubuntu22.04.1 and 1.12.2-1~ubuntu24.04.1.
Upstream STAR remains 2.7.11b, genome-index compatibility remains 2.7.4a
and legacy compatibility remains 2.7.1a. Existing indexes need no rebuild.