[glennhickey]

Summary

Makes minipoa a faster and, with the new options, as-accurate alternative to abPOA as cactus bar's base aligner. Used by ComparativeGenomicsToolkit/cactus#2026. This includes #1, merged in and ported to the new convex traceback, so merging this should mark #1 as merged too.

cactus bar, 8 Mb human/mouse, 8 threads	time	memory
before	29 s	3.1 GB
after, same settings and identical output	9 s	2.8 GB
after, cactus's new defaults (convex gap + guide tree)	12 s	4 GB
abPOA, for reference	51 s	5.5 GB

Evolver mammals (mafComparator), with cactus's new <minipoa> defaults: 0.846 / 0.646 → 0.854 / 0.652 (abPOA 0.853 / 0.652).

Speed, with identical alignments (0416daa, 89987e2)

• The horizontal-gap recurrence ran one column at a time and took 63% of poa(). It is now a vectorised decaying prefix-max scan, fused with the M/D pass, with a fast path for single-predecessor rows.
• DP scratch stays on the parameter handle between minipoa_msa() calls. A fresh buffer per call faulted the whole matrix in each time: 49 M page faults on the human/mouse run.
• The static band is centred on the edge-weighted mean position of the sequences through each node. The heaviest path's length strings together every sequence's insertions, which widened every row.
• The DP is memory-bound: at 8 threads it ran no faster than at 2. To cut the traffic:
  • Vertical gap states are stored as 16-bit saturating distances below M, which is exact. Open + extension penalties of 32768 or more are now rejected.
  • Horizontal gap states are stored as two bits per cell: "is the max" and "opened here".
  • Rows go to the matrices with non-temporal stores, while successors read a cache-resident ring of recent rows.

Accuracy (opt-in; cactus turns both on)

• An optional second gap piece makes the gap cost convex, as abPOA's is: minipoa_set_gap2(), and --gap_open2/--gap_ext2 on the command line. This adds a convex traceback.
• The guide tree skipped every minimizer found in more than a quarter of the sequences. Below 8 sequences that left every similarity at 0, so progressive mode kept the input order. Minimizers found in 16 or fewer sequences are now always used.

Robustness

• Fix traceback at the end row: "backtrack found no predecessor" #1's end-row fix (ofirr) is in both tracebacks. The traceback starts from every state of the end cell, not M alone, and a real query base scored against the end row's virtual N is emitted as an insertion.
• When the traceback still finds no path through the band, poa() retries that query unbanded instead of failing.
• mg_chain_backtrack() no longer reads t[-1] for an anchor with no predecessor.

Testing

• Identical output to the previous version: on simulated truth sets, on cactus bar's human/mouse output, and through cactus's evolver tests. SSE2, SSE4.1, AVX2 and emulated AVX-512 builds agree. ASan/UBSan are clean. An independent review checked the DP cell by cell against a scalar reference.
• The Fix traceback at the end row: "backtrack found no predecessor" #1 merge was compared against the previous commit on 203 runs: 29 datasets × 7 parameter sets, including bands down to 10 and seeding. All are identical.
• Fix traceback at the end row: "backtrack found no predecessor" #1's reproducer now aligns in its band with both gap models. With the single gap, the output is byte-identical to Fix traceback at the end row: "backtrack found no predecessor" #1's.
• Windows shaped like Fix traceback at the end row: "backtrack found no predecessor" #1's: a sequence, a short exact prefix of it, then an unrelated one. There were 150, in input order, since bar's windows are not always longest first. The previous version needed the unbanded retry on 10, 16 and 9 of them (single gap; narrow band; convex gap); this version needs it on none.
• Random prefix, suffix and internal fragments, in input order: 1072 of 1080 runs are identical to before. The other 8 are windows the previous version could align only through the retry.
• Every one of those MSAs was checked to spell out its input sequences.
• ASan/UBSan on the merged code are clean: a 200-run matrix (both gap models, bands down to 5, seeding, guide tree), plus the 150 Fix traceback at the end row: "backtrack found no predecessor" #1-shaped windows across four parameter sets.
• The CMake build gives the same 66 warnings as cactus-integration, and no new ones.

cactus builds minipoa only on x86 for now (it is off on ARM), so the ARM path is untested.

Please merge with a merge commit rather than squash/rebase, so the commits cactus pins stay reachable.

🤖 Generated with Claude Code