Linkcell neighbours and carried updates

Build the cutoff neighbour list, then keep rings when only part of the next frame moved. The scaling argument is Linkcell strong scaling. The list itself is Cutoff neighbour lists.

The cutoff list

neighListO builds one shared CSR with nneigh::orthoFullCsr when the OpenMP team has more than one thread and the caller is outside a team. The box is orthorhombic (length 3) or a restricted triclinic dump (length at least 6). On one thread, and when that CSR call is not taken, a linked linkcell fills the rows with linkcell::pairs_within. vesin is the fallback when linkcell is absent or the call fails. nneigh::preferSerialCutoff(true) forces the one-thread builder.

Published CHILL+ calls orthoFullCsr on those boxes (chill::chillPlusPublishedCutoff). Cutoff q4 and q6 do the same (chill::steinhardtQlCutoff). The ring census reads the atom-ID rows. The pair set does not depend on the thread count.

The half list keeps each pair on the lower index. The test that locks this is the Catch2 case half cutoff list is the lower index of the full list:

// excerpt: tests/test_neighbours.cpp
  REQUIRE(nneigh::orthoFullCsr(cloud, nullptr, 3.5, fullOff, fullCols, false));
  REQUIRE(nneigh::orthoFullCsr(cloud, nullptr, 3.5, halfOff, halfCols, true));

A skin reuses the wide list while every atom stays inside the Verlet shell. Bonds are the pairs still inside the cutoff.

Carried rings

A frame range on one worker visits frames in order through one FrameCarry. RingUpdater redoes only the sources near a changed bond. The result equals ringNetwork on that frame.

Carry is off for a single frame (--last omitted or equal to --frame) and whenever --rebuild is set. On a LAMMPS dump, --jobs above 1 also turns it off, because the frames are split across workers. A non-dump range keeps one carry when --jobs is above 1. More than one MPI rank on a dump leaves carry off as well.

The hop-ball test moves one atom across a periodic face and requires the carried rings to match a full search, with fewer sources recomputed than the atom count:

// excerpt: tests/test_ring_hop.cpp
  const auto &carried = updater.update(moved);
  const auto full = primitive::ringNetwork(moved, 6);
  REQUIRE(carried == full);

What was timed

Benchmarks records both windows at 4096 molecules, 200 transitions, and eight threads. Every transition stayed inside the hop ball, and the carried rings matched a full search. Speedup is rebuild divided by carry.

The mW window (10 fs step) at 0.1, 1, and 10 ps is on that page. The strict seeded graph there is the row seeded-strict.

The TIP4P/2005 ice/water slab (2 fs step) is finer: 0.01, 0.1, and 1 ps. An oxygen moves farther in one picosecond than an mW site. The seeded-strict speedups are 8.31, 3.28, and 2.59.

# excerpt: tests/cli_surface.py
        ["--frame", "1", "--last", "2", "--jobs", "1", "--rebuild", "read", cubic],

That argv is the walk with carry forced off. To keep the updater, drop --rebuild, pass a --last past --frame, and leave --jobs at 1 on a dump.