LANDISDisp: optional OpenMP threads for the spiral scan

[eliotmcintire]

Seed dispersal can now use several processor cores for its slowest step, making it about twice as fast with 4 cores. Results are identical no matter how many cores are used, and the default is still one core.

Affects: run time only, and only when more cores are turned on.

Technical details

The receiver scan in spiralLoopCpp (src/spiralSeedDispersal.cpp) is the one serial part that uses no RNG, so it now runs under OpenMP when options(LandR.LANDISDisp.threads = n) is above 1 (default 1) and a step has at least 20000 active receivers. It uses a static schedule with one hit buffer per thread, joined in thread order, so the receiver order matches the serial loop. The uniform draws, success test and compaction stay serial on the main thread, so output is bit-identical for any thread count and to useCpp = FALSE. OpenMP is guarded with #ifdef _OPENMP, and src/Makevars and src/Makevars.win add $(SHLIB_OPENMP_CXXFLAGS).

Verified with a new test-LANDISDisp-threads.R (expect_identical for threads 1, 2, 4 and the R reference, skipped without OpenMP), plus the parity, seedLocked (LANDR_SLOW_TESTS=1), wardScreen and pgv tests, all passing.

Timing on the xxxlarge fixture (9M cells, pgv supplied, verbose 0, median of 5): the whole LANDISDisp call takes 0.72 s on #255 and 0.63 s, 0.39 s and 0.35 s here with 1, 2 and 4 threads, so about twice as fast at 4 threads. The threshold made no measurable difference at 5000, 20000 or 100000, so 20000 is kept. An earlier version of this description reported a smaller gain; its benchmark timed spiralLoopCpp with a nested system.time(), whose default gcFirst = TRUE added a full garbage collection (~0.55 s) to every whole-call time.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CwcjqqK59FmTJscyi7xUqv