MadSpin: -O2 for productions with >= 4 decaying particles

[oliviermattelaer]

Summary

MadSpin compiles its standalone matrix-element trees through misc.compile, which never sets GLOBAL_FLAG, so they were always built at -O0 (MG5aMC does the same). This PR builds them at -O2, but only when a production process has at least 4 particles that MadSpin decays, because that is the only case where it pays off in a full MadSpin run.

• MadSpin/decay.py
  • MS_GLOBAL_FLAG = '-O2', MS_GLOBAL_FLAG_MIN_DECAYS = 4.
  • set_global_flag(tree) writes GLOBAL_FLAG into the tree's Source/make_opts, which every makefile of the tree includes (DHELAS, MODEL, the P directories, MadSpin/src/makefile_*). output standalone has already built Source/DHELAS and Source/MODEL at -O0 and make does not track flags, so when the flag changes those objects and libraries are removed and rebuilt.
  • decay_all_events.nb_decaying_in_production() counts the decayed final-state particles of the production processes. Decays inside a chain (the W of t > w+ b, w+ > ...) do not count. It reads all_ME (madspin_v1) or the production entries of all_me (onshell/density modes).
  • use_global_flag() is called by both build paths: compile_fortran and decay_all_events_onshell.compile.
• madgraph/interface/common_run_interface.py: update_make_opts_full now returns whether it changed the file. No existing caller uses the return value.

Why the threshold

Full MadSpin runs on the rows of the MG7 paper's tab_madspin_walltimes.

• Setup: 10^4 LO events, spinmode madspin, unweighting sequential, nb_core 6, mg7 decay backend, fixed seed.
• Statistics: median of 3 runs, all code states interleaved, Apple M5 Pro.
• Code: main with PR mg7 1 -> N processes: decaying mass without parameter_dict, decay-aware observables, no histograms in MadSpin, 1->3 CI test #228's code changes applied to both states. Without mg7 1 -> N processes: decaying mass without parameter_dict, decay-aware observables, no histograms in MadSpin, 1->3 CI test #228 the mg7 backend cannot decay Z or H, so the WH, ZZ and t t~ Z rows fail.

row	decaying particles	-O0 [s]	-O2 [s]	speed-up
WH	1	23.6	25.1	0.94
t t~	2	27.5	29.0	0.95
t t~ (decay chain)	2	30.6	32.6	0.94
ZZ	2	28.3	30.5	0.93
t t~ Z	3	33.2	35.7	0.93
t t~ Z (decay chain)	3	37.8	39.0	0.97
t t~ t t~	4	41.2	38.5	1.07
t t~ t t~ (decay chain)	4	48.2	45.8	1.05
W+W-W+W-	4	103.6	82.6	1.25

• Fixed cost: -O2 adds about 2 s of compilation to every run.
• Run-time gain: the max-weight scan and decay loop only get faster when the matrix elements dominate them. W+W-W+W-: scan 30 → 20 s, loop 52 → 34 s. On the simple rows the scan and loop are dominated by Python and do not change.
• Fortran alone: the effect is larger. For g g > t t~ g g, GET_DENSITY goes from 476 µs per call at -O0 to 174 µs at -O2.

Validation

• tests/test_manager.py -p tests/unit_tests/madspin '.*': Ran 634 tests, OK. 6 are new:
  • TestSetGlobalFlag: the flag is written, stale objects are removed, and nothing is removed once the flag is already set.
  • TestGlobalFlagGate: the decay count from both data layouts, and the flag applied only from 4 decays on.
• Real density-mode MadSpin runs, both decaying 10000 events:
  • p p > t t~ (2 decays): make_opts keeps no GLOBAL_FLAG.
  • p p > t t~ t t~ (4 decays): built with GLOBAL_FLAG=-O2.
• Object check: in an (ungated) -O2 run, the objects MadSpin built (aloha_functions.o, P1_gg_ttx/matrix.o) are byte-identical to a manual -O2 compile and differ from -O0.
• Not bit-identical to before for these productions: -O2 against -O0 changes the density matrix only at rounding level, ≤5.4e-15 relative.

Notes

• I did not run a madspin_v1 case with 4 or more decays end to end; it takes hours. That path goes through the same helper, which the unit tests cover.
• A colour-sum rewrite of the standalone GET_ALL_INTER was also measured in this work and dropped. It speeds up the Fortran colour part by 2-10x, but makes no difference on these MadSpin rows because their colour matrices are small.