mpl: stop annealing once the best result stops improving

[oharboe]

Opened for discussion.

Every annealing run in mpl took its full 2000 steps. A run now stops when its best result has not improved for a fifth of its budget (400 steps).

mpl on 18 ORFS designs at each design's own core and settings, measured with mpl_bench (#11809): median of 3 runs per arm, 8 threads pinned to 8 cores, same machine. Run-to-run spread was up to 7% of the median; read differences below 10% as noise.

	result
mpl runtime	1.26x geometric mean; 9 designs faster by 10% or more, 9 within noise, none slower
macro placement	identical on 12 designs; on 6 the macro-net wirelength moved, by 0.24% or less except asap7/riscv32i-mock-sram (10.9% shorter)

Per design

design	master s	this PR s	speedup	macro-net HPWL change
asap7/aes-block	2.85	2.25	1.27x	same
asap7/coralnpu	2.30	2.07	1.11x	same
asap7/cva6	40.29	39.54	1.02x	same
asap7/riscv32i	5.49	5.17	1.06x	+0.24%
asap7/riscv32i-mock-sram	7.68	7.46	1.03x	-10.86%
asap7/swerv_wrapper	80.18	77.88	1.03x	-0.00%
asap7/tinyRocket	12.81	11.97	1.07x	same
gf180/uart-blocks	0.85	0.18	4.84x	same
nangate45/ariane133	8.32	5.34	1.56x	-0.02%
nangate45/ariane136	9.05	7.63	1.19x	+0.01%
nangate45/black_parrot	5.46	4.53	1.20x	same
nangate45/bp_be_top	4.77	4.34	1.10x	same
nangate45/bp_fe_top	11.32	10.67	1.06x	same
nangate45/bp_multi_top	5.52	4.42	1.25x	same
nangate45/cva6	28.26	27.57	1.03x	-0.00%
nangate45/mempool_group	15.87	15.46	1.03x	same
nangate45/swerv_wrapper	44.95	44.01	1.02x	same
nangate45/tinyRocket	1.23	0.56	2.21x	same

The patience was chosen by replaying the improvement traces of every annealing run mpl made on 18 ORFS designs (asap7, gf180, nangate45) against the full budget. An early stop changes nothing before it fires, so the replay is exact:

runs	steps run	cost loss: median	p95	worst
cluster placement (179)	59%	0.00%	0.01%	0.08%
macro placement (218)	30%	0.00%	0.00%	0.00%
tilings (264)	26%	0.00%	0.00%	9.62%

A tenth of the budget (200 steps) cost up to 9.85% on a cluster placement and 20% on a tiling. The one tiling outlier loses the same 9.62% at every patience from 300 to 1000 steps.

QoR at global route, one ORFS build per arm through GRT (the full study series, whose macro placements equal this PR's on all 18 designs, so the moves are this change's): 12 identical, 2 with positive slack in both arms, and 4 whose worst slack moved: nangate45/ariane133 +87.7 ps, asap7/swerv_wrapper +20.5 ps, asap7/riscv32i -0.7 ps, nangate45/cva6 -5.8 ps. One build per arm; a different placement moves slack either way.

How these were measured, and the rest of the study: The-OpenROAD-Project/bazel-orfs#1290. The bench is offered separately in #11809.

Tests

No new test: the change is a stopping rule, and its effect is runtime and the quality numbers above, which a unit test would only pin down as golden output. The existing regressions pass unchanged.

Claude made the changes and stared at paint drying.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces an early stopping mechanism to the simulated annealing loop in SimulatedAnnealingCore::fastSA(). It defines a patience threshold of one-fifth of the maximum steps and terminates the loop early if the best result has not improved within this window, saving computation time when the temperature is too low to escape local minima. There are no review comments, so I have no feedback to provide.