[Place][InitT] Changed Default Init T Estimator to Equilibrium

[AlexandreSinger]

Ran both the equilibrium and variance initial temperature estimators and found that they both get similar quality; however, the equilibrium initial temperature estimator improves overall run time by 4-7% (up to 8% on the largest benchmarks). Setting it as the default for now.

Results on my test suites for both AP and No AP:

[AlexandreSinger]

This should be almost ready to merge. I will need to run the weekly tests since I am sure that those will have changed. Someone is currently running them; so I will wait until this evening.

[AlexandreSinger]

@amin1377 FYI

[amin1377]

Hi @AlexandreSinger,

I don’t mean to create extra work for you, but since you’re planning to change the default, I’d prefer to be a bit more cautious. If you have time, for SIV, could you please either show or share the spreadsheet containing the results so I can check the spread for CPD and WL?

Usually, when the average looks slightly worse, it suggests there might be some serious outliers. I’d like to see whether all circuits are around 1, or if there are any significant deteriorations or improvements.

Thanks!

[AlexandreSinger]

Hi @AlexandreSinger,

I don’t mean to create extra work for you, but since you’re planning to change the default, I’d prefer to be a bit more cautious. If you have time, for SIV, could you please either show or share the spreadsheet containing the results so I can check the spread for CPD and WL?

Usually, when the average looks slightly worse, it suggests there might be some serious outliers. I’d like to see whether all circuits are around 1, or if there are any significant deteriorations or improvements.

Thanks!

These are the results on AP using Equilibrium init T estimator normalized to the variance one on Titan.

circuit	total_wirelength	crit_path_delay	ap_dp_runtime
LU230_stratixiv_arch_timing.blif	0.98	0.98	0.87
LU_Network_stratixiv_arch_timing.blif	0.99	0.85	0.81
SLAM_spheric_stratixiv_arch_timing.blif	1.04	1.01	0.83
bitcoin_miner_stratixiv_arch_timing.blif	0.96	0.95	0.69
bitonic_mesh_stratixiv_arch_timing.blif	1.06	1.02	0.82
cholesky_bdti_stratixiv_arch_timing.blif	0.96	1.04	0.79
cholesky_mc_stratixiv_arch_timing.blif	1.00	1.02	0.83
dart_stratixiv_arch_timing.blif	0.99	1.06	0.86
denoise_stratixiv_arch_timing.blif	0.95	0.98	0.75
des90_stratixiv_arch_timing.blif	1.04	0.95	0.78
directrf_stratixiv_arch_timing.blif	0.94	0.93	0.72
gsm_switch_stratixiv_arch_timing.blif	1.04	1.08	0.89
mes_noc_stratixiv_arch_timing.blif	1.02	1.13	0.68
minres_stratixiv_arch_timing.blif	1.00	0.94	0.80
neuron_stratixiv_arch_timing.blif	0.99	0.97	0.85
openCV_stratixiv_arch_timing.blif	1.03	1.03	0.84
segmentation_stratixiv_arch_timing.blif	1.02	1.01	0.84
sparcT1_chip2_stratixiv_arch_timing.blif	0.97	1.15	0.66
sparcT1_core_stratixiv_arch_timing.blif	1.00	1.03	0.99
sparcT2_core_stratixiv_arch_timing.blif	1.05	0.97	0.78
stap_qrd_stratixiv_arch_timing.blif	0.85	1.06	0.81
stereo_vision_stratixiv_arch_timing.blif	1.02	1.02	0.97
gaussianblur_stratixiv_arch_timing.blif	1.03	0.98	0.77
GEOMEAN	1.00	1.00	0.81

QoR is identical on average, with around a 19% reduction in detailed placement time.

Here is a colorful picture:

AS we can see there are a couple of outliers for CPD, but their runtimes improved by a huge amount. I think this is overall a good change.

[AlexandreSinger]

@vaughnbetz FYI

[amin1377]

Thanks @AlexandreSinger. I think the results for AP make sense. If you can also include the results for the default flow, that would be great.

[AlexandreSinger]

Default flow:

circuit	total_wirelength	crit_path_delay
LU230_stratixiv_arch_timing.blif	0.97	0.99
LU_Network_stratixiv_arch_timing.blif	1.01	1.06
SLAM_spheric_stratixiv_arch_timing.blif	1.03	0.98
bitonic_mesh_stratixiv_arch_timing.blif	1.00	1.02
cholesky_bdti_stratixiv_arch_timing.blif	1.01	0.88
cholesky_mc_stratixiv_arch_timing.blif	1.02	1.02
dart_stratixiv_arch_timing.blif	1.00	1.00
denoise_stratixiv_arch_timing.blif	1.01	1.00
des90_stratixiv_arch_timing.blif	0.97	0.89
directrf_stratixiv_arch_timing.blif	1.05	1.09
gsm_switch_stratixiv_arch_timing.blif	0.99	1.00
mes_noc_stratixiv_arch_timing.blif	0.98	1.26
minres_stratixiv_arch_timing.blif	0.98	1.06
neuron_stratixiv_arch_timing.blif	0.94	1.10
openCV_stratixiv_arch_timing.blif	0.98	1.00
segmentation_stratixiv_arch_timing.blif	0.99	1.00
sparcT1_chip2_stratixiv_arch_timing.blif	0.99	1.12
sparcT1_core_stratixiv_arch_timing.blif	1.00	1.09
sparcT2_core_stratixiv_arch_timing.blif	0.94	0.86
stap_qrd_stratixiv_arch_timing.blif	1.03	1.00
stereo_vision_stratixiv_arch_timing.blif	1.00	1.01
gaussianblur_stratixiv_arch_timing.blif	1.00	1.00
GEOMEAN	1.00	1.02

[AlexandreSinger]

Weekly tests running now: https://github.com/verilog-to-routing/vtr-verilog-to-routing/actions/runs/18299908918