[AP][Timing] Added Basic Net Weighting

[AlexandreSinger]

Added basic timing awareness to the AP flow by weighting nets in the AP solver by their criticality (the max criticality of all edges through that net). This makes the solver try to minimize the length of nets that are more critical more than nets that are less critical (according to the pre-clustering timing analyzer).

Added a command-line option to trade-off between timing and wirelength in the AP flow.

[AlexandreSinger]

@vaughnbetz @amin1377 Good news regarding the timing-aware AP flow.

Results on Titan (fixed IOs, 1.3 * MinW):

	CPD	WL
No AP	1.000	1.000
AP tradeoff=0.0	0.994	0.952
AP tradeoff=0.5	0.963	0.950
AP tradeoff=1.0	0.999	0.937

Note: 1 circuit (sparcT1_chip2) failed to pack for the AP flow at tradeoff=0.0 and 0.5 due to a known issue with the IO blocks. Not all IO blocks were fixed, and the max candidate distance may have been set too low for IO blocks. For tradeoff=1.0, gsm_switch failed to route.

AP tradeoff=0.0 is the AP flow with no net weighting (it is what the AP flow would produce before this PR). With this new net weighting scheme (with a sensible tradeoff), we can see a 4.7% improvement in CPD and a 5% improvement in WL on Titan (with higher run time; this change did not improve run time)!

Increasing the tradeoff to be 1.0 makes you lose most of your gains. This is likely because the objectives of the global placer does not align with the full legalizer (0.75 tradeoff by default) or the detailed placer (0.5 tradeoff by default).

Per-circuit results for AP tradeoff=0.5, normalized to non-AP:

	CPD	WL	Run time
LU230_stratixiv_arch_timing.blif	0.979	0.947	1.189
LU_Network_stratixiv_arch_timing.blif	0.904	0.974	1.367
SLAM_spheric_stratixiv_arch_timing.blif	0.961	0.885	1.349
bitcoin_miner_stratixiv_arch_timing.blif	0.839	0.977	1.126
bitonic_mesh_stratixiv_arch_timing.blif	0.989	1.008	1.391
cholesky_bdti_stratixiv_arch_timing.blif	0.984	0.945	1.370
cholesky_mc_stratixiv_arch_timing.blif	0.928	0.948	1.351
dart_stratixiv_arch_timing.blif	1.086	0.915	1.438
denoise_stratixiv_arch_timing.blif	0.974	0.983	1.285
des90_stratixiv_arch_timing.blif	0.964	0.929	1.316
directrf_stratixiv_arch_timing.blif	0.941	0.888	1.120
gsm_switch_stratixiv_arch_timing.blif	1.008	1.053	1.393
mes_noc_stratixiv_arch_timing.blif	0.789	0.953	1.329
minres_stratixiv_arch_timing.blif	1.001	0.961	1.421
neuron_stratixiv_arch_timing.blif	0.998	1.005	1.730
openCV_stratixiv_arch_timing.blif	1.118	0.978	1.435
segmentation_stratixiv_arch_timing.blif	0.998	0.967	1.445
sparcT1_chip2_stratixiv_arch_timing.blif	Pack Failure	Pack Failure	Pack Failure
sparcT1_core_stratixiv_arch_timing.blif	0.898	0.876	1.513
sparcT2_core_stratixiv_arch_timing.blif	0.990	0.869	1.329
stap_qrd_stratixiv_arch_timing.blif	0.940	0.932	1.409
stereo_vision_stratixiv_arch_timing.blif	0.995	0.977	1.764
Geomean	0.963	0.950	1.376

As you can see, some of the circuits' CPD have gone up by as much as 12%, while WL remained mostly below 1.00. One circuit (gsm_switch) was worst in CPD and slightly worst in WL.

[vaughnbetz]

Excellent! Interesting result at timing trade off of 1.

[AlexandreSinger]

Excellent! Interesting result at timing trade off of 1.

Yeah I agree, I was not expecting the WL to improve while the CPD to get worse (yet still be better than non-AP). Perhaps the AP flow completely ignoring certain non-critical nets are causing the more critical nets to become shorter; however, this causes the originally non-critical nets to then become critical. This can probably be fixed by improving the accuracy of the pre-packed timing analysis using AP info.

[vaughnbetz]

LGTM

[amin1377]

Excellent! Interesting result at timing trade off of 1.

Not sure if it’s relevant, but for the default placement, in a set of benchmarks I tested a while ago, setting the timing trade-off to 1 actually resulted in worse CPD, while wirelength improved slightly.