NetCostHandler class

[soheilshahrouz]

This PR refactors net_cost_handler.cpp/.h files. A new class called NetCostHandler is added. All static variables in net_cost_handler.cpp are now member variables of NetCostHandler. Removing file-scoped variables is another step towards running multiple annealers at the same time.

[vaughnbetz]

Re-launching CI.

[vaughnbetz]

vtr_nightly_test5 failures all appear to be in 3D architectures with placement constraints (one big partition, left / right, and up/down partitions).
2024-08-16T02:27:57.3174001Z �[32;1m02:27:57�[0m | 3d_full_OPIN_inter_die_stratixiv_arch.timing/neuron_stratixiv_arch_timing Error: Executable vpr failed
2024-08-16T02:27:57.3182578Z �[32;1m02:27:57�[0m | full command: /usr/bin/env time -v /root/vtr-verilog-to-routing/vtr-verilog-to-routing/vpr/vpr 3d_full_OPIN_inter_die_stratixiv_arch.timing.xml neuron_stratixiv_arch_timing --circuit_file neuron_stratixiv_arch_timing.pre-vpr.blif --route_chan_width 300 --max_router_iterations 400 --router_lookahead map --initial_pres_fac 1.0 --router_profiler_astar_fac 1.5 --seed 3 --device neuron3d --sdc_file neuron_stratixiv_arch_timing.sdc --read_vpr_constraints one_big_partition.xml
2024-08-16T02:27:57.3187067Z �[32;1m02:27:57�[0m | returncode : 134
2024-08-16T02:27:57.3190615Z �[32;1m02:27:57�[0m | log file : /root/vtr-verilog-to-routing/vtr-verilog-to-routing/vtr_flow/tasks/regression_tests/vtr_reg_nightly_test5/vpr_tight_floorplan_3d/run001/3d_full_OPIN_inter_die_stratixiv_arch.timing.xml/neuron_stratixiv_arch_timing.blif/common_-sdc_file_sdc_samples_neuron_stratixiv_arch_timing_sdc_-read_vpr_constraints_tasks_regression_tests_vtr_reg_nightly_test5_vpr_tight_floorplan_3d_one_big_partition_xml/vpr.out
2024-08-16T02:27:57.3194301Z �[32;1m02:27:57�[0m | failed: Executable vpr failed (took 466.67 seconds, overall memory peak 2.82 GiB consumed by vpr run)

The return code is SIGABRT, which likely means a seg fault.

[AlexandreSinger]

Hi @soheilshahrouz , I really appreciate these kind of code refactorings! I think VTR can really improve with more classes like this!

I have left a few comments; however, I recognize that a lot of this code is not yours and you are just moving it from one place to another. So, feel free to ignore the comments if you think they are not necessary for what you are doing.

[soheilshahrouz]

vtr_reg_nightly_test3/vtr_reg_qor_chain_large

VTR Large	pack-time	place time	place WL	place cpd	place_mem
vtr-master	59.4	70.41168848	235102.554	20.72379515	678.4687356
vtr-pr	59.7	70.23339596	235102.554	20.72379515	675.5519583
	1.00	0.997467856	1	1	0.995700941

[soheilshahrouz]

vtr_reg_nightly_test2/titan_quick_qor

titan_quick_qor	pack-time	place time	place WL	place cpd	place_mem
titan-master	616.2897924	1259.088278	2495045.479	17.23294049	5334.326215
titan-pr	617.727757	1270.405134	2495045.479	17.23294049	5334.333385
ratop	1.00233326	1.008988135	1	1	1.000001344

[vaughnbetz]

Minor changes requested and some TODOs. Biggest one is 3D / cube_bb -- we seem to be missing the 3rd dimension computation. @amin1377 : is the layer part of cube_bb merged in?

[vaughnbetz]

TODO: make these line up with the edge of the grid so we get rid of some if statements (probably width-1 or width instead of width-2).
Also, we should have a 3rd dimension normalization somewhere using the same technique and it seems that variable should also go here.

[vaughnbetz]

We should have a 3rd term:
ncost += (layer_max-layer_min+1) * chanz_place_cost_fac

Big TODO!

[vaughnbetz]

QoR looks good -- good to merge.

[AlexandreSinger]

@soheilshahrouz @vaughnbetz The failed CI test is the one that I fixed in my most recent PR. I think this can just be merged.

[soheilshahrouz]

@vaughnbetz I applied your comments and replaced get_net_bb_cost_() and get_non_updatable_bb_() methods with functors initialized in the constructor. All CI tests passed and the PR is ready to be merged.