Improved parallel router: add NetlistRouter

[duck2]

This adds a common NetlistRouter interface for the serial and parallel routers and removes the code duplication between route_timing.cpp and route_parallel.cpp. route_timing.* is divided into three files:

• route_net.* (net routing utils)
• route_utils.cpp (top-level routing utils)
• route.cpp (only the top-level route())

Unlike try_timing_driven_route_tmpl(), route() no longer manages a ConnectionRouter, so it doesn't need to be templated.

ConnectionRouter no longer retries with a full-device bounding box by itself. The "retry with full BB" flag is bubbled up and it's up to the NetlistRouter to retry the connection/net properly.

Will experiment with high-fanout behavior in ConnectionRouter and whether clipping HF BBs hurts QoR in the serial case. It's mandatory in the parallel case, so would be nice if it doesn't affect QoR and we won't have to add branches there.

How Has This Been Tested?

• CI tests
• Will attach QoR from Titan & VTR

Types of changes

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• My change requires a change to the documentation
• I have updated the documentation accordingly
• I have added tests to cover my changes
• All new and existing tests passed

[duck2]

It looks like there's a single min_chan_width_route_time failure and a segfault in a unit test... no idea why. Can update the golden result for the first failure, since that can vary wildly. Waiting for initial titan results.

[duck2]

titan_quick_qor, serial routing, placement seed=10 (both runs did placement to make sure we didn't break it):
titan_10_t1.xlsx

Metric	15c5340	This
max_vpr_mem	1	1.000
pack_time	1	0.991
placed_wirelength_est	1	1.000
place_time	1	1.001
placed_CPD_est	1	1.000
routed_wirelength	1	1.000
critical_path_delay	1	0.990
geomean_nonvirtual_intradomain_critical_path_delay	1	0.992
crit_path_route_time	1	1.014

[duck2]

Nightly koios, extracted from GH runners. No .xlsx file -- had to manually get the data from the artifacts because it looks like the parser isn't getting the CPD & runtime values. Koios runs are fixed chan width, so they should be parsed by something like vpr_titan.txt, not vpr_standard.txt.

Circuit	Runtime (8e59d59)	This	Ratio	CPD (8e59d59)	This	Ratio
bwave_like.float.small.v	98.37	98.06	0.997	7.849	7.563	0.964
dnnweaver.v	91.02	124.54	1.368	10.287	11.632	1.131
tdarknet_like.small.v	85.81	82.670	0.963	19.866	17.566	0.884
proxy.5.v	23.60	23.06	0.977	8.712	8.712	1.000
proxy.7.v	45.94	46.25	1.007	9.053	9.079	1.003
proxy.8.v	24.04	24.63	1.025	8.632	8.632	1.000
Geomean			1.048			0.994

[duck2]

VTR benchmarks extracted from GH runners: vtr_gh_t1.xlsx

circuit	Runtime (8e59d59)	This	Ratio	CPD (8e59d59)	This	Ratio	WL (8e59d59)	This	Ratio	W_min (8e59d59)	This	Ratio
arm_core.v	13.94	14.95	1.072	22.914	22.914	1.000	184307	184307	1.000	116	116	1.000
bgm.v	15.43	9.88	0.640	20.283	20.283	1.000	402334	402334	1.000	74	74	1.000
stereovision1.v	9.31	9.1	0.977	5.441	5.441	1.000	131011	131011	1.000	88	88	1.000
stereovision2.v	17.92	18.57	1.036	13.719	13.719	1.000	422372	422372	1.000	96	96	1.000
LU8PEEng.v	20.18	21.8	1.080	74.240	74.240	1.000	340308	340308	1.000	96	96	1.000
LU32PEEng.v	81.52	83.43	1.023	73.170	73.170	1.000	1348522	1348522	1.000	132	132	1.000
mcml.v	36.91	36.71	0.995	45.477	45.456	1.000	999026	997319	0.998	142	144	1.014
Geomean			0.962			1.000			1.000			1.002

[vaughnbetz]

Commenting changes and moving functions around.
QoR looks good; once these updates are made it's good to merge.

[vaughnbetz]

@duck2: is this ready to merge?

• there was a unit test seg fault --> resolved?
• serial QoR looks good.
• Any need to test parallel QoR, or just merge and test later?

[vaughnbetz]

@soheilshahrouz : the unit test failure is a NoC test that removes and adds a link (it seg faults). It's hard to see how Fahri's code could cause that so I suspect it's a pre-existing weakness. Can you take a look?

[soheilshahrouz]

@soheilshahrouz : the unit test failure is a NoC test that removes and adds a link (it seg faults). It's hard to see how Fahri's code could cause that so I suspect it's a pre-existing weakness. Can you take a look?

Is this test failure different than the one mentioned in the link below?
#2357 (comment)

[vaughnbetz]

Not sure. @ducks2?