sweep: do a sweep of routing layer adjustments

[oharboe]

No description provided.

[oharboe]

@maliberty @jeffng-or I thought grt::have_routes should have returned 0 here, i.e. that there are conflicts?

Run bazel run BoomTile_grt /tmp/grt gui_grt on this PR:

[oharboe]

Hmm... Is there a simple way to extract a single metric to put in a table for a sweep on how well the global route went?

02:36:03  Layer         Resource        Demand        Usage (%)    Max H / Max V / Total Overflow
02:36:03  ---------------------------------------------------------------------------------------
02:36:03  M1                   0             0            0.00%             0 /  0 /  0
02:36:03  M2            65833386       9983742           15.17%             1 /  0 / 12
02:36:03  M3            93444600       14160770           15.15%             2 /  2 / 13
02:36:03  M4            60500272       5722594            9.46%             3 /  0 / 87
02:36:03  M5            57725066       4516355            7.82%             1 /  1 /  7
02:36:03  M6            39750131       3263934            8.21%             3 /  1 / 172
02:36:03  M7            54848832       1635775            2.98%             2 /  2 / 53
02:36:03  ---------------------------------------------------------------------------------------
02:36:03  Total        372102287       39283170           10.56%            12 /  6 / 344

[oharboe]

@jeffng-or @maliberty Is the on the macros is a red herring? It has previously been able to deal with that. Does grt leave behind only impossible global routing DRC messages after 30 iterations or can it leave behind others as well as those it can't resolve?

[oharboe]

@maliberty @jeffng-or grt sweeps no longer make the server fall over. megaboom runs all variants up to and including cts in with default number of jobs, which makes as much as possible run in parallel and gives good utilization of the server. However, grt variants needs to be run in sequence of the server runs out of memory.

Finding: ROUTING_LAYER_ADJUSTMENT=0.3, 0,4, 0.5, 0.6 all fail.

Note, I'm assuming 0.6 is a dead build walking, being higher than 0.5 which is already known to fail. It is still running after 6 hours compared to the others finishing in 2 hours, so I don't hold out much hope.

Placement density is already pretty low, the design is flattened with macro placement from hierarchical synthesis.

Is this an interesting test case?

Thoughts?

Standalone global routing test case https://drive.google.com/file/d/1hf4l9pwqxPbzMsvLkr-URIdOXI6nZMXZ/view?usp=sharing

[oharboe]

ROUTING_LAYER_ADJUSTMENT=0.6 is still running, the others failed too. Nothing more to learn here, closing.