GRT: improved NDR resources management during Fast Route incremental routing

[openroad-ci]

Summary

Nets with NDR can have different edge costs depending on the metal layer (3D). In this example for the sky130hd/hs PDK, the 2D edge cost is always 2 but the 3D edge cost varies from 1 to 2 for NDR nets.

Net: clk NDR cost in met1: 2  Edge cost: 2
Net: clk NDR cost in met2: 1  Edge cost: 2
Net: clk NDR cost in met3: 2  Edge cost: 2
Net: clk NDR cost in met4: 1  Edge cost: 2
Net: clk NDR cost in met5: 2  Edge cost: 2

The MazeRoute3D can create routes that are only possible during the 3D stage.
During incremental routing, the initial 2D state is already congested due to the NDR higher edge cost.
If FastRoute can't solve congestion, it triggers a 2D congestion error (even if 3D finds a solution).

Main Features

1. Soft NDR on 2D-3D congestion mismatch

• Instead of checking 2D overflow, checks 3D congestion for a more accurate signal during updateDirtyRoutes
• Disables NDR (demotes to soft-NDR, edge_cost = 1) for nets where the 2D vs 3D usage is mismatched at the end of FastRoute run
• Skips clock nets when rerouting nets involved in congestion to protect timing

2. NDR soft-disable when repair_antennas starts uninitialized

When repair_antennas starts from a state where GRT was never run (guides loaded directly from ODB):

• Fixes a uint16_t underflow bug caused by mismatched 3D add/release paths (areSegmentsRestored not set)
• Applies soft-NDR to nets occupying congested 2D edges using routes_ segments (real coords) instead of sttrees_ (which are empty in this flow), via the new disableCongestedNDRNetsFromRoutes / disableNDRNetsFromGridRoutes functions

3. Regression test loading guides from ODB for repair_antennas

• This is the first test to evaluate the execution of incremental routing starting from an uninitialized state.
• It would be interesting to include more tests like this in the future.

Type of Change

• Bug fix
• Improvements

Impact

Don't throw errors on false routing congestion (check for 3D route congestion).
Apply soft NDR to solve congestion during incremental routing.

Verification

• I have verified that the local build succeeds (./etc/Build.sh).
• I have run the relevant tests and they pass.
• My code follows the repository's formatting guidelines.
• I have signed my commits (DCO).

Related Issues

Underflow error after antenna repair

[gemini-code-assist]

Code Review

This pull request enhances the global router's ability to handle Non-Default Rule (NDR) nets during antenna repair, specifically when starting from ODB guides where Steiner trees are not yet initialized. It introduces logic to demote NDR nets to soft NDR (unit edge cost) if they pass through congested 2D edges, ensuring resource usage is correctly tracked in both 2D and 3D stages. Feedback focuses on improving the robustness of grid index calculations and segment traversal to prevent potential out-of-bounds accesses and to correctly ignore via transitions during congestion checks.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[maliberty]

Please explain how "Nets with NDR can have different edge costs depending on the metal layer (3D)."

[eder-matheus]

Please explain how "Nets with NDR can have different edge costs depending on the metal layer (3D)."

I'm not sure how common it is, but the create_ndr allows the user to set per-layer width and spacing values. So we can have an NDR that changes only a subset of the routing layers (e.g., changing the spacing of metal2, metal3 and metal4) while preserving the rules for the other layers.

We can also define different width/spacing ratios per layer. So we can have some metal layers with 2x the cost, while other layers with 3x, for example.

[jfgava]

@maliberty @eder-matheus I just found a bug in CTS when processing the spacing value to calculate NDR spacing. We are picking the wrong value for some layers on sky130... Other public and private PDKs were not affected.
I'll run a CI to evaluate the impact on sky130 designs... The edge cost will be 2 for 2D and 3D edges (more congestion).
But note that this issue with different edge cost can still occur with different NDRs (more spacing and width). I'll make a new unit test for this

[maliberty]

I think such NDRs are uncommon but at least I understand the issue. Usually people just say something like double-spacing for all layers. This is a case where doing it in 3d would win as you could cost each 3d edge differently.

[maliberty]

Merge?

[eder-matheus]

Merge?

Not yet. Jonas is checking CI one last time before merging it.