[jfgava]

Problem

During the 2D stages, FastRoute checks that each NDR net fits in a single layer of every edge it uses. When no layer could host the net, Graph2D::getCostNDRAware added 100 × the net's edge cost to the shared edge usage. Every later NDR net on that edge was charged the same, even when it fit.

For example, an edge with three layers of 2 tracks each cannot host an NDR net of cost 3. That edge looked about 300 tracks over capacity, although it still had room for three regular nets. As a result:

• Regular nets were pushed off edges they fit in.
• The overflow iterations chased overflow that did not exist.
• The inflated usage could exceed the 2D usage limit, triggering the GRT-0228.

Implementation

• True usage. The shared usage always gets the net's real edge cost. The tracks an NDR net misses on its best layer are recorded as its NDR deficit on the edge (NDRUsage::deficit, Edge::ndr_deficit, plus a running total).
• Pricing for NDR nets only. FastRouteCore::getUsageNDRAware makes an edge with a deficit look overflowed by kNdrDeficitCost (100) tracks per missing track, but only to NDR nets. Regular nets see the true usage. It feeds the maze cost and the L, Z, spiral and monotonic pattern routes. The NDR nets on an edge share its deficit (getNDRDeficitIfRouted), so a net that fits also sees the nets it leaves without room.
• Rip-up. NDR nets are rerouted off edges with a deficit, even when the edge is below capacity.
• Overflow. The 2D overflow includes the total deficit, in both the incremental count and its full-scan reference. The iteration loop, best-solution tracking and soft-NDR demotion therefore still react to it. The NDR fallback after layer assignment also triggers whenever a deficit is left.
• Net order. StNetOrder adds 100 × an NDR net's own deficit to the net's congestion key. The key is now capped at the int16_t maximum instead of wrapping.
• Stale records. initNDRnets clears the per-edge NDR records of the previous run. Without this, a new net allocated at the address of a deleted one was taken for a net already using the edge.
• Cleanup. The multiplier, hasNDRCapacity, and the unused net name copy in getCostNDRAware are removed.

Tests

• New: NDR deficits. Two tests in the NDRDeficitTest fixture of TestOverflow:
  • the 3 layers × 2 tracks example with an NDR net of cost 3: usage 3, deficit 1, three regular nets still fit, 2D overflow 1;
  • two NDR nets sharing one edge's deficit.
• New: re-init. A TestGraph2D test checks that a same-size re-init drops the previous run's NDR records. It fails without the fix.
• Updated. AccountsForNDRDeficitsAndCancellation (previously AccountsForNDRPenaltiesAndCancellation), and few goldens.

Type of Change

• New feature

Impact

Better handling of routing resources when dealing with NDR nets during FastRoute 2D phases.

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 included tests to prevent regressions.
• I have signed my commits (DCO).

[gemini-code-assist]

Code Review

This pull request introduces the concept of NDR deficits to track and price Non-Default Routing (NDR) nets that cannot fit on any single layer of an edge, separating deficits from edge usage and integrating them into routing costs, rip-up checks, and net ordering. Feedback on these changes highlights two key improvements: first, in Graph2D::initNDRnets(), the multi-arrays should be resized before calling clearNDRnets() to prevent potential out-of-bounds accesses when grid dimensions change; second, in getUsageNDRAware(), using edge-specific capacities via getEdgeCapacity() instead of nominal global capacities would more accurately represent congestion on edges with reduced capacity.

[github-actions]

This PR exceeds the recommended size of 1000 lines. Please make sure you are NOT addressing multiple issues with one PR. Note this PR might be rejected due to its size.

[luis201420]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[jfgava]

@codex review

[jfgava]

The last commits addressed these comments from Eder:
PR #11963: grt: enhance FastRoute NDR handling
Modules: grt (FastRoute), test/ (Bazel flow metrics)

== Priority 1 (correctness) ==

src/grt/src/fastroute/src/graph2d.cpp:1171 bug (minor): an NDR deficit can gostale. A net's deficit is set once, in reserveNDRCapLayer, and only removedwith that net. When another net's layer is released, the remaining nets arenot re-checked. Example with tracks {2,2,3} and two NDR nets of width 3:A takes L2, B gets deficit 1. Rip up A: L2 has 3 free tracks again, butedge.ndr_deficit and the total stay at 1, so the 2D overflow counts 1.getNDRDeficitIfRouted(A) also returns 1, so A is priced off an edge it fitson. It heals when B is ripped on the next maze pass, or when copyBR rebuildsall usage. The old charged_overflow flag had the same staleness, but the 2Doverflow total now depends on the deficit. Suggest re-checking the remainingdeficit nets on the edge at release (the map is tiny). Test:graph.updateUsageH(1, 0, ndr_net_, -kNdrWidth); // the net that fits leavesEXPECT_EQ(graph.getTotalNDRDeficit(), 0);
src/grt/src/fastroute/src/FastRoute.cpp:2759,2769 bug: "An NDR deficit leftin 2D always becomes 3D overflow" isn't true. reserveNDRCapLayer packs netsfirst-fit, so a deficit can exist even when some layer assignment fits.Example: L0 = 3 tracks, L1 = 2 tracks. A (width 2) is added first and takesL0, so B (width 3) gets deficit 1, but A on L1 and B on L0 would fit. Iflayer assignment finds that, no net is demoted: disableNDRForCongestedNets(unchanged by this PR) only checks 2D usage overflow and 3D overflow.past_cong (getOverflow2Dmaze) still counts the deficit, so thepast_cong == 0 check fails and mazeRouteMSMDOrder3D is skipped for thewhole design. The old code had inflated usage on that edge, so the netswere demoted and the 3D maze still ran. Suggest leaving the deficit out ofthe 3D-maze check (or recomputing it from the 3D assignment), and fixingthe comment.
src/grt/src/fastroute/src/maze.cpp:2310,2392 suggestion: the full-scanreference adds the same running ndr_deficit_total_ as the incremental path,so the overflowcheck (GRT-904/905) can't catch the total drifting from thesum of Edge::ndr_deficit. Summing edge.ndr_deficit in the scan loops wouldmake it a real check.

== Priority 2 (QoR) ==

src/grt/test/soft_ndr_escalation.ok question: wirelength goes 1814 -> 2052 um(+13%), and NDR demotion now happens at the iteration limit (15) instead ofon stagnation (12). The PR body doesn't mention it, and the commit messagestill says 1872. Expected?
test/aes_sky130hd.metrics_limits, test/jpeg_sky130hd.metrics_limits question:DRT::ant::errors is raised 0 -> 1 on both. A commit message explainsjpeg_sky130hd, but nothing explains aes_sky130hd, and the PR body mentionsneither. The base .metrics files are stale (gpl metrics change even thoughthe flow's gpl uses RUDY, not FastRoute), so the metrics diff doesn't showthis PR's own effect. Could you compare before/after on the same master,through DRT? The 21-checkpoint replay only covers the first global_route.

== Priority 3 (testing) ==

Missing tests: the stale-deficit case above, and newRipupCheck /newRipupCongestedL ripping up an NDR net on a deficit edge that is undercapacity. No test calls either function.

== Priority 6 (process) ==

Some commit messages don't match the final diff: jpeg_sky130hs and theaes_sky130hs antenna limit are mentioned but neither changed, and the goldenwirelength numbers are out of date. Worth fixing, since merge commits keep them.

== Overall ==
Needs changes. Keeping the NDR penalty out of the shared usage is the right
fix, and the bookkeeping is easy to follow. Main ask: stop the deficit from
blocking the 3D maze when layer assignment resolved it, and explain the
soft_ndr_escalation wirelength and the raised antenna limits.

[luis201420]

Just to clarify, my comments are only suggestions and nothing blocking. The PR looks good to me.

[jfgava]

@osamahammad21 QoR metrics degraded for some flow tests due to more detours caused by fewer soft-NDR demotion. The next PR should fix this by restoring the routing state. No issues for public and private CI