drt: preserve top-layer wrong-way tracks

[naveenvenk17]

Fixes #10873.

The regression came from dropping all track data above TOP_ROUTING_LAYER. The top routing layer uses coordinates from the next routing layer to create wrong-way edges and to build the maze x/y coordinate union. Removing those coordinates left valid routes without a path.

Keep the reduced z dimension, but collect tracks from the first routing layer above TOP_ROUTING_LAYER. That preserves the memory optimization while restoring the missing top-layer topology.

Tested on the GCP Linux VM:

• The supplied tiny_rocket/IHP130 replay passed the original 50% and 70% failure points without DRT-0255.
• Replay completed optimization iterations 0 and 1, then entered iteration 2 before the 20-minute cap.
• Peak reported memory was 1650 MB. The full-grid rollback reached about 1900 MB.
• All 26 //src/drt/test:all targets passed after rebasing onto current master.
• Bazel --config=lint build for //src/drt:drt passed.
• Stamped //:openroad build passed.
• clang-format and git diff --check passed.

[github-actions]

Welcome to OpenROAD! Thanks for opening your first PR.
Before we review:

• Contribution Guide: https://openroad.readthedocs.io/en/latest/contrib/contributing.html
• Build Instructions: https://openroad.readthedocs.io/en/latest/contrib/BuildWithCMake.html

Please ensure:

• CI passes
• Code is properly formatted
• Tests are included where applicable
  A maintainer will review shortly!

[gemini-code-assist]

Code Review

This pull request removes the logic that limits grid graph layers and fixed-shape costing to the TOP_ROUTING_LAYER threshold. Specifically, it removes the layer-skipping checks based on the grid graph's minimum/maximum layer numbers and the top routing layer configuration in both FlexDR_init.cpp and FlexGridGraph.cpp. There are no review comments, so I have no feedback to provide.

[maliberty]

@mguthaus FYI

[mguthaus]

Sounds like it's best to keep it removed. I'm not an expert in the DRT module and was just going for an easy reduction but obviously this isn't that.

[osamahammad21]

The regression is narrower than the grid needing the full stack. The top routing layer gets its wrong-way edges from the tracks of the layer above it, and those track coords also feed the maze x/y coordinate union. So trimming the upper layers out of initTracks deleted all wrong-way edges on the top routing layer plus some nodes below, which is what breaks the maze search here.

None of that needs grid z-planes above the top routing layer, and that's where the memory win comes from. So we can keep collecting tracks one layer above TOP_ROUTING_LAYER while still trimming the z dimension.

[naveenvenk17]

Thanks, this was the right split. I updated the patch to keep the reduced z dimension and collect track coordinates from only the first routing layer above TOP_ROUTING_LAYER.

The supplied replay now passes both original failure points, completes iterations 0 and 1, and enters iteration 2 before the 20-minute cap. No DRT-0255 appears. Peak reported memory was 1650 MB, compared with about 1900 MB for the full-grid rollback. I also rebased onto current master and reran all 26 DRT tests, the DRT lint build, and the stamped OpenROAD build.

[osamahammad21]

Easier to read and interpret.

Suggested change

	// Do not add z planes above TOP_ROUTING_LAYER. Keep tracks from the first
	// routing layer above it because initEdges uses those coordinates for
	// wrong-way edges on the top routing layer.
	const bool aboveTop = currLayerNum > router_cfg_->TOP_ROUTING_LAYER;
	if (aboveTop && collectedAboveTopTracks) {
	continue;
	}
	collectedAboveTopTracks = aboveTop;
	if (currLayerNum - 2 > router_cfg_->TOP_ROUTING_LAYER) {
	continue;
	}