grt: implement CUGR net merge mirroring FastRoute path by sparsh-karna · Pull Request #10899 · The-OpenROAD-Project/OpenROAD

sparsh-karna · 2026-07-13T17:17:08Z

Summary

When GlobalRouter::mergeNetsRouting() fires for the CUGR path, this PR enables attempting to stitch the two existing route segments at the former buffer pin position (connectCUGRRouting) rather than ripping up and rerouting the survivor from scratch.

Key changes:

GlobalRouter::connectCUGRRouting(preserved, removed): Mirrors connectRouting() but uses cugr_->hasAvailableResources() instead of fastroute_->hasAvailableResources() for the capacity check. Reuses the existing geometry helpers findBufferPinPostions, findTopLayerOverPosition, and createConnectionForPositions, which are FastRoute-agnostic.
GlobalRouter::mergeNetsRouting() (CUGR branch): If stitching succeeds, calls cugr_->mergeNet() to transfer tree ownership and save guides. On failure, falls back to addDirtyNet() so the survivor is rerouted incrementally (matching the previous safe behavior).
CUGR::mergeNet(preserved, removed): Attaches the removed net's GRTreeNode subtree as a child of the preserved net's routing tree so getRoutes() emits all wire segments. Inserts the removed net into merged_nets_ so removeNet() knows not to decrement GridGraph demand for it (the wires are still physically there, just owned by the preserved net).
CUGR::removeNet() (merged-net guard): Skips removeTreeUsage() for nets that were transferred via mergeNet, preventing double-decrement of GridGraph congestion numbers.
CUGR::hasAvailableResources(layer, tile_x, tile_y): A thin wrapper over GridGraph::getEdge().getResource() >= 1.0, equivalent to FastRouteCore::hasAvailableResources().

Type of Change

New feature

Impact

Improves QoR (Quality of Results) and runtime during incremental routing by preserving existing routing topology during buffer removal ECOs when using CUGR, matching FastRoute's behavior.

Verification

I have verified that the local build succeeds (./etc/Build.sh).
I have run the relevant tests and they pass. (All 143 GRT regression tests pass.)
My code follows the repository's formatting guidelines.
I have included tests to prevent regressions.
I have signed my commits (DCO).

Related Issues

None.

gemini-code-assist

Code Review

This pull request implements support for merging nets in CUGR rather than always falling back to a full dirty-net reroute. It introduces connectCUGRRouting to stitch two CUGR routes together at the former buffer pin position and perform capacity checks against the CUGR GridGraph, and updates the CUGR class to handle routing tree ownership transfer during net merges. Feedback was provided to add an early return check if either route is empty to prevent a potential crash when finding the top layer over a position.

eder-matheus · 2026-07-13T17:22:39Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 98e1a41ad8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

eder-matheus

Please include unit tests in this PR. You can use the existing remove_buffer tests to create the new tests for the CUGR path.

sparsh-karna · 2026-07-22T17:30:51Z

Please include unit tests in this PR. You can use the existing remove_buffer tests to create the new tests for the CUGR path.

I used the already available remove_buffer1 and remove_buffer2 tests, and added the -use_cugr flag to check the new merging functionality

eder-matheus · 2026-07-22T20:58:53Z

@sparsh-karna It seems you have conflicts on the CUGR.cpp file. Please resolve

sparsh-karna · 2026-07-23T12:32:57Z

@sparsh-karna It seems you have conflicts on the CUGR.cpp file. Please resolve

I have rebased the branch on top of upstream/master and resolved the conflict in CUGR.cpp.

When GlobalRouter::mergeNetsRouting() fires for the CUGR path, attempt to stitch the two existing route segments at the former buffer pin position (connectCUGRRouting) rather than ripping up and rerouting the survivor from scratch. Key changes: GlobalRouter::connectCUGRRouting(preserved, removed) Mirrors connectRouting() but uses cugr_->hasAvailableResources() instead of fastroute_->hasAvailableResources() for the capacity check. Reuses the existing geometry helpers findBufferPinPostions, findTopLayerOverPosition, and createConnectionForPositions which are FastRoute-agnostic. GlobalRouter::mergeNetsRouting() -- CUGR branch If stitching succeeds, call cugr_->mergeNet() to transfer tree ownership and save guides. On failure, fall back to addDirtyNet so the survivor is rerouted incrementally (matching the previous safe behaviour). CUGR::mergeNet(preserved, removed) Attaches the removed net's GRTreeNode subtree as a child of the preserved net's routing tree so getRoutes() emits all wire segments. Inserts the removed net into merged_nets_ so removeNet() knows not to decrement GridGraph demand for it (the wires are still there, now owned by the preserved net). CUGR::removeNet() -- merged-net guard Skips removeTreeUsage() for nets that were transferred via mergeNet, preventing double-decrement of GridGraph congestion numbers. CUGR::hasAvailableResources(layer, tile_x, tile_y) Thin wrapper over GridGraph::getEdge().getResource() >= 1.0, equivalent to FastRouteCore::hasAvailableResources. All 143 GRT regression tests pass. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Remove the standalone connectCUGRRouting() function and fold the CUGR-specific capacity check (against GridGraph) into the existing connectRouting() function via a use_cugr_ branch. A new optional connection_out parameter receives the stitching segments when the caller needs to update the CUGR routing tree. The mergeNetsRouting() function now uses a single connectRouting() call. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Consolidate mergeNetsRouting() to use a single connectRouting() logic block for both FastRoute and CUGR paths, applying CUGR-specific capacity checks inside connectRouting() and performing the final CUGR tree merge conditionally after routing is connected. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Add error logging with ID 705 to prevent negative layer indices from failing silently in CUGR::hasAvailableResources(). Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Rename nodeA/B, target_layerA/B, target_xA/B, target_yA/B, and findPath to snake_case equivalent to comply with OpenROAD coding guidelines. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Mirrors the remove_buffers3 test (disjoint layer range merge) using -use_cugr flag to exercise the CUGR incremental net merge path. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

eder-matheus · 2026-07-23T18:18:58Z

@codex review

chatgpt-codex-connector

💡 Codex Review

OpenROAD/src/grt/src/GlobalRouter.cpp

Lines 5753 to 5760 in ad2a2c8

    
           if (max1 != -1 && max2 != -1) { 
        
             if (max1 < min2) { 
        
               insertViasForConnection(net1_route, pin_pos1, max1, min2); 
        
             } else if (max2 < min1) { 
        
               insertViasForConnection(net1_route, pin_pos1, max2, min1); 
        
             } 
        
           } 
        
           net1_route.insert(net1_route.end(), net2_route.begin(), net2_route.end());

Merge same-gcell CUGR trees before deleting the removed net

For CUGR merges where the buffer pins are in the same gcell, this branch updates only routes_; it never calls cugr_->mergeNet() or marks the removed net in merged_nets_. In that scenario, the following net-destroy callback removes the deleted net's existing tree usage from GridGraph even though those wires were just folded into the survivor's guides, so congestion accounting and later CUGR incremental routes are based on missing demand. Please transfer the CUGR tree/ownership in this branch as well, including any via bridge inserted above.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…ting In connectRouting(), when both buffer pins land in the same gcell the else-branch folded net2 route segments into net1_route but never called cugr_->mergeNet(). This left the removed net absent from merged_nets_, so the subsequent inDbNetDestroy callback subtracted its tree usage from GridGraph even though those wires had already been transferred to the survivor net, corrupting congestion demand for later incremental routes. Pass an empty connection vector to cugr_->mergeNet() so that: - merged_nets_ is updated, suppressing the erroneous removeTreeUsage - The removed tree is attached as a direct child of the preserved tree Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

In CUGR::mergeNet() the connection chain nodes (node_a -> ... -> node_b) were appended to the preserved tree without recording their demand in GridGraph. Subsequent calls to totalOverflow() / saveCongestion() and incremental routing therefore treated the stitched edges as free, and a later updateNet() or removeNet() of the survivor would subtract demand that was never added, resulting in negative edge usage. Fix: build a temporary mini-tree for the connection segments and call grid_graph_->addTreeUsage() on it (with the survivor net NDR costs) before attaching the real chain into the preserved tree. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

hasAvailableResources() always compared available resource against 1.0, but CUGR routes NDR nets with per-layer demand factors > 1 via GRNet::getNdrCosts(). A connection stitch could therefore be accepted on an edge with only 1 free track even though the NDR factor required more, causing congestion undercount and incorrect incremental routes. Changes: - Add demand parameter (default 1.0) to CUGR::hasAvailableResources(). - Add CUGR::getNdrCosts(odb::dbNet*) public getter so GlobalRouter can retrieve the survivor nets per-layer demand vector. - In connectRouting() CUGR capacity loop, look up the survivor nets NDR costs and pass the per-layer demand into hasAvailableResources() so NDR nets are checked against their true edge demand. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

sparsh-karna requested a review from a team as a code owner July 13, 2026 17:17

sparsh-karna requested a review from jfgava July 13, 2026 17:17

github-actions Bot added the size/M label Jul 13, 2026

gemini-code-assist Bot reviewed Jul 13, 2026

View reviewed changes

Comment thread src/grt/src/GlobalRouter.cpp Outdated

chatgpt-codex-connector Bot reviewed Jul 13, 2026

View reviewed changes

Comment thread src/grt/src/cugr/src/CUGR.cpp Outdated

eder-matheus self-requested a review July 13, 2026 17:29

sparsh-karna force-pushed the cugr-merge-net branch from 74660a9 to b2232b6 Compare July 13, 2026 17:47

eder-matheus requested changes Jul 14, 2026

View reviewed changes

Comment thread src/grt/src/GlobalRouter.cpp Outdated

Comment thread src/grt/src/GlobalRouter.cpp Outdated

Comment thread src/grt/src/cugr/src/CUGR.cpp

Comment thread src/grt/src/cugr/src/CUGR.cpp Outdated

eder-matheus requested changes Jul 15, 2026

View reviewed changes

Comment thread src/grt/src/GlobalRouter.cpp Outdated

Comment thread src/grt/src/GlobalRouter.cpp Outdated

Comment thread src/grt/src/GlobalRouter.cpp Outdated

sparsh-karna force-pushed the cugr-merge-net branch from 05adcb6 to 602f4f8 Compare July 23, 2026 12:30

sparsh-karna added 13 commits July 23, 2026 17:46

grt: add early return check for empty routes in CUGR net merge

d2b930c

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: preserve connection geometry in CUGR tree on merge

8cefb8e

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: run clang-format to fix format check violations

53efa9f

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: error on invalid layer index in CUGR hasAvailableResources

746f1a3

Add error logging with ID 705 to prevent negative layer indices from failing silently in CUGR::hasAvailableResources(). Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: fix snake_case variable names in CUGR::mergeNet

930cb0f

Rename nodeA/B, target_layerA/B, target_xA/B, target_yA/B, and findPath to snake_case equivalent to comply with OpenROAD coding guidelines. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: move CUGR mergeNet call into connectRouting

35214f7

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: apply merged net flags to both paths

76486de

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: unify fallback dirty net logic for mergeNetsRouting

b30628e

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: add remove_buffers integration tests for CUGR path

8cd749d

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

grt: regolden remove_buffers2_cugr segs after CUGR rebase

d265533

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

sparsh-karna force-pushed the cugr-merge-net branch from 23b3564 to d265533 Compare July 23, 2026 17:47

grt: add remove_buffers3 integration test for CUGR path

ad2a2c8

Mirrors the remove_buffers3 test (disjoint layer range merge) using -use_cugr flag to exercise the CUGR incremental net merge path. Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

chatgpt-codex-connector Bot reviewed Jul 23, 2026

View reviewed changes

Comment thread src/grt/src/cugr/src/CUGR.cpp Outdated

Comment thread src/grt/src/cugr/src/CUGR.cpp Outdated

sparsh-karna added 4 commits July 24, 2026 03:48

Fix GRT-1252 by properly building CUGR merge connection tree via DFS

b7ef516

Signed-off-by: Sparsh Karna <sparsh2005karna@gmail.com>

	if (max1 != -1 && max2 != -1) {
	if (max1 < min2) {
	insertViasForConnection(net1_route, pin_pos1, max1, min2);
	} else if (max2 < min1) {
	insertViasForConnection(net1_route, pin_pos1, max2, min1);
	}
	}
	net1_route.insert(net1_route.end(), net2_route.begin(), net2_route.end());

Uh oh!

Conversation

sparsh-karna commented Jul 13, 2026

Summary

Type of Change

Impact

Verification

Related Issues

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

eder-matheus commented Jul 13, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

eder-matheus left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

sparsh-karna commented Jul 22, 2026

Uh oh!

eder-matheus commented Jul 22, 2026

Uh oh!

sparsh-karna commented Jul 23, 2026

Uh oh!

eder-matheus commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants