grt: FastRoute maze-routing performance (bit-identical) by saurav-fermions · Pull Request #10848 · The-OpenROAD-Project/OpenROAD

saurav-fermions · 2026-07-09T06:55:10Z

Summary

Two constant-factor optimizations to the FastRoute global router. Global-route guide output is byte-for-byte identical before and after.

Reuse 3D maze BFS scratch buffers — reuse two member scratch buffers (heapVisited, heapQueue) in addNeighborPoints instead of allocating net-node-sized local vectors on every call. Mirrors the existing maze2D visited_2D_/queue_2D_ reuse pattern.
Hoist current-cell state in mazeRouteMSMDOrder3D — hoist the popped cell's distance / path-length / direction out of the boost multi_array 3D grids into locals, removing ~5 redundant 3D index computations per popped node.

Testing

Built openroad; full grt regression suite passes (140/140, incl. golden log_compare). Guide output bit-identical to baseline.

Notes

Pure constant-factor optimizations; no routing behavior change.
The related O(N^2)->O(1) coordinate-dedup optimization for the same router is in a separate PR (grt: O(1) coordinate dedup in setTreeNodesVariables (FastRoute) #10838). No overlap with this PR.
No src/sta change; no cross-repo dependency.

addNeighborPoints allocated two local vectors (heapVisited, heapQueue) sized to the net node count on every call. On medium03 (98.6K cells) that is ~175K calls and ~125M node-slots of alloc+init churn per global_route, and the function is ~9% of the GRT profile. Reuse two member scratch buffers instead, clearing/resizing the used region each call rather than reallocating. This mirrors the existing maze2D visited_2D_/queue_2D_ reuse pattern. heapVisited is reset with assign(numNodes,false); heapQueue is write-before-read so only its size matters. Pure constant-factor change, no algorithmic difference: the global_route guide output is byte-identical (same MD5) before and after, and the grt regression suite stays 131/131. ~4.6% median global_route speedup on medium03 (10385ms -> 9910ms, median of 3 runs each). Signed-off-by: Saurav Singh <saurav.singh@fermions.co>

…bit-identical routing) The 3D maze Dijkstra inner loop in mazeRouteMSMDOrder3D re-read the popped ("current") cell's distance, path length, and direction from the boost multi_array 3D distance grids once per direction block -- up to six times per popped node. These values are loop-invariant during a node's relaxation: only the neighbor cells (tmpX/tmpY/tmpL) are written, never [curL][curY][curX] itself. Each read pays a full boost::multi_array 3D index computation. Hoist d1_3D_, path_len_3D_, and directions_3D_ for the current cell into locals (cur_d1, cur_path_len, cur_dir) immediately after removeMin3D, and use them in the six direction blocks. This removes ~5 redundant 3D index computations per popped node. On aes_nangate45 the 3D maze pops ~22.4M nodes per global_route, so the saved indexing is substantial. Pure constant-factor change, no algorithmic difference. cur_d1 is int and is promoted to float in `cur_d1 + cost + penalty` exactly as the original multi_array read was, so arithmetic is identical. Bit-identity proof (write_guides MD5, baseline bc114e8639 vs this change): gcd_nangate45 : da0e9f94e649c9f65443ce70d79ec7a9 (identical) aes_nangate45 : ccd6428a729cfea6c7228ac4d341c892 (identical, 156K cells) Timing (global_route only, interleaved base vs opt, median of 6, aes 156K cells): base 4790 ms -> opt 4469 ms (~6.7% faster, -321 ms) grt regression: 131/131 pass (suite includes guideok comparisons across many designs, an additional bit-identity check). Signed-off-by: Saurav Singh <saurav.singh@fermions.co>

gemini-code-assist

Code Review

This pull request optimizes 3D maze routing performance by reusing member scratch buffers (visited_3D_ and queue_3D_) to avoid per-call allocations, and caching current cell states to eliminate redundant 3D index computations. The reviewer suggests replacing std::vector<bool> with std::vector<char> for the scratch buffer to avoid bit-packing overhead and further improve performance.

gemini-code-assist · 2026-07-09T06:57:03Z

+  // Reusable BFS scratch for addNeighborPoints (3D). Mirrors the maze2D
+  // visited_2D_/queue_2D_ reuse pattern: allocate once, reset the used region
+  // each call instead of reallocating per call.
+  std::vector<bool> visited_3D_;


Using std::vector<bool> can be slower due to the overhead of bit-packing and bit-manipulation operations. Since this PR focuses on performance optimizations, using std::vector<char> (or std::vector<uint8_t>) instead of std::vector<bool> is a great opportunity to improve performance by allowing direct byte access.

Suggested change

std::vector<bool> visited_3D_;

std::vector<char> visited_3D_;

gemini-code-assist · 2026-07-09T06:57:03Z

+  visited_3D_.assign(numNodes, false);
+  queue_3D_.resize(numNodes);
+  std::vector<bool>& heapVisited = visited_3D_;
+  std::vector<int>& heapQueue = queue_3D_;


Update the type of heapVisited to std::vector<char>& to match the change of visited_3D_ from std::vector<bool> to std::vector<char> for better performance.

Suggested change

visited_3D_.assign(numNodes, false);

queue_3D_.resize(numNodes);

std::vector<bool>& heapVisited = visited_3D_;

std::vector<int>& heapQueue = queue_3D_;

visited_3D_.assign(numNodes, 0);

queue_3D_.resize(numNodes);

std::vector<char>& heapVisited = visited_3D_;

std::vector<int>& heapQueue = queue_3D_;

osamahammad21

revert the STA commit change. I assume this was accidental

saurav-fermions · 2026-07-09T22:35:09Z

@osamahammad21 good catch — the src/sta change was accidental (a stray submodule-pointer bump). Removed and force-pushed; the PR is now src/sta-clean.

On the gemini suggestion to switch visited_3D_ from std::vector<bool> to std::vector<char>: I've kept std::vector<bool> deliberately, because it mirrors the existing maze2D visited_2D_ buffer (same type) that this change is modeled on. This PR is about reusing the scratch buffer to cut allocations, not changing its element type; routing output is bit-identical either way, and switching only the 3D buffer would make it inconsistent with the 2D one. Happy to do a separate vector<bool>→vector<char> sweep across both if the team prefers.

saurav-fermions added 2 commits July 9, 2026 06:45

saurav-fermions requested review from a team as code owners July 9, 2026 06:55

saurav-fermions requested review from jfgava and osamahammad21 July 9, 2026 06:55

github-actions Bot added the size/S label Jul 9, 2026

gemini-code-assist Bot reviewed Jul 9, 2026

View reviewed changes

osamahammad21 requested changes Jul 9, 2026

View reviewed changes

saurav-fermions force-pushed the feat/grt-perf branch from 48a25fd to e317fc0 Compare July 9, 2026 22:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

grt: FastRoute maze-routing performance (bit-identical)#10848

grt: FastRoute maze-routing performance (bit-identical)#10848
saurav-fermions wants to merge 2 commits into
The-OpenROAD-Project:masterfrom
Fermions-ASI:feat/grt-perf

saurav-fermions commented Jul 9, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Jul 9, 2026

Uh oh!

gemini-code-assist Bot Jul 9, 2026

Uh oh!

osamahammad21 left a comment

Uh oh!

saurav-fermions commented Jul 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	std::vector<bool> visited_3D_;
	std::vector<char> visited_3D_;

Uh oh!

Conversation

saurav-fermions commented Jul 9, 2026

Summary

Testing

Notes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jul 9, 2026

Choose a reason for hiding this comment

Uh oh!

osamahammad21 left a comment

Choose a reason for hiding this comment

Uh oh!

saurav-fermions commented Jul 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants