fix(gpu-router): dispatch-limit 2D grid + keep-best ripup rounds#669
Merged
Conversation
Two fixes from the M6 reproducibility investigation: 1. wavefront_batch dispatched (n*total)/256 workgroups in one dimension; full-board batches need ~129k against the 65535/dim wgpu cap, so the first dispatch poisoned the encoder and every subsequent GPU batch silently fell back to CPU (observed: 'proposed 7/83' on the 40-net subset, 72 dead batches per full CM5 run — every published M6 run was effectively CPU-bound). Split over a 2D grid; the shader flattens via num_workgroups. After: 'proposed 83/83', zero validation errors. 2. The greedy pass's rip-up loop kept regressing rounds (CM5: 504 -> 489 placed over a 46-minute round, damage kept). Snapshot before each round and restore on regression, matching the negotiation loop's keep-best. Plus a board_export example (gerbers/excellon/kicad/bom/pick-place from a routed .pcb.json). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
3 Skipped Deployments
|
Choji has used up today's reviewsYour team has reached the free plan's limit of 5 reviews per day, so Choji stepped aside on this pull request. That's a plan limit, not a reflection of the change itself. Your free reviews reset tomorrow. To have Choji review every push without waiting, upgrade to Pro for unlimited reviews. You're on the free plan · 5 reviews/day. |
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
From the M6 reproducibility investigation:
GPU dispatch limit (the big one):
relax_batchdispatched all workgroups in one dimension — full-board batches need ~129k against the 65535/dim cap, so the first dispatch poisoned the command encoder and every subsequent GPU batch silently fell back to CPU. Every published M6 run (including the 67-min one) ran with the GPU effectively dead after the opening dispatch. Fix: 2D dispatch grid, shader flattens vianum_workgroups. 40-net subset flips from 'proposed 7/83' to 'proposed 83/83', zero validation errors, 1.000 routability. Hardware parity tests pass.Keep-best rip-up rounds: the greedy pass's rip-up loop kept regressing rounds (504→489 placed over a 46-minute round on CM5). Now snapshots and restores on regression, same semantics as the negotiation loop.
Also adds a
board_exportexample (gerbers/excellon/KiCad/BOM/pick-place from a routed.pcb.json).Full-board CM5 measurement with both fixes is running; numbers to follow in a comment.
🤖 Generated with Claude Code