Skip to content

Integration fixes and dual-PDK requalification on the native GDS

Stream b0, branch feat/rtl2gds-b0 (from feat/rtl2gds b813c34), 2026-10-24. Scope is block/macro level: no pad ring, no seal ring; IR/EM/SI/OCV/CPPR are unchecked; density is reported separately and is never folded into DRC. Third-party tools ran only as black boxes (signoff image 4c48415d8bd0: Magic 8.3.678, Netgen 1.5.323, KLayout 0.30.9; OpenRCX image f152917ebc45; OpenSTA 3.1.0 host binary). No reference source was read.

Evidence (all tracked): docs/evidence/requalification.json, docs/evidence/gds-verification.json, docs/evidence/def-segmentation.json, docs/evidence/extraction-requalification.json, docs/evidence/timing-baseline.json.

What changed

  1. export-def writes maximal collinear runs. The router stores one record per lattice edge; the DEF used to write one NEW path per record, each with its own half-width end extension. The new bounded helper rustroad_core::wire_runs::collinear_runs merges same-net, same-layer records that touch or overlap on one track line and splits a run only at a strictly interior via landing (lower or upper metal), pin-access point or same-layer branch. Vias stay singleton paths and patches stay RECT statements. Every via and branch remains a path end, which is what endpoint-based readers need.
  2. export-gds streams patches and writes the same runs. Routes.patches were dropped; they are now rectangles on the routing layer's net purpose. Wires use the DEF runs. Every used LEF via variant already had its own VIA_<name> structure (Sky130 VIA_M1M2_PR_MR; GF180 VIA_Via1_VH, VIA_Via2_HH, VIA_Via2_HV, VIA_Via2_VH); the XOR confirms them. KLayout's DEF reader does draw routing RECTs in NETS, so the reference streamout covers patches without an overlay: on the same routed workspace, the b813c34 native GDS (no patches) differed from the reference only on the patch layers, and the new native GDS does not differ. verify_gds.py keeps this as a permanent control.
  3. extract includes patches. A patch adds ground, plane and lateral coupling capacitance on all four edges of its pieces not already covered by same-net metal, attached to the node of the same-net conductor it overlaps (no new node or resistor). Details: docs/extraction.md (physical model step 5). Workspaces without patches extract byte-identically to b813c34.
  4. Harness fixes found while requalifying:
  5. scripts/verify_gds.py required the old router DRC baseline. It now requires Magic 0 and KLayout 0, adds a planted sub-minimum square (both decks must count it), checks that every DEF RECT patch is streamed on its net purpose, and deletes those rectangles as a control (the XOR must fail). Its basic KLayout LVS check stops at "Netlists match": deleting the wire of port z[1] still read match on both PDKs (a labelled net without devices is left out of that comparison). It now also runs both LVS checks of scripts/signoff, which catch it (top pin count 11 != 12, z[1] unpaired); the deleted-wire control requires Netgen and the signoff KLayout check to fail. The victim is the largest signal wire (a DEF NETS path), not a supply spine. The missing-cell control picks a cell the producer actually used. The odd-GF180-coordinate plant is now refused by the workspace reader (INVALID_WORKSPACE, LEF-rule via enclosure / supply via grid) before streaming; the control records which layer caught it and that no file was written.
  6. scripts/verify_router_signoff.py assumed the first path of the first net is an access-layer wire; it now picks the first vertical access-layer wire of at least four widths.
  7. scripts/compare_spef.py merge-def failed on DEF RECT lines; they are now kept verbatim (and checked identical). campaign --sta-bin DIR selects the RustSTA binaries.
  8. New: scripts/def_segmentation.py (old vs new DEF proof), scripts/timing_baseline.py (multi-corner baseline), scripts/requalification_evidence.py (evidence assembly).

The stored routing state (routing.rbin, schema 5) is not compacted. The lattice records carry the junction structure that validate() and the extractor's graph use; merging at export gives the identical DEF/GDS metal with no schema, mirror or old-revision risk.

Geometry proof for the merged DEF

scripts/def_segmentation.py on the final flow workspaces compares OLD = export-def of the b813c34 binary with NEW = the flow's export-def (same routed revision). It checks that the text outside NETS, net headers, vias and RECT patches is identical; that KLayout reference streamouts of both are XOR-identical on all layers (--layer-details); that the text labels (layer, string, position) are identical; that deleting the longest NEW path is detected by the XOR (control); and that merge-def of OLD and NEW gives the same paths. It also runs OpenRCX on both: the per-edge OLD DEF made the OpenRCX reference depend on path segmentation, which is why the flow's DEF now writes maximal runs. The receipt is docs/evidence/def-segmentation.json.

Requalification (both PDKs, integrated tools)

Flow: scripts/verify_routing.py, policy producer (RustSys c5ec23c, 3de13b3a…) with fixtures/flow/cell-policy.json, signoff manifest, RustSTA 0.4 (rusteda c8f71b63…, ruststa ac76f0f3…), RustROAD ff523a9d… (built from this branch). Pass 1 used --signoff-reference; pass 2 used --signoff-gds <pdk>=<pass 1 native export-gds>. Both passes produced byte-identical DEF; pass 2's workspaces export the same GDS bytes and the same native SPEF bytes (min/nom/max) as pass 1.

Pass 2 also feeds scripts/verify_router_signoff.py (planted notch and sub-area controls, swapped LVS pins, a placement-only copy, pin-access test layouts) and scripts/verify_gds.py (read-only, byte-determinism including an independent copy, XOR and labels, verbatim cell bytes, patches streamed, planted DRC, four LVS verdicts on native and reference, negative controls). The per-check verdicts of this stream are in docs/evidence/requalification.json and docs/evidence/gds-verification.json, kept as historical evidence for reproducibility, not as current results.

The Sky130 clock-input antenna item found in this stream was handed to the antenna ECO stream (docs/hold-closure.md); its record is antenna_eco_handoff in docs/evidence/requalification.json.

Post-integration timing baseline

scripts/timing_baseline.py times pass 2 at tt+nom, ss+max and ff+min with the unmodified PDK Liberty and the flow's SDC, on the native SPEF from rustroad extract and on OpenRCX SPEF of the flow's merged DEF. One RustSTA 0.4 workspace per SPEF source holds all three corners (--require-wire-model spef_rc, OpenSTA-compatible pulse rules); OpenSTA 3.1.0 runs the same inputs per corner. The parity gate is 1 ps per endpoint. The per-endpoint hold_fix rows (required delay, net, driver, worst path stages) for every corner and source are in docs/evidence/timing-baseline.json; hold closure is docs/hold-closure.md.

Resources (Rule 9)

Estimates before running: the native commands are small (export-gds < 32 MiB); RustSTA three-corner import < 256 MiB; containers capped at 4 GiB (signoff) / 2 GiB (OpenRCX), one at a time per harness; no run above 4 GiB, so no scale lock.

Reproduction

free -h
export PATH=/home/cgao/.cargo/bin:$PATH
systemd-run --user --scope -p MemoryMax=8G -p MemorySwapMax=2G --setenv=PATH=$PATH \
  timeout 1200s cargo build -j 4 --release --workspace
STA=/mnt/data2/cgao/git/RustROAD/.work/sta-signoff/bin
python3 scripts/verify_routing.py --workspace .work/runs/r1 --base-rev empty \
  --rustroad target/release/rustroad --rusteda $STA/rusteda --ruststa $STA/ruststa \
  --producer policy --signoff-reference
for p in sky130 gf180; do target/release/rustroad --workspace .work/runs/r1/${p}_physical \
  export-gds --cell-gds <pdk cell gds> --layer-map fixtures/pdk/$p-gds-map.json \
  --out .work/runs/r1-native/$p.gds; done
python3 scripts/verify_routing.py --workspace .work/runs/r2 --base-rev empty \
  --rustroad target/release/rustroad --rusteda $STA/rusteda --ruststa $STA/ruststa \
  --producer policy --signoff-gds sky130=.work/runs/r1-native/sky130.gds \
  --signoff-gds gf180=.work/runs/r1-native/gf180.gds
python3 scripts/verify_router_signoff.py --run-dir .work/runs/rsig2 --flow .work/runs/r2 \
  --rustroad target/release/rustroad
python3 scripts/verify_gds.py --run .work/runs/gds5 --source .work/runs/r2 \
  --rustroad target/release/rustroad --evidence docs/evidence/gds-verification.json
python3 scripts/compare_spef.py campaign --workspace .work/runs/spef2 --base-rev empty \
  --ready .work/runs/r2 --sta-bin $STA
python3 scripts/compare_spef.py evidence --summary .work/runs/spef2/summary.json \
  --out docs/evidence/extraction-requalification.json
# OLD DEF: export-def of the same workspaces with the b813c34 binary
python3 scripts/def_segmentation.py --run .work/runs/seg2 \
  --case sky130:OLD/sky130.def:.work/runs/r2/sky130.def:ROUTES/sky130_routes.json \
  --case gf180:OLD/gf180.def:.work/runs/r2/gf180.def:ROUTES/gf180_routes.json \
  --evidence docs/evidence/def-segmentation.json
python3 scripts/timing_baseline.py --flow .work/runs/r2 --out .work/runs/tb4 \
  --evidence docs/evidence/timing-baseline.json
python3 scripts/requalification_evidence.py --pass1 .work/runs/r1 --pass2 .work/runs/r2 \
  --native-gds .work/runs/r1-native --router-signoff .work/runs/rsig2 \
  --rustroad target/release/rustroad --scratch .work/runs/evidence-scratch \
  --out docs/evidence/requalification.json

Open items at the end of this stream

  • Hold closure and the Sky130 clock-input antenna item went to the hold and antenna ECO streams (docs/hold-closure.md).
  • verify_routing.py/flow_status.py gated only tt+nom; the three-corner gate is scripts/rtl2gds.py (docs/rtl2gds.md).
  • GF180 density needs fill; Sky130 windowed density needs a chip-level layout. Both are reported separately.
  • Extraction accuracy items (PDK corner R data, omitted via pads and end caps) are described in docs/extraction.md.