Skip to content

Stage-current placement diagnostic

This opt-in research command observes at most three current optimizer stage-end iterates. It completes the ordinary global-best selection first, then applies the existing retarget, legalize and detail helpers serially to those observations. It never chooses a trial as the ordinary result, inserts physical cells, changes the source HEAD, or asserts foundry/timing/equivalence acceptance.

rustroad --workspace /private/import --threads 2 place-p1-diagnose \
  --base-rev rev1 --out-workspace /private/new-observations --out-base-rev empty

Only a fresh disjoint output path is accepted. An existing different HEAD gives REVISION_CONFLICT; an existing matching or partial output is not adopted. The output base accepts the empty sentinel independently of source revision handles. An existing published HEAD must be exactly one positive canonical revN followed by one newline before a conflict is reported; malformed HEADs remain invalid workspaces. The bounded read and both identity checks precede any output creation, and the shared legacy source-base logic is unchanged. Source SH and output EX locks span the calculation and commit. Source data must be regular, unique-inode, read-only files below non-symlink ancestors. The selected inline imported schema3/4 revision has exactly design.json and the four canonical instance/connectivity side tables. Unknown files, routed/materialized or seeded state, or unsupported metadata reject explicitly. Nonempty generated tap/filler arrays reject in a lexical seed before their values are restored.

Observed phases and authority

The ordinary legacy baseline is observed once, followed by the unchanged ordinary selected geometry. Up to three reached stages retain their current iterate, even if it differs from the optimizer's global best. Identical f64 payloads share storage but retain separate stage labels and stage reports. Each reached stage then emits integer projection, after-retarget, after-legalize and after-detail coordinates. A zero-horizon stage is not invented. A successful late stall retains its actual reached current iterate; numerical/budget/observer failures are unverified errors, never successful baseline fallback.

observations.json uses rustroad.p1-stage-observations.v1 in its header. It contains borrowed imported metadata, initial fixed-order packing, the exact movable-to-instance mapping, native centre-net occurrence order and free intervals, plus the already computed named/layered effective I/O rectangles, followed by streaming global integer movable positions. Imported master pin shapes, fixed placement and N/FS rows remain available for an independent adapter. This raw document is not the separate tiny semantic schema. Legal/detail orientations follow the actual row. Projected/retargeted targets have no assigned physical orientation; any terminal-pin metric at those targets needs a separately declared orientation scope. Only native centre cost is reported here. Pin HPWL, displacement, DRC, LVS and timing remain unchecked.

Every set has a unique position_set_id; header and summary explicitly identify ordinary_selected as the ordinary reference. The writer checks required baseline/default and reached-stage phase order before commit. Effective I/O is retained only in diagnostic preparation, with<=4096 bits and<=1MiB aggregate name/layer bytes admitted before the existing helper clones names. The ordinary None path retains no additional I/O view. Both admission-record passes and the retained-record visit are charged; no I/O placement is recomputed for output.

Every observation includes actual cumulative defined work and numerical counters. Reference geometry/status/selection equal the ordinary path when the diagnostic completes; its serialized work does not equal an ordinary call: diagnostic preparation and capture work already incurred are included. The reference counters are frozen before the later trial/metric passes. Final diagnostic counters include those later passes. No trial resets the meter.

The independently authored tiny oracle was frozen before this implementation: 25 cases, wide integer geometry/native comparability, 40 rejection controls. Its data and API may be consumed; its validator implementation was deliberately not read before source freeze. Existing self-authored numerical tests and a successful compilation alone cannot qualify this change. Tiny independent geometry, dual-PDK controls and the old qualified default geometry remain gates.

Numerical admission and simultaneous buffers

Let O=N+F, P=all pin occurrences, p=modelled pin occurrences, B=4096, T=2B(nx+ny), J=max_iterations and E=max_evaluations. The existing requested-work bound is W=E(p+D+T+10B)+20OJ+P(J+2), where D counts both deposition directions. The diagnostic explicitly admits W+E+H, with H=6O+6. The additional E covers the evaluation-control charge omitted by the old conservative formula. H holds up to three O-sized copies, 0+O+2O comparisons, and three entered/finished pairs. Held units exclude competing work but are not completed work. Consumption is atomic; unused hold is released on both successful and failed optimizer exit. Retarget/legal/detail and all exact metric passes share the same remaining 20-billion total and existing numerical counters, without reserve reset.

The old engine model is 256O+64P+96M+112I+1MiB. The diagnostic adds at most48O for three snapshots,48N for one serial trial's target/local/global trio, and 128KiB of writer/control scratch. The old model already includes the ordinary baseline, translated view, optimizer histories/global-best outcome, ordinary targets/global candidate and selected coordinates that can coexist with trials. During each trial projected/retargeted locals are dropped before subsequent locals, and float targets are dropped before detail. No DesignDb/PDK/library is deep-cloned. Model/metadata/config borrows end before physical materialization; the new command does not call that materialization at all.

At N=100000,F=4096,P=1000000,M=100000,I=250000, the admitted old+new engine estimate is139224832 bytes. The selected workspace front door is narrower: 20000 cells,1024 rows,20000 nets,100000 pins,256 masters,131072 aggregate pin shapes,200000 netlist records; every array also obeys one million aggregate elements and two million nodes. Rows*(fixed+1)<=250000 is admitted before initial packing; tap sites and blockage split bounds are separately admitted before their diagnostic output vectors grow. Source JSON<=16MiB and all five revision extents<=32MiB, with raw/decoded string, depth and duplicate-key limits before typed restoration. Unsupported arithmetic rejects before legacy native costs: primitives<=2^40, i64 midpoint/span safety and u64 cost/weight upper bounds.

The whole-process expected model is1850500864 bytes: parsed IR768MiB, executable image128MiB, parser/physical preparation128MiB, numerical139224832 bytes, source clean file cache256MiB, output cache64MiB, four8MiB stacks, allocator128MiB and controller128MiB. This is a conservative engineering estimate, not a proof about allocator or kernel memory. A separately authorized2GiB physical+512MiB swap/no-AS foreground guard, measured RSS and cleanup are required. OOM or timeout is unverified. No combinatorial paths or graph histories are materialized. The new API requires an explicitly installed pool of at most four workers; checking it does not initialize an unconstrained global Rayon pool. The CLI rejects diagnostic thread counts above four before building its pool.

Explicit byte passes and commit

One2GiB ledger covers named read/hash/scan/encode/write byte passes. The old loader's admitted read/mapped extents are charged once, not misrepresented as syscall tracing. Lexical nodes, initial packing expansions and the bounded physical preparation's defined record/tap visits are carried into the numerical total. Existing loader/fixed validation retain their own finite work guards; the reported number is not every CPU instruction.

Each observation file is<=60MiB, each summary/failure/CLI JSON<=64KiB, and the entire output inventory remains below64MiB and16files. The193MiB held tail is released once the variable writer closes, then pays actual observation read/SHA<=120MiB, one source read/SHA<=64MiB+128bytes, summary encode/write/read/SHA <=256KiB, ten HEAD encode/write units, and admits the fixed CLI<=128KiB tail. The receipt labels its charged total before CLI stdout; it does not call held or unexecuted work completed. Errors report attempted/charged scope.

All variable-sized validation, source/namespace/identity checks and counter admission precede HEAD rename. Precommit failures retain private partial files without HEAD. Rename/fsync/kill/stdout uncertainty may leave a provisional HEAD; only actual exit0, the typed completion receipt and current source/output identities establish diagnostic completion. HEAD is never a physical design revision or acceptance gate. There is no rollback loop or failed-evidence deletion.

Source and execution gates

The source-only freeze, formatter logs and mechanical default-seam review are separate from later compiler/test qualification. A fresh isolated target and owned guards are required for every future build stage. No EDA/native/reference execution or promotion is authorized by the presence of this source document. The old qualified P1 checkouts, binaries and all failed research receipts remain unchanged. In particular, inherited OpenROAD exit42 remains unverified.