Skip to content

asap7: DATABASE MICRONS 4000, so the 1x tech LEFs' quarter-nanometre rules hold - #4663

Closed
jeffhsu3 wants to merge 2 commits into
The-OpenROAD-Project:masterfrom
jeffhsu3:asap7-tech-lef-rounding
Closed

jeffhsu3 wants to merge 2 commits into
The-OpenROAD-Project:masterfrom
jeffhsu3:asap7-tech-lef-rounding

Conversation

@jeffhsu3

@jeffhsu3 jeffhsu3 commented Oct 10, 2026 •

Copy link
Copy Markdown

Summary

Both 1x tech LEFs (flow/platforms/asap7/lef/asap7_tech_1x_201209.lef and asap7_tech_1x_260907.lef) declare DATABASE MICRONS 1000 but write several rule values on a quarter nanometre. OpenROAD stores LEF distances as lround(value * dbu), so those values are rounded to whole nanometres, and for three of them the rounded value stops the rule from doing what it was written to do. This PR changes the unit to DATABASE MICRONS 4000 in both files, at which every value is exact, and leaves the rule values as ASAP7 wrote them.

Revised from an earlier version that substituted whole-nanometre values instead. The review of the same change in siliconcompiler/lambdapdk#246 asked for the unit instead, and The-OpenROAD-Project/OpenROAD#11695 (merged) already made this change to OpenROAD's own copies of the 201209 LEF under test/asap7.

Layer: rule LEF value Stored at 1000 Stored at 4000 Router comparison Effect at 1000
M1 LEF58_EOLKEEPOUT EOL width 0.01825 18 (= min width) 73 (min width 72) length < eolWidth An 18 nm (minimum-width) tip is never a line end, so the keep-out never applies.
M4/M5 LEF58_EOLKEEPOUT side extension 0.02425 24 (= adjacent edge) 97 (edge at 96) corner strictly inside the box Ends on the adjacent track's edge, so adjacent-track tips are never caught.
M6/M7 LEF58_EOLKEEPOUT side extension 0.03225 32 (= adjacent edge) 129 (edge at 128) same Same, against the 32 nm adjacent track.
M8/M9 SPACINGTABLE thresholds 0.05975 … 0.99975, 0.39975 / 1.19975 / 1.79975 rounded up exact row/column applies when greater than the value A width or run exactly on the intended 60 … 1000 nm or 400 / 1200 / 1800 nm threshold falls in the row below.

Also now exact: M6/M7 AREA 0.0021875 (35000 units², was rounded to 2188 nm²) and the remaining quarter-nanometre values (M2/M3 and M3 0.0125, M6/M7 ENDOFLINE 0.0375, V3/V5 enclosure 0.02425, the M4_M3widePWR0p864 width ranges), which rounded harmlessly before.

MANUFACTURINGGRID 0.001 is unchanged (now 4 units), so routed geometry stays on the nanometre; the router checks shapes, not rule values, against that grid. The platform's cell, DFF and SRAM LEFs carry no UNITS block and load unchanged (asap7sc7p5t site 216 x 1080 units, i.e. 0.054 x 0.27 µm).

Evidence

  • Stored values. Reading either LEF into OpenROAD 26Q3 with the platform's cell LEFs and querying the EOLKEEPOUT rules gives M1 EOL width 18, M4/M5 side extension 24 and M6/M7 32 at 1000; with this patch 73, 97 and 129. A scan of every number in both files finds all exact at 4000.
  • check_drc on a routed SHA-256 (stock 201209 LEF, SiliconCompiler 0.38.10 / OpenROAD 26Q3). drt::check_drc reports no M1 or M4/M5 keep-out violations on that database. Raising the stored M1 EOL width above 18 makes it report 1,164 M1 keep-out violations; raising the M4/M5 side extension above 24 makes it report 249 on M4 and 2 on M5.
  • Independent checker. These include every M1 tip-spacing and M4/M5 adjacent-track tip/parallel-run violation found by an independent DRM-based checker (gdscheck's ASAP7 deck).
  • Full flow with the rules active (measured on the earlier whole-nanometre revision, which enables the same rules): detailed routing converges to 0 violations; wirelength 53,790 µm (was 53,798); vias +0.17%; detailed-route time +22% (156 s → 191 s); timing unchanged.
  • QoR. The earlier revision's QoR check flagged two asap7 designs, since detailed routing starts with more violations once the keep-outs apply. This revision enables the same rules, so the same effect is expected; the check will rerun on this push.

The investigation is written up in ASAP7_ECOSYSTEM_DISCREPANCIES.md, sections 5–7.

Notes

🤖 Generated with Claude Code

The ASAP7 1x tech LEFs (asap7_tech_1x_201209.lef and asap7_tech_1x_260907.lef)
declare DATABASE MICRONS 1000 but write several rule values on a quarter
nanometre. OpenROAD stores them as lround(value * 1000), and three of them
then stop the rule doing what it was written to do:

- M1 EOLKEEPOUT 0.01825 becomes 18 nm. The EOL test is length < width, so no
  18 nm (minimum-width) tip is a line end and the keep-out never applies.
  Now 0.019.
- M4/M5 EOLKEEPOUT side extension 0.02425 becomes 24 nm and M6/M7's 0.03225
  becomes 32 nm, exactly the edge of the wire on the adjacent track. The
  CORNERONLY check needs the corner strictly inside, so adjacent-track tips
  are never caught. Now 0.025 and 0.033.
- M8/M9 SPACINGTABLE thresholds (0.05975 ... 0.99975, 0.39975, 1.19975,
  1.79975) round up, so a width or run exactly on the intended 60 ... 1000 nm
  or 400/1200/1800 nm threshold falls in the row below. Now rounded down
  (0.059 ... 0.999, 0.399, 1.199, 1.799).

On a routed SHA-256, OpenROAD's check_drc reported no M1 or M4/M5 keep-out
violations with the stored values; with the M1 width at 19 and the M4/M5
side extension at 25 it reported 1164 on M1 and 249 + 2 on M4/M5.

The other quarter-nanometre values (M2/M3 0.0125, M6/M7 ENDOFLINE 0.0375,
V3/V5 EOL enclosure 0.02425, M6/M7 AREA, M4_M3widePWR0p864 width ranges)
round to values that behave the same on a 1 nm grid and are unchanged. The
4x LEF holds these values as whole nanometres and is unaffected.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: jeffhsu3 <jeffhsu3@gmail.com>

@gemini-code-assist gemini-code-assist Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request updates the LEF technology files (asap7_tech_1x_201209.lef and asap7_tech_1x_260907.lef) to adjust spacing and EOL keepout rules. The modifications address rounding issues introduced by the DATABASE MICRONS 1000 setting, ensuring that values do not round to exact grid boundaries or incorrect thresholds. Specifically, keepout and extension values for layers M1, M4, M5, M6, and M7 are slightly increased, and parallel run lengths and widths in the spacing tables for M8 and M9 are adjusted. No review comments were provided, and there is no additional feedback.

jeffhsu3 added a commit to jeffhsu3/asap7_reference_design that referenced this pull request Oct 10, 2026
§6 now traces where LambdaPDK's ASAP7 tech LEF comes from.
asap7sc7p5t_28's techlef_misc copy is byte-identical, and ORFS's platform
LEF matches it apart from implant layers. It also lists the same fix
proposed at every copy, covering both the 201209 LEF and the newer 260907
LEF:

- siliconcompiler/lambdapdk#246
- The-OpenROAD-Project/OpenROAD-flow-scripts#4663
- The-OpenROAD-Project/asap7sc7p5t_28#10

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@openroad-ci

openroad-ci commented Oct 10, 2026 •

Copy link
Copy Markdown
Member

🔍 QoR check

Metrics reflect the PR merge build — i.e. what will land on the target branch.

Commit 3be3e58 · Jenkins build #3 · Baseline: build · View build on dashboard

62 design(s) checked — 3 with regression(s), 1 without a comparable baseline.

❌ asap7/aes-mbff base — 1 failing metric(s)
Metric target base delta limit band
globalroute__timing__setup__tns -341.223 -263.679 -29.408% -333.679 PeriodPadding 20.000%
❌ asap7/riscv32i base — 3 failing metric(s)
Metric target base delta limit band
cts__timing__setup__tns -1151.680 -61.627 -1768.806% -251.627 PeriodPadding 20.000%
finish__timing__setup__tns -18448.500 -960.116 -1821.487% -1152.139 PeriodPadding 20.000%
globalroute__timing__setup__tns -376.405 -23.904 -1474.633% -213.904 PeriodPadding 20.000%
❌ asap7/uart base — 2 failing metric(s)
Metric target base delta limit band
finish__timing__setup__tns -1239.000 -920.871 -34.547% -1105.045 PeriodPadding 20.000%
globalroute__timing__setup__tns -1892.230 -1502.450 -25.943% -1802.940 PeriodPadding 20.000%

Both 1x tech LEFs declare DATABASE MICRONS 1000 but write several rule
values on a quarter nanometre, which a 1 nm database rounds off.  OpenROAD
stores lround(value * dbu), and then no 18 nm M1 tip is a line end for
EOLKEEPOUT (0.01825 -> 18, the test being length < width), the M4-M7 corner
keep-out's side extension (0.02425, 0.03225 -> 24, 32) ends exactly on the
adjacent track's wire edge, which a corner must lie strictly inside, and the
M8/M9 SPACINGTABLE thresholds (0.05975 ... 1.79975) round up past the
manual's.

At 4000 units per micron every value in both files is exact, so the rules
hold as ASAP7 wrote them: the stored M1 line-end width is 73 against a
minimum width of 72, the side extensions 97 and 129 against adjacent edges
at 96 and 128, and M6/M7's AREA 0.0021875 is 35000 rather than 2188.  The
manufacturing grid stays 0.001, now 4 units, so geometry is still on the
nanometre.  The platform's cell, DFF and SRAM LEFs carry no UNITS block and
load unchanged.

This replaces the earlier whole-nanometre substitutes, as the review of
siliconcompiler/lambdapdk#246 asked; that copy and
The-OpenROAD-Project/asap7sc7p5t_28#10 now make the same change, and
The-OpenROAD-Project/OpenROAD#11695 made it to OpenROAD's own copies of
this LEF and warns when a LEF value rounds (ODB-1220).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Signed-off-by: jeffhsu3 <jeffhsu3@gmail.com>
@jeffhsu3 jeffhsu3 changed the title asap7: whole-nanometre values where the tech LEF rounds off a rule asap7: DATABASE MICRONS 4000, so the 1x tech LEFs' quarter-nanometre rules hold Oct 13, 2026
@jeffhsu3 jeffhsu3 closed this Oct 25, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants