drt: allow NDR auto-taper to be disabled per net by simi1505 · Pull Request #10805 · The-OpenROAD-Project/OpenROAD

simi1505 · 2026-07-04T00:36:04Z

Summary

Adds runtime, per-net control over the detailed router's auto-tapering of non-default-rule (NDR) nets. By default, DRT tapers wide NDR wires down to minimum width near pin connections. Needed for digital pins, but unwanted for wide analog/pad traces that must keep full width all the way to the pin.

Previously the only control was the compile-time AUTO_TAPER_NDR_NETS global. This exposes it per net at runtime, no recompile required.

Addresses #9995 and #10032.

Changes

odb: new dbNet flag disable_auto_taper (disableAutoTaper() / setDisableAutoTaper()), with a schema-revision bump.
drt: honours the flag during pin access and maze routing (frNet::isAutoTaperDisabled, drNet::autoTaperEnabled).
Tcl: new command set_routing_disable_auto_taper (-net <name> | -all_clocks) [-disable] [-enable].
Adds regression test ndr_no_auto_taper.

Notes

Default behaviour is unchanged (auto-taper stays on for all nets).
The flag is subtractive: it can disable taper on a net, but never forces taper on when the global is off; -enable restores the default.
Continues the work in drt: allow auto-taper to be disabled per net (#9995) #10721, rebased onto current master with review fixes (schema revision, split include, ord::get_db_block).

Testing

Verified end-to-end via LibreLane on the IHP sg13g2 AMS chip template:

Top-level chip assembly: wide NDR analog/pad traces keep their full width all the way to the pins (previously tapered to minimum width at the pin).
Digital macro (counter): routes cleanly with no regression when the option is unused (the default path is behaviour-preserving).

Config comparison:

Without DRT_DISABLE_AUTO_TAPER: https://github.com/iic-jku/ihp-sg13g2-ams-chip-template/tree/main
With DRT_DISABLE_AUTO_TAPER: https://github.com/iic-jku/ihp-sg13g2-ams-chip-template/tree/fixed-auto-taper

Companion LibreLane PR: librelane/librelane#978

github-actions

Welcome to OpenROAD! Thanks for opening your first PR.
Before we review:

Contribution Guide: https://openroad.readthedocs.io/en/latest/contrib/contributing.html
Build Instructions: https://openroad.readthedocs.io/en/latest/contrib/BuildWithCMake.html

Please ensure:

CI passes
Code is properly formatted
Tests are included where applicable
A maintainer will review shortly!

gemini-code-assist · 2026-07-04T00:51:53Z

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.

Signed-off-by: Simon Dorrer <simon.dorrer@jku.at>

maliberty · 2026-07-04T02:42:53Z

I'm not sure what a "pad trace" refers to but dbSigType can be used to detect analog nets. Does that suffice or is a new bit really needed?

mole99 · 2026-07-04T07:50:24Z

  return false;
 }

+bool UniqueInsts::isNoAutoTaperNDRInst(frInst* inst) const


I'm not sure about this one. Does that mean that, as soon as an instance is connected to a net with auto-taper disabled, that the entire instance has auto taper disabled?

Also, why not use isNDRInst(inst) instead of net->getNondefaultRule() as in the original code?

isNDRInst(inst) only tells us whether the instance touches some NDR net. It reduces the per-terminal information to a single bool, so it can't tell us whether the same net is both NDR and has auto-taper suppressed.

Here I need that conjunction on one net. Take an instance with net A (NDR, taper enabled) and net B (auto-taper disabled, but no NDR): isNDRInst(inst) would be true (because of A) and "the instance has an auto-taper-disabled net" would also be true (because of B), yet no single net qualifies. So the instance should not be treated as no-auto-taper. Auto-taper only affects NDR nets, so the per-net flag is only meaningful when the same net also carries an NDR.

That's why the loop checks net->getNondefaultRule() && (!AUTO_TAPER_NDR_NETS || net->isAutoTaperDisabled()) together per terminal, rather than combining two instance-level predicates.

However, the condition was duplicated. I've factored it into frNet::isAutoTaperSuppressed(bool auto_taper_enabled) and reused it at all three sites. Here in isNoAutoTaperNDRInst, and the two in FlexPA_acc_point. So the predicate now lives in one place. Behaviour is unchanged. See commit e638dd6

To your other question, "Does that mean that, as soon as an instance is connected to a net with auto-taper disabled, that the entire instance has auto taper disabled?":
Yes, for pin-access purposes, if an instance touches at least one NDR net whose auto-taper is suppressed, the whole instance gets its own unique class. That's intentional and matches the pre-existing behaviour: the line this replaced was if (!AUTO_TAPER_NDR_NETS && isNDRInst(inst)), which likewise classified the entire instance. Pin-access patterns are computed per unique-instance-class, so a no-taper NDR instance needs its own class to get access points computed against the untapered (full-width NDR) geometry. The actual tapering still happens per-net/per-pin during routing. This only affects how instances are grouped for pin access.

Alright. So as long as you can have instances with both non-NDR nets and NDR nets connected to them, it with tapering individually enabled, this should be fine.

Address review: the double negative ("disable" in the command name plus -disable/-enable flags) was confusing. Rename to `set_routing_auto_taper` and require exactly one of -enable/-disable (no default), so it reads as a direct on/off switch. Behavior and the odb flag (`disable_auto_taper`) are unchanged.

@mole99

…essed` Address review from @mole99: the 'NDR net whose auto-taper is suppressed' predicate was duplicated inline in three places (FlexPA_unique once, FlexPA_acc_point twice). Extract it into `frNet::isAutoTaperSuppressed(bool auto_taper_enabled)` and reuse it. Behavior-preserving.

simi1505 · 2026-07-04T12:06:40Z

I'm not sure what a "pad trace" refers to but dbSigType can be used to detect analog nets. Does that suffice or is a new bit really needed?

Good question. I asked Claude to help me with this. This is his thinking summarized. What do you say?

Let me clarify "pad trace" first. In top-level chip assembly (padframe / chip integration), the wide NDR connections from I/O pads to the core (or pad-to-pad), must keep their full NDR width all the way to the pad/pin rather than taper. These are typically SIGNAL-type nets with an assigned NDR, not marked ANALOG.

I looked at using dbSigType, but I don't think it covers the cases here:
It's orthogonal to signal type. "Don't taper / keep full NDR width" is a routing directive, not a classification. The nets that need it are usually SIGNAL (pad/assembly/high-current), and conversely not every analog net wants no-taper.
Flows don't set ANALOG for these nets. They're selected by NDR assignment (e.g. LibreLane's DRT_ASSIGN_NDR by name regex) and remain SIGNAL. Keying off ANALOG would require flows to additionally classify nets as analog in odb, which they don't do today (including the top-level case in #10032).
It can't express per-net choice among NDR nets. The regression test keeps 2 of 6 NDR clock nets untapered while the rest still taper. sigType is too coarse for that.
On the cost: this doesn't grow dbNet. It reuses the existing free unused2 bit, so the only overhead is the schema-revision bump. I modelled it after dbNet::dont_touch (a per-net routing directive stored as a flag), which felt like the closest precedent.

simi1505 · 2026-07-04T12:09:59Z

@mole99 @maliberty I tried to answer your review as well as possible. If you are not happy with this solution, we can further discuss it at FSiC in two days. Just let me know. However, we really would appreciate this feature. Especially for the HeiChips 2026 in the first week of August, we would need this feature. Thanks!

Address review: the double negative ("disable" in the command name plus -disable/-enable flags) was confusing. Rename to `set_routing_auto_taper` and require exactly one of -enable/-disable (no default), so it reads as a direct on/off switch. Behavior and the odb flag (`disable_auto_taper`) are unchanged. Signed-off-by: Simon Dorrer <simon.dorrer@jku.at>

@mole99

…essed` Address review from @mole99: the 'NDR net whose auto-taper is suppressed' predicate was duplicated inline in three places (FlexPA_unique once, FlexPA_acc_point twice). Extract it into `frNet::isAutoTaperSuppressed(bool auto_taper_enabled)` and reuse it. Behavior-preserving. Signed-off-by: Simon Dorrer <simon.dorrer@jku.at>

…imi1505/OpenROAD into drt-disable-auto-taper-per-net

maliberty · 2026-07-05T04:40:20Z

  }
 }

+sta::define_cmd_args "set_routing_auto_taper" \


Add documentation to the odb/README

maliberty · 2026-07-05T04:42:30Z

+  if { !([info exists keys(-net)] ^ [info exists flags(-all_clocks)]) } {
+    utl::error ORD 1016 "Either -net or -all_clocks need to be defined."
+  }
+  if { [info exists flags(-enable)] == [info exists flags(-disable)] } {
+    utl::error ORD 1017 "Exactly one of -enable or -disable must be specified."
+  }


Why use two different idioms and different words (Either vs Exactly one) for the same one-hot style check?

maliberty · 2026-07-05T04:43:35Z

+# nets without recompiling.  Use -enable to restore the default behavior.
+proc set_routing_auto_taper { args } {
+  sta::parse_key_args "set_routing_auto_taper" args \
+    keys {-net} flags {-all_clocks -enable -disable}


The -enable flag is unused

maliberty · 2026-07-05T04:45:47Z

+  bool disableAutoTaper();
+
+  void setDisableAutoTaper(bool disable_auto_taper);


Using positive names avoids double negatives like setDisableAutoTaper(false) which just means setAutoTaper() [using bool enable=true arg]

simi1505 requested review from a team as code owners July 4, 2026 00:36

simi1505 requested a review from maliberty July 4, 2026 00:36

simi1505 mentioned this pull request Jul 4, 2026

added DRT_DISABLE_AUTO_TAPER to disable detailed-routing auto-taper per net librelane/librelane#978

Open

github-actions Bot added the size/M label Jul 4, 2026

github-actions Bot reviewed Jul 4, 2026

View reviewed changes

simi1505 mentioned this pull request Jul 4, 2026

Allow auto taper to be disabled per net #9995

Open

added option for "Disable Auto-Taper Per Net"

6f6803d

Signed-off-by: Simon Dorrer <simon.dorrer@jku.at>

simi1505 force-pushed the drt-disable-auto-taper-per-net branch from 9a79975 to 6f6803d Compare July 4, 2026 01:03

simi1505 mentioned this pull request Jul 4, 2026

drt: add -disable_auto_taper_ndr_nets flag to detailed_route hpretl/OpenROAD#1

Draft

mole99 reviewed Jul 4, 2026

View reviewed changes

Comment thread src/OpenRoad.tcl Outdated

mole99 reviewed Jul 4, 2026

View reviewed changes

simi1505 added 2 commits July 4, 2026 10:50

simi1505 added 3 commits July 4, 2026 14:16

Merge branch 'drt-disable-auto-taper-per-net' of https://github.com/s…

c27572e

…imi1505/OpenROAD into drt-disable-auto-taper-per-net

maliberty requested a review from osamahammad21 July 5, 2026 04:38

maliberty reviewed Jul 5, 2026

View reviewed changes

		bool disableAutoTaper();

		void setDisableAutoTaper(bool disable_auto_taper);

Uh oh!

Conversation

simi1505 commented Jul 4, 2026

Summary

Changes

Notes

Testing

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot commented Jul 4, 2026

Uh oh!

maliberty commented Jul 4, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

simi1505 commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

simi1505 commented Jul 4, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

simi1505 commented Jul 4, 2026 •

edited

Loading