rsz: Merge rsz-buffering-boost for Rebuffer QoR gain by minjukim55 · Pull Request #10866 · The-OpenROAD-Project/OpenROAD

minjukim55 · 2026-07-10T00:52:04Z

Summary

Port of the upstream rsz rebuffering-accuracy series (original author
Martin Povišer). Core idea: rebuffer builds its buffer trees from an internal
delay/slew estimate that had drifted away from the model repair_design (and
signoff STA) actually use — so rebuffer would optimize for numbers that later
stages disagree with. The series realigns that estimate and fixes several
errors in the buffered-net cost model.

The 8 substantive commits:

Move the shape factor to Resizer — relocate the slew "shape factor"
(the NLDM-derived slew/load correction) into the shared Resizer so
rebuffer and repair_design read one consistent value instead of keeping
separate copies that can diverge.
Use the new slew degradation model in rebuffer — repair_design
recently got a new slew degradation model; adopt the identical model inside
rebuffer so its internal estimate matches the signed-off timing model.
Fix the wire-splitting formula — stop scaling r_drvr by the shape
factor. r_drvr is the asymptotic slew/load ratio derived from NLDM data
and already contains the shape factor, so the extra scaling double-counted
it.
Adjust the buffering-test methodology — when measuring the delay on a
buffered net, include the driving gate delay so the measurement uses the
same delay definition as the estimator.
Improve junction placement in the buffer topology rewrite — place the
Steiner junctions of the rewritten buffer tree more accurately, tightening
the estimated wirelengths and loads the cost model sees.
Fix buffer-load miscount — pass loads to the gate-delay calculator
exclusive of the driving port capacitance (the driver's own pin cap was
being counted into the load it drives).
Redo the long-wire algorithm — rework the long-wire buffering /
length-estimation path so the iterative length estimate stays accurate on
long, high-fanout nets.
Add zero-RC guards — guard the RC math against zero wire RC (0 * inf
→ NaN); fixes the STA-1350 "gate delay load cap is NaN" crash in
fully_rebuffer on nets with no parasitics.

(The branch also carries small housekeeping commits: two clang-format fixes and
one test-golden update for the numeric churn below.)

QoR

5 random placement seeds × 2 binaries (base = pre-series, test = this branch)
on a customer design (advanced-node PDK), full flow place → CTS → route.
Everything downstream of the binary is identical (same floorplan, same seeds),
so the delta isolates the change.

WS = global setup worst slack (ns). % = test vs base (+ = better).
fmax % is derived from the global WNS so it always tracks WS — the
per-clock fmax metric picks a different worst path and can move opposite to
WS.

PLACE

seed	WS base	WS test	WS %	fmax %
1	-0.0959	-0.0638	+33.4%	+7.39%
2	-0.0906	-0.0876	+3.3%	+0.65%
3	-0.1222	-0.0707	+42.2%	+11.69%
4	-0.0825	-0.1017	-23.2%	-4.06%
5	-0.0871	-0.1140	-30.8%	-5.55%
mean	-0.0957	-0.0876	+8.5%	+1.85%

CTS

seed	WS base	WS test	WS %	fmax %
1	-0.0892	-0.0569	+36.2%	+7.56%
2	-0.0785	-0.0691	+12.0%	+2.14%
3	-0.0885	-0.0607	+31.4%	+6.45%
4	-0.1004	-0.0553	+44.9%	+10.58%
5	-0.1132	-0.0366	+67.7%	+18.83%
mean	-0.0939	-0.0557	+40.7%	+8.98%

GRT (global route)

seed	WS base	WS test	WS %	fmax %
1	-0.1433	-0.0971	+32.3%	+9.89%
2	-0.0948	-0.0724	+23.6%	+5.06%
3	-0.0907	-0.0889	+2.0%	+0.40%
4	-0.1571	-0.0890	+43.3%	+14.84%
5	-0.1201	-0.0906	+24.6%	+6.42%
mean	-0.1212	-0.0876	+27.7%	+7.08%

Takeaway: consistent setup-WS / fmax improvement from CTS onward
(CTS +8.98%, GRT +7.08% fmax mean; 5/5 seeds improve at both stages). PLACE is
seed-noisy and nets to a small gain — the estimate matches the signed-off timing
model once CTS/routing make that model accurate.

Related Issues

Upstream rsz rebuffering-accuracy series. QoR validation tracked internally
(private).

Signed-off-by: Martin Povišer <povik@cutebit.org>

Recently repair_design received a new slew degradation model, use the same model in rebuffer. Signed-off-by: Martin Povišer <povik@cutebit.org>

r_drvr should not be scaled by the shape factor as it is derived from NLDM data as the asymptotic slew/load ratio and as such it already includes the shape factor. Signed-off-by: Martin Povišer <povik@cutebit.org>

When measuring the delay on a buffered net include the driving gate delay. Signed-off-by: Martin Povišer <povik@cutebit.org>

Signed-off-by: Martin Povišer <povik@cutebit.org>

Fix load values passed to the gate delay calculator to be exclusive of the driving port capacitance. Signed-off-by: Martin Povišer <povik@cutebit.org>

Signed-off-by: Martin Povišer <povik@cutebit.org>

Signed-off-by: Minju Kim <mkim@precisioninno.com>

Cherry-picked from 86b8145b48 to fix STA-1350 gate delay load cap is NaN crash in fully_rebuffer when a net has zero wire RC (0 * inf). Signed-off-by: Martin Povišer <povik@cutebit.org> Signed-off-by: Minju Kim <mkim@precisioninno.com>

Regenerate unit .ok logs (rsz/gpl/openroad), their secondary golden outputs (.defok/.vok DEF/verilog), and aes/jpeg/ibex sky130 flow .metrics + .metrics_limits for the numeric churn from the estimation-accuracy rebuffering changes. The flow pass/fail gate is .metrics_limits (check_test_metrics compares the result against it; .metrics is the reference used to recompute the limits). Both are regenerated from a local run of the branch binary. Secondary DEF/verilog goldens are regenerated alongside the .ok logs so the in-test diff_file/diff_files comparisons pass and no "Differences found" text leaks into the committed .ok (No-Diffs-In-Ok-Files check). Note: ibex_sky130hs DRT::ANT::errors limit moves 0 -> 1 (one antenna violation from the changed routing). All other diffs are expected metric shifts (buffer counts, slew/cap slack, clock skew). No crashes; verified locally: rsz 241/241 pass, gpl tcl+py pass, openroad.upf_aes pass. Signed-off-by: Minju Kim <mkim@precisioninno.com>

gemini-code-assist · 2026-07-10T00:52:30Z

Note

The number of changes in this pull request is too large for Gemini Code Assist to generate a review.

…g-boost Signed-off-by: Minju Kim <mkim@precisioninno.com> # Conflicts: # src/rsz/src/RepairDesign.cc

repair_design6 is a master-added test (reroute feature). After merging the rsz rebuffering estimation-accuracy series, the improved slew/RC estimation inserts 215 buffers (was 285) with slightly better slack (+115.5% vs +112.4%); no new violations. Regenerate the golden. Signed-off-by: Minju Kim <mkim@precisioninno.com>

minjukim55 · 2026-07-13T06:43:36Z

ORFS: The-OpenROAD-Project/OpenROAD-flow-scripts#4347

Pin tools/OpenROAD at the rsz rebuffering estimation-accuracy series head (The-OpenROAD-Project/OpenROAD#10866, upstream author Martin Poviser). This drops the earlier SparseLU-guard bump so the submodule pin resolves against public OpenROAD for ORFS CI. Signed-off-by: Minju Kim <mkim@precisioninno.com>

gudeh

Changes to GPL are only golden test files.

jhkim-pii · 2026-07-15T05:32:21Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 17d1e1d0f2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-15T05:39:32Z

  const float load_slew
-      = (size.driver_resistance
-         + resizer_->dbuToMeters(bnet->maxLoadWireLength() + extra_wire_length)
-               * wire_res)
-        * (bnet->cap() + extra_cap) * elmore_skew_factor_;
+      = size.driver_resistance * (bnet->cap() + segment_cap)
+        + segment_cap * segment_res * resizer_->slew_shape_factor_ / 2;


Include downstream cap in slew check

When extra_wire_length is nonzero, this check is deciding whether a buffer can drive the added wire segment, but it only charges the wire's own capacitance and omits the Elmore term for the downstream load (segment_res * bnet->cap()). addWire() and annotateTiming() propagate slew limits with wire_res * (wire_cap / 2 + p->cap()), so on long/high-cap fanout branches this underestimates the slew and can keep/export buffer options that still violate the sink max-slew budget.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-15T05:39:32Z

  int rejected_move_count_ = 0;

+  // For Elmore slew modeling, see computeSlewShapeFactor()
+  float slew_shape_factor_ = 0.0;


Initialize slew factor before direct rebuffering

For standalone rsz::fully_rebuffer callers that do not run resizePreamble() first, this cached factor remains at 0 because Resizer::fullyRebuffer() only delegates to Rebuffer::fullyRebuffer() and Rebuffer::init() does not call computeSlewShapeFactor(). Existing direct flows such as src/rsz/test/rebuffer1.tcl call fully_rebuffer after parasitic estimation, so all new slew_shape_factor_ uses suppress wire/via slew degradation and can silently choose incorrect buffer trees; the previous lazy RepairDesign::getSlewRCFactor() path initialized the factor on demand.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-07-15T05:39:32Z

+      float a = wire_res * wire_cap * resizer_->slew_shape_factor_ / 2;
+      float b = (r_drvr * wire_cap)
+                + (wire_res * ref_cap * resizer_->slew_shape_factor_);
+      float c = (r_drvr * ref_cap) - max_load_slew_margined;


Solve the same slew equation used for the violation

When repairing a max-slew wire violation, this quadratic should solve the same load_slew expression used to enter and update the loop a few lines above, where the entire (r_drvr * ... + wire ...) sum is multiplied by resizer_->slew_shape_factor_. Leaving the driver terms unscaled here makes the computed split length correspond to a different equation whenever the shape factor is not 1, so repeaters can be placed too far from the load and the segment can remain over the max-slew budget or require extra repair iterations.

Useful? React with 👍 / 👎.

Signed-off-by: Minju Kim <mkim@precisioninno.com> # Conflicts: # test/aes_sky130hd.metrics # test/aes_sky130hd.metrics_limits # test/ibex_sky130hs.metrics # test/ibex_sky130hs.metrics_limits # test/jpeg_sky130hs.metrics # test/jpeg_sky130hs.metrics_limits # test/upf_aes.defok # test/upf_aes.ok

jhkim-pii · 2026-07-15T06:53:18Z

  double load_slew
      = (r_drvr * (c_wire + ref_cap) + r_wire * ref_cap + r_wire * c_wire / 2)
-        * (*slew_rc_factor_);
+        * resizer_->slew_shape_factor_;


r_drvr considers the slew shape factor already. Multiplication by slew_shape_factor_ (double multiplication of slew_shape_factor_) looks incorrect.

@jhkim-pii r_drvr ultimately comes from LibertyPort::driveResistance(). What makes you say it considers the shape factor already?

Oops, I forgot how driveResistance is defined. So I agree, in effect it has the shape factor baked in and should not be multiplied by it.

jhkim-pii · 2026-07-15T08:35:09Z

      load_slew = (r_drvr * (c_wire + ref_cap) + r_wire * ref_cap
                   + r_wire * c_wire / 2)
-                  * (*slew_rc_factor_);
+                  * resizer_->slew_shape_factor_;


The same issue

jhkim-pii · 2026-07-15T08:35:42Z

  float r_drvr = resizer_->driveResistance(drvr_pin_);
-  assert(slew_rc_factor_.has_value());
-  float load_slew = r_drvr * load_cap * (*slew_rc_factor_);
+  float load_slew = r_drvr * load_cap * resizer_->slew_shape_factor_;


The same issue

jhkim-pii · 2026-07-15T08:36:26Z

+    double slew_left = r_drvr * cap_left * resizer_->slew_shape_factor_;
    double slew_slack_left = maxSlewMargined(max_load_slew_left) - slew_left;
-    double slew_right = r_drvr * cap_right * (*slew_rc_factor_);
+    double slew_right = r_drvr * cap_right * resizer_->slew_shape_factor_;


The same issue

jhkim-pii · 2026-07-16T07:10:24Z

-      const BnetPtr in2 = addWire(aux2, node->location(), -1);
+      odb::Point p1 = aux1->location();
+      odb::Point p2 = aux2->location();
+      odb::Point p3 = crit2->location();


I think the third point should be node->location().
crit2 is not related to the new junction point of aux1, aux2, and node.

povik · 2026-07-16T14:46:31Z

 Pin                                     Limit     Slew    Slack
 ---------------------------------------------------------------
-_548_/B                               320.000  317.570    2.430 (MET)
+_326_/B                               320.000  405.257  -85.257 (VIOLATED)


@minjukim55 I see this test where the violation has shown up has repair_design -reroute. Perhaps that matters

Integrate the rsz policy/optimizer refactor, reroute feature, and other master changes. The rebuffering estimation-accuracy series source is resolved against the refactored rsz; the affected test goldens are regenerated in the following commit. Signed-off-by: Minju Kim <mkim@precisioninno.com>

Regenerate unit .ok logs (rsz/gpl/openroad), their secondary golden outputs (.defok/.vok DEF/verilog), and aes/jpeg/ibex sky130 flow .metrics + .metrics_limits for the numeric churn from the estimation-accuracy rebuffering changes. The flow pass/fail gate is .metrics_limits (check_test_metrics compares the result against it; .metrics is the reference used to recompute the limits). Both are regenerated from a local run of the branch binary. Secondary DEF/verilog goldens are regenerated alongside the .ok logs so the in-test diff_file/diff_files comparisons pass and no "Differences found" text leaks into the committed .ok (No-Diffs-In-Ok-Files check). Also regenerate the repair_design6 golden (a master-added reroute test): the improved slew/RC estimation inserts fewer buffers with slightly better slack and no new violations. Note: ibex_sky130hs DRT::ANT::errors limit moves 0 -> 1 (one antenna violation from the changed routing). All other diffs are expected metric shifts (buffer counts, slew/cap slack, clock skew). Signed-off-by: Minju Kim <mkim@precisioninno.com>

Advance the branch to current OR master so PR The-OpenROAD-Project#3585 stays mergeable (master moved 20d2d5c -> 06c1421: rsz max-iter/drive-resistance/cell-order, dpl negotiation, rcx). Golden conflicts (aes/jpeg sky130hd metrics) are resolved and regenerated against the merged binary in the following commit. Signed-off-by: Minju Kim <mkim@precisioninno.com>

After advancing to OR master 06c1421 (merge M2) and with the slew_shape_factor fix reverted, regenerate the flow goldens against the merged binary: - aes_sky130hd / ibex_sky130hs / jpeg_sky130hd: recompute .metrics_limits (DRV slew/cap slack + clock_skew shift from the reverted slew estimate). - tinyRocket_nangate45: reverting the fix restores the slew_shape_factor double-count, which over-estimates slew -> extra buffering -> global-route congestion overflow (GRT-0228). Enlarge the pipe-cleaner floorplan (die 500->540, core 490->530) so it routes; regen .metrics/.metrics_limits. Signed-off-by: Minju Kim <mkim@precisioninno.com>

povik and others added 11 commits July 7, 2026 15:12

rsz: Move the shape factor to Resizer

edeaf0a

Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Use new slew degradation model in rebuffer

88e0eb2

Recently repair_design received a new slew degradation model, use the same model in rebuffer. Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Fix details of wire splitting formula

80b61fb

r_drvr should not be scaled by the shape factor as it is derived from NLDM data as the asymptotic slew/load ratio and as such it already includes the shape factor. Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Adjust methodology in buffering test

9240c5b

When measuring the delay on a buffered net include the driving gate delay. Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Improve junction placement in buffer topo rewrite

c1136b2

Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Fix miscounting buffer load

1de9203

Fix load values passed to the gate delay calculator to be exclusive of the driving port capacitance. Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: Redo long wire algorithm

0388c94

Signed-off-by: Martin Povišer <povik@cutebit.org>

rsz: clang-format getSlewRCFactor inline body

c38f7e6

Signed-off-by: Minju Kim <mkim@precisioninno.com>

rsz: clang-format Asymptotics struct brace

e31da6c

Signed-off-by: Minju Kim <mkim@precisioninno.com>

rsz: Add zero RC guards

63c8fcc

Cherry-picked from 86b8145b48 to fix STA-1350 gate delay load cap is NaN crash in fully_rebuffer when a net has zero wire RC (0 * inf). Signed-off-by: Martin Povišer <povik@cutebit.org> Signed-off-by: Minju Kim <mkim@precisioninno.com>

minjukim55 self-assigned this Jul 10, 2026

github-actions Bot added the size/XL label Jul 10, 2026

minjukim55 mentioned this pull request Jul 12, 2026

Bump OpenROAD: rsz rebuffering estimation-accuracy series The-OpenROAD-Project/OpenROAD-flow-scripts#4347

Open

1 task

minjukim55 added 2 commits July 12, 2026 23:27

Merge remote-tracking branch 'origin/master' into secure-rsz-bufferin…

878184f

…g-boost Signed-off-by: Minju Kim <mkim@precisioninno.com> # Conflicts: # src/rsz/src/RepairDesign.cc

minjukim55 mentioned this pull request Jul 13, 2026

PrimaDelayCalc::simulate1 SIGSEGV: SparseLU solve() on failed compute() (missing info() guard) #10893

Open

minjukim55 marked this pull request as ready for review July 13, 2026 06:43

minjukim55 requested review from a team as code owners July 13, 2026 06:43

minjukim55 requested review from gudeh and precisionmoon July 13, 2026 06:43

gudeh approved these changes Jul 13, 2026

View reviewed changes

chatgpt-codex-connector Bot reviewed Jul 15, 2026

View reviewed changes

jhkim-pii reviewed Jul 15, 2026

View reviewed changes

jhkim-pii reviewed Jul 16, 2026

View reviewed changes

github-actions Bot added size/L and removed size/XL labels Jul 16, 2026

povik reviewed Jul 16, 2026

View reviewed changes

minjukim55 force-pushed the secure-rsz-buffering-boost branch from c340f92 to b383f98 Compare July 16, 2026 15:07

github-actions Bot added size/XL and removed size/L labels Jul 17, 2026

minjukim55 force-pushed the secure-rsz-buffering-boost branch from 09ec48a to 20a3489 Compare July 20, 2026 12:12

github-actions Bot added size/L and removed size/XL labels Jul 20, 2026

minjukim55 force-pushed the secure-rsz-buffering-boost branch from 20a3489 to 0cb1ec3 Compare July 20, 2026 12:41

github-actions Bot added size/XL and removed size/L labels Jul 20, 2026

minjukim55 added 2 commits July 20, 2026 23:00

Uh oh!

Conversation

minjukim55 commented Jul 10, 2026

Summary

QoR

PLACE

CTS

GRT (global route)

Related Issues

Uh oh!

gemini-code-assist Bot commented Jul 10, 2026

Uh oh!

minjukim55 commented Jul 13, 2026

Uh oh!

gudeh left a comment

Choose a reason for hiding this comment

Uh oh!

jhkim-pii commented Jul 15, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jul 15, 2026

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!

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!

4 participants