[GPL] Add randomized initial placement perturbation in GPL by calewis · Pull Request #10642 · The-OpenROAD-Project/OpenROAD

calewis · 2026-06-11T16:37:05Z

Summary

Implemented a randomized initial placement perturbation feature in the OpenROAD Global Placer (GPL). This feature shifts standard cells by discrete increments of site widths (X direction) and row heights (Y direction), bounded to a maximum of 0.5 microns in both directions.

This replacement of the legacy srand(42) and rand() non-grid-aligned cell perturbation is controlled by a seed configured via the new -random_seed flag (default is 1)

Impact

Replaces the srand(42) based random cell coordinate perturbation loops in global placement with a new implementation.
Adds the -random_seed option to the global_placement Tcl command to customize the random seed (default is 1).
Perturbs standard cell initial coordinates by up to 0.5 microns in both X and Y directions, yielding slightly different but deterministic placements.
Updates all affected global placement regression golden files (.ok and .defok files) to align with the new default seed 1 layout expectations.

Verification

I have verified that the local build succeeds (./etc/Build.sh).
I have run the relevant tests and they pass.
My code follows the repository's formatting guidelines.
I have included tests to prevent regressions.
I have signed my commits (DCO).

gemini-code-assist · 2026-06-11T16:37:31Z

Note

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

maliberty · 2026-06-16T21:24:29Z

+  // microns. We guarantee at least 1 site/row step of shift if perturbation is
+  // enabled.


Why is that guarantee needed? Global placement doesn't legal the result.

It's not needed, you want to simplify to something else?

We could use https://www.boost.org/doc/libs/latest/libs/math/doc/html/math_toolkit/dist_ref/dists/normal_dist.html with a <good value> radius if you'd prefer.

Signed-off-by: Drew Lewis <cannada@google.com> TAG=agy CONV=0fed8794-10c3-4c6f-bdf2-3dedf6c02c30

Update the golden file to match the deterministic placement results produced by the global placement perturbation logic on this branch. The test has been verified to be 100% deterministic and independent of thread count (verified with both 4 and 8 threads). Signed-off-by: Drew Lewis <cannada@google.com>

calewis · 2026-06-24T15:06:56Z

@maliberty all checks now pass and I'm happy with where it's at.

maliberty · 2026-06-24T20:10:54Z

I started a secure CI. This has a merge conflict.

maliberty · 2026-06-24T20:23:38Z

+  // To ensure that 99.5% of the perturbations fall within a circle of radius
+  // r_dbu, we use a Rayleigh distribution for the radial distance R = sqrt(X^2
+  // + Y^2). CDF(r) = 1 - exp(-r^2 / (2 * sigma^2)) = 0.995 => sigma = r_dbu *
+  // sqrt(-1 / (2 * ln(0.005))) sigma \approx 0.3071963263271184 * r_dbu.
+  const double sigma = r_dbu * 0.3071963263271184;
+
+  boost::random::normal_distribution<double> dist(0.0, sigma);


I thought we talked about using a bounded offset. This gives some chance for a large jump that is unexpected.

If you'd like we can just resample until the offset is <= the radius? There is only a 0.5% of a point falling outside the radius so only a 0.003% chance you'd need to sample more than twice. Gaussian is nice because the most probable outcome is that you barely move the point at all.

The perturbation distances in microns look like (0.5 micron radius):

Vector Length Histogram (Rayleigh Distribution) Based on 2D Normal (mean=0, sigma=0.153598) Points: 50000 Min length: 0.000769 Max length: 0.786546 ------------------------------------------ 0.00 | ** (267) 0.03 | **************** (1900) 0.06 | ****************************** (3537) 0.09 | ****************************************** (4884) 0.12 | *********************************************** (5460) 0.15 | ************************************************** (5792) 0.18 | ************************************************* (5752) 0.21 | ********************************************* (5246) 0.24 | ************************************** (4517) 0.27 | ****************************** (3548) 0.30 | ************************* (2952) 0.33 | ****************** (2111) 0.36 | ************ (1458) 0.39 | ******** (1016) 0.42 | ***** (621) 0.45 | *** (400) 0.48 | ** (263) 0.51 | * (124) 0.54 | (67) 0.57 | (38) 0.60 | (22) 0.63 | (8) 0.66 | (9) 0.69 | (4) 0.72 | (2) 0.78 | (2)

With the current code if instead we use rejection based sampling for distances over 0.5 micron:

Truncated Vector Length Histogram Based on 2D Normal (mean=0, sigma=0.153598) Points Accepted: 50000 Points Rejected: 256 (fell outside 0.5 radius) Min length: 0.001130 Max length: 0.499774 ------------------------------------------ 0.00 | * (103) 0.02 | ********** (826) 0.04 | ******************** (1627) 0.06 | **************************** (2315) 0.08 | ************************************* (2976) 0.10 | ****************************************** (3423) 0.12 | *********************************************** (3836) 0.14 | ************************************************** (4009) 0.16 | ************************************************* (3954) 0.18 | ************************************************ (3864) 0.20 | ******************************************** (3558) 0.22 | ***************************************** (3323) 0.24 | ************************************** (3063) 0.26 | ********************************* (2708) 0.28 | *************************** (2225) 0.30 | *********************** (1919) 0.32 | ******************* (1543) 0.34 | *************** (1217) 0.36 | ************ (991) 0.38 | ********* (759) 0.40 | ******* (582) 0.42 | ***** (426) 0.44 | *** (297) 0.46 | ** (231) 0.48 | * (150) 0.50 | (75)

Finally, for completeness the old code, assuming it was using uniform sampling between 0 and 0.5 microns for both x and y:

Uniform Vector Length Histogram [x, y from U(0.0, 0.5)] ------------------------------------------ 0.00 | (18) 0.02 | * (111) 0.04 | **** (255) 0.06 | ****** (368) 0.08 | ******** (509) 0.10 | ********* (572) 0.12 | ************ (766) 0.14 | ************** (875) 0.16 | **************** (1007) 0.18 | ****************** (1092) 0.20 | ******************** (1271) 0.22 | ********************** (1353) 0.24 | ************************ (1475) 0.26 | ************************* (1556) 0.28 | *************************** (1684) 0.30 | ******************************* (1899) 0.32 | ********************************* (2003) 0.34 | ********************************* (2035) 0.36 | ************************************** (2321) 0.38 | *************************************** (2423) 0.40 | ****************************************** (2556) 0.42 | ******************************************** (2692) 0.44 | ********************************************** (2824) 0.46 | *********************************************** (2862) 0.48 | ************************************************** (3031) 0.50 | *********************************************** (2903) 0.52 | *********************************** (2132) 0.54 | **************************** (1742) 0.56 | *********************** (1442) 0.58 | ******************* (1195) 0.60 | **************** (992) 0.62 | ************* (810) 0.64 | ********* (566) 0.66 | ****** (379) 0.68 | *** (216) 0.70 | * (65)

I guess it still falls off sharply enough to not be worth changing.

…turbation Signed-off-by: Drew Lewis <cannada@google.com> # Conflicts: # test/ibex_sky130hs.metrics_limits

Signed-off-by: Drew Lewis <cannada@google.com>

calewis requested review from a team as code owners June 11, 2026 16:37

calewis requested review from LucasYuki and osamahammad21 June 11, 2026 16:37

github-actions Bot added the size/S label Jun 11, 2026

calewis force-pushed the gpl-placement-perturbation branch from fbc9177 to 0eddb5e Compare June 11, 2026 19:21

github-actions Bot added size/M and removed size/S labels Jun 11, 2026

maliberty reviewed Jun 16, 2026

View reviewed changes

calewis marked this pull request as draft June 16, 2026 23:28

calewis force-pushed the gpl-placement-perturbation branch 6 times, most recently from 018a4f0 to ca5112e Compare June 22, 2026 19:57

Implement randomized initial placement perturbation in GPL

616a13d

Signed-off-by: Drew Lewis <cannada@google.com> TAG=agy CONV=0fed8794-10c3-4c6f-bdf2-3dedf6c02c30

calewis force-pushed the gpl-placement-perturbation branch from ca5112e to 616a13d Compare June 22, 2026 20:18

calewis marked this pull request as ready for review June 22, 2026 20:20

calewis force-pushed the gpl-placement-perturbation branch from 173be28 to c86c58e Compare June 24, 2026 11:40

maliberty reviewed Jun 24, 2026

View reviewed changes

calewis added 3 commits June 24, 2026 21:22

Merge remote-tracking branch 'upstream/master' into gpl-placement-per…

985f91f

…turbation Signed-off-by: Drew Lewis <cannada@google.com> # Conflicts: # test/ibex_sky130hs.metrics_limits

test: Update ibex_sky130hs.metrics_limits after upstream merge

18620ba

Signed-off-by: Drew Lewis <cannada@google.com>

Merge upstream/master into gpl-placement-perturbation

9606a74

Signed-off-by: Drew Lewis <cannada@google.com>

calewis requested a review from a team as a code owner June 25, 2026 16:58

calewis requested a review from maliberty June 25, 2026 16:58

run the tests one more time

3280e93

Signed-off-by: Drew Lewis <cannada@google.com>

calewis force-pushed the gpl-placement-perturbation branch from dd4a0ab to 3280e93 Compare June 25, 2026 17:00

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[GPL] Add randomized initial placement perturbation in GPL#10642

[GPL] Add randomized initial placement perturbation in GPL#10642
calewis wants to merge 6 commits into
The-OpenROAD-Project:masterfrom
calewis:gpl-placement-perturbation

calewis commented Jun 11, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot commented Jun 11, 2026

Uh oh!

maliberty Jun 16, 2026

Uh oh!

calewis Jun 16, 2026

Uh oh!

calewis Jun 16, 2026

Uh oh!

calewis commented Jun 24, 2026

Uh oh!

maliberty commented Jun 24, 2026

Uh oh!

maliberty Jun 24, 2026

Uh oh!

calewis Jun 24, 2026

Uh oh!

calewis Jun 24, 2026

Uh oh!

calewis Jun 24, 2026

Uh oh!

maliberty Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		// microns. We guarantee at least 1 site/row step of shift if perturbation is
		// enabled.

Uh oh!

Conversation

calewis commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Impact

Verification

Uh oh!

gemini-code-assist Bot commented Jun 11, 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!

calewis commented Jun 24, 2026

Uh oh!

maliberty commented Jun 24, 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!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

calewis commented Jun 11, 2026 •

edited

Loading