Skip to content

rsz: move fully_rebuffer toward min-delay buffering - #12368

Open
maliberty wants to merge 14 commits into
The-OpenROAD-Project:masterfrom
The-OpenROAD-Project-staging:rsz-rebuffer-size-driver-slack
Open

maliberty wants to merge 14 commits into
The-OpenROAD-Project:masterfrom
The-OpenROAD-Project-staging:rsz-rebuffer-size-driver-slack

Conversation

@maliberty

Copy link
Copy Markdown
Member

Improves fully_rebuffer toward minimum-delay buffering, measured against the buffer_varying_lengths reference harness (openroad/rsz/test/rebuffer_reference/) on Nangate45 and asap7, with the driver fixed and with -rebuffer_size_driver. Each commit message gives its own results.

Commits

  1. Unlock drivers in buffer_varying_lengths_size_driver. The test locked its drivers, so driver sizing could never resize them.
  2. Check the driver cell's max capacitance in rebuffer sizing. Candidate trees are built against their buffers' capacitance limits, so sizing could pick a driver whose limit the tree exceeds. On master, fully_rebuffer with sizing leaves 445 max capacitance violations on black_parrot and 19 on swerv_wrapper; with this commit, none.
  3. Check every driver input before upsizing. Upsizing a multi-input cell slows the nets on its other inputs too.
  4. Aim driver sizing at the best reachable slack on failing paths. The target was anchored to the current driver, so a resize that gained slack lost to cheaper combinations.
  5. Offer trees without their first buffer to driver sizing. Lets an upsized driver replace the buffer placed at a weak driver.
  6. Apply the sizing leakage limit to the whole net. The per-cell limit refused upsizes that lower the net's leakage, e.g. BUF_X1 to BUF_X4 in Nangate45 (ratio 4.02).
  7. Add asap7 buffer_varying_lengths tests and references; PDK settings move to buffer_varying_lengths_pdk.tcl.
  8. No area-recovery relaxation near the worst slack. Area recovery gave up about 1% of the path delay on every net; failing nets within 10% of the worst slack now keep their fastest tree. Removing it on every failing net instead doubled the timing-driven placement buffering area on sky130hs/gcd and broke detailed placement.
  9. Offer evenly spaced chains on long wires. Removes the periodic spikes left by stepping back from the load end.
  10. Plan rebuffer trees for each driver cell on failing paths. With sizing, also buffer the net as though the driver were each faster allowed cell.
  11. Check side loads against the same target in both sizing passes. The first pass checked them against the original target, then raised it, so the cell that set the best slack could fail the second check and leave sizing with the unsized area tree.

Results

Gap to the reference, mean / worst:

Case Before After
Nangate45, driver fixed 3.17 / 18.9 ps 2.02 / 5.6 ps
Nangate45, driver sized 23.9 / 41.4 ps 1.66 / 5.3 ps
asap7, driver fixed 0.81 / 5.9 ps 0.23 / 1.8 ps
asap7, driver sized 9.1 / 15.8 ps 0.53 / 3.1 ps

"Before" is master (for asap7, master with the tests of commit 7).

fully_rebuffer alone on placed designs, sizing on: swerv_wrapper (asap7) TNS -284 -> -83 ps; black_parrot (Nangate45, 4.5 ns) TNS -655 -> -504 ns, failing endpoints 3616 -> 1761, max capacitance violations 445 -> 2 (2 before rebuffering). With sizing off, commit 9 costs black_parrot 21 ns of TNS (-720 -> -741 ns) while improving the reference gaps and swerv_wrapper. Runtime changes are within +8% per commit; see the commit messages. sky130hs/gcd passes through CTS. repair_timing -setup is unchanged on gcd, swerv_wrapper, and black_parrot.

Golden updates follow from the rebuffer changes: the buffer_varying_lengths tests, rebuffer_size_driver*, and, through the repair_timing rebuffer move, clone_flat, repair_fanout7*, repair_fanout8*, repair_setup2*, repair_setup_invalid_phase, and repair_setup_path4; and, through timing-driven placement (commit 8), the gpl tests simple01-td-repair, simple01-td-scan-net, and simple01-vcts-clamp.

The test builds its drivers as LOCKED, and rebuffering does not resize
fixed instances, so -rebuffer_size_driver could never resize them. Unlock
the drivers after repair_design in the driver-sizing variant, and check
after each rebuffering step that no driver moved, which the lock was
there to prevent.

Rebuffering now resizes 22 drivers. The mean gap to the reference
(rebuffer_reference/run_reference.py --size-driver) drops from 23.9 ps
to 21.4 ps; the worst gap is unchanged at 41.4 ps.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
With -rebuffer_size_driver, slackWithDriverCell() checked the driver
cell's output slew and load slew but not its max capacitance. The
candidate trees are built against the capacitance limits of their
buffers, so a tree could be chosen with a driver cell whose limit it
exceeds, leaving max capacitance violations for repair_design to fix
again. Reject such combinations, with the same margin as the buffers.

fully_rebuffer alone, sizing on, before -> after:
- black_parrot (nangate45, 4.5 ns clock): max capacitance violations
  445 -> 0, TNS -655 -> -670 ns, failing endpoints 3616 -> 3953,
  area +0.5%, runtime +8%
- swerv_wrapper (asap7): max capacitance violations 19 -> 0, TNS
  unchanged at -284 ps
- gcd and the rsz unit tests: unchanged

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
upstreamSideLoadsOk() checked only the other loads of the net driving
the input on the timed path. Upsizing a multi-input cell raises the
capacitance of every input, slowing the nets on the other inputs and
the paths through them. Check those nets' loads too.

With the current area-first selection this rejects a few upsizes
(15 instead of 19 on rebuffer_size_driver) and leaves TNS unchanged.
It is needed for slack-first selection: without it, upsizing on gcd
worsens TNS at every clock period tried.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
With -rebuffer_size_driver, sizeDriverWithTree() picks the smallest
(tree + driver) area that meets slack_target. That target is the slack
of the timing-optimized tree with the current driver, less the
relaxation fullyRebuffer() allows for area recovery (about 1% of the
path delay when slack is negative). Because it is anchored to the
current driver, any resize that gains slack competes with cheaper
combinations that only match the current driver, and loses on area: on
a failing path sizing could trade slack for area but not area for
slack.

When slack_target is negative, anchor it instead to the best slack any
candidate tree reaches with any driver cell that passes the side-load
check, less the same relaxation. The selection is unchanged: the
smallest area meeting the target wins, but only among combinations
within the relaxation of the best. For example, with the current driver
at -100 ps, a relaxation of 2 ps, and an upsized driver reaching -80 ps,
the target moves from -102 ps to -82 ps. Passing paths are unchanged.

Sizing on, before -> after:
- rebuffer_size_driver (gcd, 0.5 ns): TNS -1.43 -> -1.15 ns, area +3.2%
- swerv_wrapper (asap7): TNS -284 -> -93 ps (sizing off: -126 ps)
- black_parrot (nangate45, 4.5 ns clock): TNS -678 -> -585 ns,
  failing endpoints 4145 -> 1947
- buffer_varying_lengths_size_driver: mean gap to the reference
  21.4 -> 15.5 ps, worst 41.4 -> 26.3 ps
Rebuffer runtime is unchanged on swerv_wrapper and +5% on black_parrot.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
Rebuffering often places a buffer right at a weak driver, which does the
job of a larger driver. Driver sizing only re-timed the trees built for
the current driver, so it never saw the same tree without that buffer
and could not replace the buffer with an upsized driver. For each
candidate tree whose first node below the driver is a buffer reached
through wires only, also offer the tree with that buffer removed and
the driver wired straight to the buffer's load.

Sizing on, before -> after:
- buffer_varying_lengths_size_driver: mean gap to the reference
  15.5 -> 11.6 ps
- swerv_wrapper (asap7): TNS unchanged at -93 ps, 8% fewer buffers
- black_parrot (nangate45, 4.5 ns clock): TNS -585 -> -576 ns,
  failing endpoints 1947 -> 1876, 8% fewer buffers
- rebuffer_size_driver (gcd): unchanged
Rebuffer runtime is unchanged on swerv_wrapper and +2% on black_parrot.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
The sizing leakage limit filters swappable cells by the ratio of their
leakage to the current cell's. For a driver this refuses an upsize even
when it replaces buffers and lowers the net's leakage. With the default
4x limit it also bars BUF_X1 from growing to BUF_X4 in Nangate45, whose
leakage ratio is 4.02.

In rebuffer's driver sizing, consider every swappable cell except those
only the leakage limit excludes, and instead limit the leakage of the
driver and its buffers together to the limit times that of the current
driver with the timing tree. getSwappableCells() keeps the per-cell
limit for all other callers; its cache now holds both lists.

Sizing on, before -> after:
- buffer_varying_lengths_size_driver: mean gap to the reference
  11.6 -> 2.0 ps, worst 25.4 -> 8.2 ps; leakage of drivers and buffers
  +7% (the extra comes from buffer choices, not drivers)
- black_parrot (nangate45, 4.5 ns clock): TNS -576 -> -554 ns,
  failing endpoints 1876 -> 1810
- swerv_wrapper (asap7), rebuffer_size_driver (gcd): unchanged

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
Add buffer_varying_lengths_asap7 and buffer_varying_lengths_asap7_size_driver
so rebuffering is measured on a second PDK. The PDK-specific settings
(libraries, buffers, layers, length step, clock period, and the
reference search seeds) move to buffer_varying_lengths_pdk.tcl, shared
by the tests and rebuffer_reference/; the nangate45 tests' output is
unchanged. On asap7 the paths are 3.75 um * i long (up to 375 um, about
five buffer spacings on M3) and the clock period is 1000 ps.

run_reference.py takes --pdk; references are now
reference_<pdk>_<case>.txt, with new asap7 references. Gaps to the
reference with this branch:
- asap7 fixed driver: mean 0.81 ps (0.6%), worst 5.9 ps
- asap7 driver sizing: mean 0.68 ps (0.5%), worst 3.2 ps
  (master: mean 9.1 ps, worst 15.8 ps)

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
fully_rebuffer let area recovery give up about 1% of the path delay on
every net, so on the most critical nets a smaller tree that the model
rated only slightly slower won over the fastest one. A failing net
whose slack before rebuffering is within 10% of the design's worst
slack now gets no relaxation. Other failing nets keep it: removing it
on every failing net doubled the rebuffer area in timing-driven global
placement on sky130hs/gcd (+21% -> +47%), and detailed placement then
failed after CTS repair_timing at 99.9% utilization.

Gap to the buffer_varying_lengths reference (mean / worst):
  nangate45 fixed driver  3.17 / 18.9 ps -> 3.09 / 18.9 ps
  nangate45 sized driver  1.96 /  8.2 ps -> unchanged
  asap7 fixed driver      0.81 /  5.9 ps -> 0.76 /  5.8 ps
  asap7 sized driver      0.68 /  3.2 ps -> 0.52 /  2.7 ps
Each length is its own path, so only the longest are near the worst
slack.

fully_rebuffer alone on placed designs, before -> after:
  swerv_wrapper asap7, sizing on:
    WNS -31.7 -> -28.4 ps, TNS -93.3 -> -83.0 ps
  black_parrot nangate45 at 4.5 ns, sizing on:
    WNS -5.06 -> -4.99 ns, TNS unchanged at -554 ns
  swerv_wrapper, black_parrot, and gcd, sizing off: unchanged
  gcd 0.40-0.60 ns, sizing on: unchanged
sky130hs/gcd through CTS: rebuffer area in global placement unchanged
(+21%); after CTS repair_timing, utilization 76.6% -> 76.5%, WNS
-0.208 -> -0.216 ns, TNS -5.2 -> -5.8 ns.
Runtime and area are within noise.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
Long-wire stepping places buffers back from the load end at the
asymptotic spacing, leaving the remainder on the upstream stage. On a
weak driver that stage should be shorter than the others, and on some
lengths one more buffer fits. For each buffer size, also offer the
stepped buffer count and one more, evenly spaced, with the upstream
stage 1, 0.5 or 0.25 of the others. Only segments longer than one
spacing build these chains. The long-wire stepping timer is now
accumulated.

Gap to the buffer_varying_lengths reference (mean / worst):
  nangate45 fixed driver  3.09 / 18.9 ps -> 2.02 / 5.6 ps
  nangate45 sized driver  1.96 /  8.2 ps -> 2.10 / 9.2 ps
  asap7 fixed driver      0.76 /  5.8 ps -> 0.23 / 1.8 ps
  asap7 sized driver      0.52 /  2.7 ps -> 1.08 / 3.2 ps
With driver sizing, the short upstream stage suits only the current
driver; planning per driver cell recovers it in the next commit.

fully_rebuffer alone on placed designs, before -> after:
  black_parrot nangate45 at 4.5 ns, sizing on:
    TNS -554 -> -541 ns, failing endpoints 1810 -> 1803
  black_parrot, sizing off:
    TNS -720 -> -741 ns, failing endpoints 4595 -> 4715
  swerv_wrapper asap7, sizing off: TNS -126.5 -> -123.8 ps
  swerv_wrapper, sizing on: unchanged
  gcd: unchanged
Runtime +2-3%, area within 0.05%.

repair_timing -setup is unchanged on gcd, swerv_wrapper, and
black_parrot. In the unit tests its rebuffer move shifts results both
ways: repair_setup2 WNS -0.12 -> -0.11 ns, repair_fanout7/8 and
clone_flat WNS 0.01-0.06 ns worse.

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
With -rebuffer_size_driver, every candidate tree was planned for the
current driver cell and only retimed with the others. With the evenly
spaced chains of the previous commit, the tree planned for a weak
driver starts with a short stage that a larger driver does not need.
On a failing path, also buffer the net from scratch (timing and area
recovery) as though the driver were each allowed cell that speeds up
the timing tree, and offer those trees to driver sizing. Driver timing
for another cell is shared with slackWithDriverCell() through
cellArcTiming().

Gap to the buffer_varying_lengths reference, driver sized
(mean / worst):
  nangate45  2.10 / 9.2 ps -> 1.66 / 5.3 ps
  asap7      1.08 / 3.2 ps -> 0.53 / 3.1 ps
Fixed-driver results are unchanged.

fully_rebuffer alone on placed designs, sizing on, before -> after:
  black_parrot nangate45 at 4.5 ns:
    TNS -541 -> -508 ns, failing endpoints 1803 -> 1754,
    runtime 39.6 -> 42.7 s (+8%)
  swerv_wrapper asap7: unchanged
  gcd: unchanged

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
On a failing path, bestSlackWithDriver() accepted a driver cell whose
upstream side loads stayed above the original slack target, then raised
the target to that cell's slack.  chooseTreeAndDriver() checks the side
loads against the raised target, so the cell that set it could be
rejected there.  When no other combination met the target, sizing fell
back to the unsized area tree.

upstreamSideLoadsOk() becomes upstreamSideLoadSlack(), which returns the
worst slack the resize leaves on a side load it slows.
bestSlackWithDriver() now accepts a candidate only when that slack meets
the target the candidate would set (its slack less the relaxation), so
the cell that sets the best slack always passes the second check.

Sizing on, before -> after:
- black_parrot (nangate45): TNS -508.11 -> -503.50 ns, failing
  endpoints 1754 -> 1761, area -0.002%
- swerv_wrapper (asap7), rebuffer_size_driver (gcd sweep), and the
  buffer_varying_lengths tests: unchanged

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
…river-slack

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>

# Conflicts:
#	openroad/rsz/test/buffer_varying_lengths_asap7.ok
#	openroad/rsz/test/buffer_varying_lengths_asap7.tcl
#	openroad/rsz/test/buffer_varying_lengths_asap7_size_driver.ok
#	openroad/rsz/test/buffer_varying_lengths_asap7_size_driver.tcl
#	openroad/rsz/test/buffer_varying_lengths_pdk.tcl
#	openroad/rsz/test/rebuffer_reference/common.tcl
#	openroad/rsz/test/rebuffer_reference/reference_asap7_fixed.txt
#	openroad/rsz/test/rebuffer_reference/reference_asap7_size_driver.txt
#	openroad/rsz/test/rebuffer_reference/reference_fixed.txt
#	openroad/rsz/test/rebuffer_reference/reference_nangate45_fixed.txt
#	openroad/rsz/test/rebuffer_reference/reference_nangate45_size_driver.txt
#	openroad/rsz/test/rebuffer_reference/reference_size_driver.txt
#	openroad/rsz/test/rebuffer_reference/rsz_solution.tcl
#	openroad/rsz/test/rebuffer_reference/sweep.tcl
#	src/rsz/test/rebuffer_reference/reference_fixed.txt
#	src/rsz/test/rebuffer_reference/reference_size_driver.txt
…river-slack

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
@maliberty maliberty self-assigned this Nov 19, 2026
@github-actions

Copy link
Copy Markdown
Contributor

This PR exceeds the recommended size of 1000 lines. Please make sure you are NOT addressing multiple issues with one PR. Note this PR might be rejected due to its size.

@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 integrates driver sizing with buffer tree generation during rebuffering in the Resizer tool, enabling timing, area, and leakage budget evaluations for driver resizes. It also optimizes long wire buffering by offering evenly spaced buffer chains. The review feedback suggests using .value() instead of .value_or() to avoid silently hiding impossible states with default values, and adding a guard in stepToward to prevent a potential division by zero when the distance is zero.

Comment on lines +239 to +240
const CellArcTiming timing = cellArcTiming(rf, plan_cell_, bnet->cap())
.value_or(CellArcTiming{});

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.

medium

Using .value_or(CellArcTiming{}) here silently defaults to a zero-valued CellArcTiming if cellArcTiming returns std::nullopt. Since planForDriverCell only plans for cells with matching arcs, cellArcTiming returning std::nullopt represents an impossible state. According to the OpenROAD code-review guidance, we should avoid guards that hide impossible states with silent defaults. Using .value() instead will ensure a loud failure if this invariant is ever broken.

      const CellArcTiming timing = cellArcTiming(rf, plan_cell_, bnet->cap()).value();
References
  1. Never suggest a guard that hides an impossible state: a null check, an empty, zero or negative check, or a missing-key check that returns a default, skips, clamps, continues or logs a warning and carries on. A null dereference that is reachable is reported with its triggering input; the fix is logger_->error(...) or an assert, never a silent default. (link)

Comment on lines +1205 to +1210
static odb::Point stepToward(const odb::Point& from,
const odb::Point& to,
const int distance)
{
int dx = to.x() - from.x();
int dy = to.y() - from.y();

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.

medium

If from and to are the same point (i.e., dx == 0 and dy == 0) and distance is 0, the condition abs(dx) + abs(dy) >= distance evaluates to 0 >= 0 (true). This leads to a division by zero when calculating ratio. Adding a guard to return from immediately when distance <= 0 prevents this potential division by zero and makes the utility function more robust.

static odb::Point stepToward(const odb::Point& from,
                             const odb::Point& to,
                             const int distance)
{
  if (distance <= 0) {
    return from;
  }
  int dx = to.x() - from.x();
  int dy = to.y() - from.y();

@maliberty
maliberty marked this pull request as ready for review November 19, 2026 21:41
@maliberty
maliberty requested review from a team as code owners November 19, 2026 21:41
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 19, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ❌ Failed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ❌ Failed 2026-11-19T21:41:11.254822Z
We couldn’t complete this review. Comment @codex review to try again.
fa42fc4 Draft marked ready
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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@maliberty
maliberty requested review from jhkim-pii and removed request for precisionmoon November 19, 2026 21:41
@maliberty

Copy link
Copy Markdown
Member Author

I've been looking at the CI failures and they are mostly noise amplifying (ie bad luck).

With -rebuffer_size_driver, bestSlackWithDriver() and
chooseTreeAndDriver() each timed every candidate tree with every
candidate driver cell, and the trees planned per driver cell grow with
the cell count.  Each evaluation recomputed the driver's input slew and
the previous stage's drive resistance, looked up the cell's arc, and
walked the tree for its buffer leakage.

Now the input slew and previous-stage drive resistance are cached per
net and the equivalent arc per cell, each tree's buffer leakages are
collected once, and a tree/cell slack and a cell's side-load slack are
computed on first use and shared by both passes.  The leakage sum keeps
its order, so results are unchanged.

Global placement with -rebuffer_size_driver true on rapidus2hp
(3_3_place_gp.odb bit-identical before and after):
  jpeg              7:09 -> 3:35  (master 1:20)
  cva6              5:15 -> 3:34  (master 2:12)
  hercules_idecode 19:13 -> 10:34 (master 4:42)

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant