Add Python API for Resizer (rsz)

[rohithsiddi]

Summary

Adds a Python API for the Resizer (rsz) tool so that flows can use rsz_aux and the rsz module instead of Tcl for buffer_ports, repair_design, repair_timing, and related commands.

Motivation

The Resizer had no Python bindings; Python flows had to call Tcl commands via evalTclString. This change brings rsz in line with other tools (e.g. grt, cts, drt, pdn) that expose a Python API.

Changes

• Build: New rsz_py SWIG module in src/rsz/src/CMakeLists.txt using Resizer-py.i; linked in src/CMakeLists.txt; rsz added to Python module load list in src/Main.cc.
• Bindings: Resizer-py.i exposes snake_case wrappers for Resizer commands (opaque STA types; no Tcl-specific code).
• Wrapper: src/rsz/test/rsz_aux.py with full command coverage, docstrings, and correct unit conversions (microns↔metres, ns↔seconds, percent↔fraction).
• Tests: helpers.py and gcd_resize.py mirroring the existing Tcl test; run via CTest as rsz.gcd_resize.py.
• Docs: New "Using the Python interface to rsz" section in src/rsz/README.md.

Testing

• ctest -R "rsz.gcd_resize.py" — pass
• ctest -R "rsz" — all RSZ tests (Tcl + Python) pass

[gemini-code-assist]

Code Review

This pull request introduces a Python API for the Resizer (rsz) tool, which is a great step towards unifying the scripting interface across OpenROAD tools. The changes are well-structured, including build system updates, SWIG bindings, a Pythonic wrapper, tests, and documentation.

My review focuses on correctness and memory safety. I've found a few areas in the SWIG interface file (Resizer-py.i) where using smart pointers (std::unique_ptr) would improve exception safety and prevent potential memory leaks. These are detailed in the specific comments.

Overall, this is a solid contribution that significantly improves the usability of the Resizer from Python-based flows.

_{Note: Security Review is unavailable for this PR.}

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[maliberty]

We should have a python API for rsz. This PR however tries to impose an interface the mirrors tcl which is not desirable. In Python we intend to expose an object-oriented API rather than a command oriented API. Please look at other modules for examples.

[rohithsiddi]

Thanks for the feedback — understood. I’ll pivot the architecture to follow the object-oriented Python API pattern used in the other modules instead of the Tcl-like command wrappers.

I'll study the preferred modules to align with your exact design patterns and push an updated structure to this branch shortly.

[rohithsiddi]

@maliberty I've addressed the review comments. Summary of the changes below.

What I updated

• Kept the RSZ Python interface object-oriented (design.getResizer().<method>), aligned with other tool bindings.
• Simplified Resizer-py.i style to match existing *-py.i files (minimal comments, no behavior change).
• Refined rsz_aux.py interface details:
  • removed unused verbose argument from report_floating_nets
  • clarified Python-facing error messages (removed Tcl-style flag prefixes)
• Updated gcd_resize.py call site accordingly (report_floating_nets(design)).

Validation

• Rebuilt successfully.
• RSZ Python test passes:
  • ctest --output-on-failure -R "rsz.gcd_resize.py"

[maliberty]

Why are there so many %ignore ?

[rohithsiddi]

The higher %ignore count in Resizer-py.i compared to GRT/CTS is because
Resizer.hh is deeply integrated with OpenSTA internals:

• ~21 methods return/take STA collection types (NetSeq*, PinSet*, vector)
  that have no Python SWIG typemaps — these are either ignored or replaced with
  %extend wrappers (e.g. findFloatingNetsCount(), string-based repairSetup).
• ~5 methods use output-by-reference params (Delay&, Slew&, map<Pin*, float>&)
  not cleanly expressible in Python.
• Internal/GUI hooks (postReadLiberty, setDebugGraphics, lib_data_) follow
  the same pattern as GRT's init/initGui/setRenderer ignores.

GlobalRouter gets away with only 5 %ignores because its API mostly uses
odb::db* types already registered by the odb SWIG module.

[maliberty]

Look at odb where we have

  void getGridPatternX(int i, int& origin_x, int& line_count, int& step);

and map it with

%apply int& OUTPUT { int& origin_x, int& origin_y, int& line_count, int& step };

[rohithsiddi]

@maliberty Thanks for the feedback — I updated Resizer-py.i accordingly.

Resizer-py.i: replaced %ignore with %apply OUTPUT where applicable

For bufferWireDelay, I removed %ignore and added:

%apply float& OUTPUT { sta::Delay& delay, sta::Slew& slew };

sta::Delay and sta::Slew are aliases of float, so this follows SWIG's standard float& OUTPUT mapping pattern (same style as in odb/src/swig/common/dbtypes_common.i). From Python, bufferWireDelay(...) now returns (delay, slew).

I also removed %ignore from methods that SWIG can wrap directly without extra typemaps:

• parseMove (static, returns MoveType)
• repairSetup(const sta::Pin*) (single-pin testing overload)
• resetInputSlews (no args)
• getSlewRCFactor (returns float)

Why the remaining %ignore entries stay

The remaining %ignore methods are cases where %apply is not applicable:

• Return STA container/internal types (sta::NetSeq, sta::PinSet, sta::VertexSeq, std::optional<sta::Slack>) that are not directly wrapped in this module
• Use C-style fixed-size output arrays (e.g. ArcDelay[2]) — %apply works on scalar T&, not T[N]
• Use std::map<const Pin*, float>& output params
• Require sta::Scene* inputs, which are internal STA context objects not available from Python
• Internal lifecycle/GUI hooks (init, constructor, postReadLiberty, setDebugGraphics), consistent with other tool bindings

For findFloatingNets, findFloatingPins, findOverdrivenNets, and repairSetup(vector<MoveType>), %ignore is paired with %extend wrappers that expose Python-usable alternatives (count helpers and string-based sequence input).