EDA Standard-Cell Gate Sizing Optimization

Given a placed netlist described by CircuitOps IR tables, assign a library cell to every instance so as to minimize one score charging timing (TNS) and design-rule (slew / capacitance) violations, plus the leakage you spend clearing them. Your submission is re-run on sealed hidden cases; lower is better.

/app/methods/main/ ships a working but crude reference implementation. It is there to define the zero point and to show the plumbing, not to be the starting shape of your answer. You may delete all of it and design your method from scratch. The only fixed parts of a submission are the solve(...) contract and the .size output format below. What your method looks at, what it decides per instance, and how it searches are entirely yours to choose.

Hard Constraints

• Submit an algorithm (solve), not precomputed answers. Do not key on case names.
• Keep the exact signature solve(input_dir: str, output_path: str) -> None: read one case from input_dir, write a valid .size file to output_path.
• Sequential cells and macros keep their original library cell. Only combinational cells may be resized.
• A combinational cell may only move to a cell with the same func_id (logically equivalent, different drive strength or threshold), read from libcell_properties.csv.bz2.
• List every instance exactly once, with no unknown instance or library-cell names. An illegal sizing scores nothing.
• solve must be deterministic and case-name-independent.

What You Have

• Visible cases at /app/data/<case>/ for each case in /app/data/manifest.json, each with IR_Tables/ (cell_properties, libcell_properties, pin_properties, net_properties and the graph edge tables, all *.csv.bz2) and design/ (compressed .v/.def/.sdc). The shared library is at /app/data/platform/ASAP7. Hidden cases are different designs, same format.
• /app/methods/main/ — this directory is what gets graded. The shipped solver.py handles the contract and the .size format and applies a uniform strongest-drive upsize. Its score is the zero point: matching it earns nothing, and scoring worse earns nothing either. Rewrite it, throw it away and start over, or add any helper Python next to it (python3, numpy, pandas, scikit-learn available).
• Your self-check surface (free, unlimited): python /app/selfcheck.py runs your current solver.py on the visible cases through the same legality gate and the same OpenROAD score the sealed grader uses, and prints the breakdown into 10*|TNS|, 20*slew, 20*cap and leakage. A case name limits it to that case (python /app/selfcheck.py <case>); the scoring pass takes minutes on the large cases.

What You Submit

/app/methods/main/solver.py must expose this exact signature. It and the .size format below are the only parts of a submission that are fixed:

def solve(input_dir: str, output_path: str) -> None:
    ...

The .size output is one line per instance:

<instance name> <library cell name>

There is no submit step and no per-attempt feedback. Work and self-check for as long as your run window allows, then leave your best solver.py in place; it is graded once at the end.

How It Is Judged

The grader copies methods/main/ into a clean sandbox, runs solve(...) on each hidden case, checks legality, then evaluates the sizing with a timing engine. The per-case score is

score = leakage_delta_uW + 10*|TNS_ns| + 20*slew_violation + 20*cap_violation

leakage_delta_uW is your leakage minus the original netlist's, so leaving a cell alone costs nothing. TNS is total negative slack; slew and cap are the summed amounts by which pins exceed their transition-time and load-capacitance limits. Each penalty term switches off entirely once its violation reaches zero, so driving a category to exactly 0 is worth more than driving it low. Scoring is per case against that case's own reference points, then averaged, so a case you ignore cannot be carried by one you optimize. Lower is better. Your solver's wall-clock time is not scored, though it is capped.