Timing Closure

Timing Closure

A disciplined loop that drives a Vivado design to timing closure using the SynthPilot MCP tools. You are the conductor: measure → diagnose → classify → apply the safest fix → re-measure → repeat. Never make a number go green by hiding a real violation.

When to use

• "close timing", "fix WNS/TNS", "why doesn't this meet timing", "resolve setup violations"
• After synthesis or implementation reports negative slack.

When NOT to use

• No project open / no design to analyze → set the project up first (or use full-flow-demo).
• The user wants a one-shot report only, not iteration → just call report_timing_summary.
• Hold (min-delay) violations dominate and routing is already done → these usually need Vivado's own hold-fix; surface that rather than looping.

Prerequisites (verify first)

• test_connection succeeds (Vivado + Tcl server up).
• A project is open (get_project_info). If only RTL exists, run synthesis first.

The loop

Step 0 — Baseline (measure, don't guess)

1. Ensure a timed design exists. If none, run_synthesis (or run_synthesis_async + poll get_run_status for large designs). Synthesis WNS is optimistic — true signoff is post-implementation, but synth WNS is a fast first signal.
2. Read the numbers: report_timing_summary, then extract_timing_metrics for structured WNS/TNS/WHS/THS, failing-endpoint count, and per-clock breakdown.
3. Record the baseline (WNS, TNS, #failing endpoints, clocks involved). Every later iteration is compared against this. State it to the user.

If WNS ≥ 0 post-implementation: report closed, show the margin, stop. (Post-synth positive is necessary but not sufficient — confirm with run_implementation.)

Step 1 — Diagnose the worst paths

• analyze_critical_paths and report_worst_timing_paths → the top failing paths.
• For the worst path, report_timing_path / report_datapath_delay to split delay into logic vs net vs clock-skew/uncertainty.
• report_high_fanout_nets → fanout-driven net delay.
• get_all_clocks / analyze_clock_domain_timing / report_clock_interaction → is the failing path intra-clock, or a cross-clock / async path that shouldn't be timed at all?

Step 2 — Classify the failure (this decides the fix)

Map the worst paths to one of these, in this priority order:

Step 3 — Apply the smallest safe change

• Constraints (A/B): apply the specific exception, then save_constraints. Before any set_false_path/set_multicycle_path, state the assumption ("treating CDC path reg_a→reg_b as async; valid only if a synchronizer exists") and get confirmation if the path could be real.
• Strategy (C/E): set_implementation_strategy (e.g. Performance_ExplorePostRoutePhysOpt) / set_synthesis_strategy. Cheapest lever, try before touching RTL.
• RTL (D): describe the pipelining/retiming change and the expected slack gain; recommend, don't silently rewrite — hand back to the user to approve RTL edits.
• Apply one class of change per iteration so cause→effect is attributable.

Step 4 — Re-measure

• Re-run the affected stage (run_implementation for signoff; run_synthesis if you changed RTL/synth strategy). Use the async variants + get_run_status for long runs.
• Re-read WNS/TNS via extract_timing_metrics. Compare to baseline and to the previous iteration. Did it improve, regress, or stall?

Step 5 — Loop or stop

Repeat Steps 1–4 until one stop condition:

• ✅ Closed: WNS ≥ 0 (and WHS ≥ 0) post-implementation → report margin, list the changes applied, save constraints, stop.
• ⛔ No safe move left: remaining failures are class D (real logic depth) or require risky exceptions → stop, present the options + trade-offs, hand to the human.
• ⛔ Stalled: 2 consecutive iterations with no WNS improvement → stop, don't thrash.
• ⛔ Budget: a caller-specified max iterations reached.

Safety rails (do not violate)

1. No fake-pass. Never apply set_false_path/set_multicycle_path/waivers to a path that could be real just to clear a violation. If unsure it's safe, ask.
2. Evidence before claims. "Timing met" requires a fresh post-implementation extract_timing_metrics showing WNS ≥ 0 — never infer it.
3. Show your work. Each iteration: report what you changed, why (which class), and the before/after WNS/TNS. Keep a running summary table.
4. One change class per loop so a regression is traceable.
5. Hands off RTL unless the user approves the specific edit.

Output (each iteration + final)

Iter N | change: <what> (class <X>) | WNS <before> → <after> ns | TNS <before> → <after> | failing <n>→<m>

Final: closed/not-closed verdict, the ordered list of changes that worked, remaining risks, and (if not closed) the recommended RTL/architectural next step.