constraints-authoring

Author a correct Vivado XDC timing constraint set from scratch — primary clocks, generated (derived) clocks, input/output delays, and timing exceptions (clock groups / false paths / multicycle paths)

SKILL.md

GitHub

Constraints Authoring (XDC from scratch)

A methodology playbook for writing a correct, complete, and honest XDC constraint set. The goal of constraints is to tell the timing engine the truth about the design's real timing requirements — not to make WNS look good. A constraint that lies (a too-loose clock, a false_path on a real path) produces a green report that does not survive on silicon. Correctness first, optimism never.

This skill stops at a timed, sane, signed-off constraint set plus a clear list of any assumptions you had to make. It does NOT attempt timing closure (that is the timing-closure skill's job). It hands off once timing is defined, not met.


When to use

  • A project has no XDC at all, or no timing constraints (only pin/IO assignments) — the classic "no constraints yet".
  • check_timing / report_timing_summary reports unconstrained clocks, ports, or combinational endpoints (e.g. "There are N input/output ports with no input/output delay specified").
  • The design has PLL/MMCM/clock-divider outputs that are missing generated-clock definitions (derived clocks defaulting to the source period, or showing up as unconstrained).
  • Multiple asynchronous clock domains exist and inter-clock paths are being analyzed (and failing) when they should be cut with set_clock_groups.
  • You need to add I/O timing budgets (set_input_delay / set_output_delay) for a board interface.
  • A legitimate multicycle or false relationship exists that the tool cannot infer and is currently over-constraining.

When NOT to use

  • Timing is already fully constrained and you are trying to close it (improve WNS/TNS). Use the timing-closure skill — do NOT "fix" timing by adding exceptions here.
  • CDC is the question ("is my clock crossing safe?"). Constraining async clocks with set_clock_groups silences the cross-domain timing report; it does not make the crossing safe. Use the cdc-analysis skill to verify synchronizers exist before/after you group clocks. (This skill will tell you when to stop and call it.)
  • Pure pin/placement assignment (LOC/IOSTANDARD only, no timing). Those are create_io_constraint; that part is in-scope here but if there is no timing question at all this skill is overkill.
  • You don't know the board/interface timing numbers. Do not invent input/ output delays. Ask the user for the datasheet/board numbers, or state the assumption explicitly and flag it.

Prerequisites (verify first)

Run these BEFORE writing anything. Do not skip — a wrong period or a clock on a non-existent port wastes a full synthesis.

  1. test_connection — confirm the MCP↔Vivado Tcl bridge is alive. If this fails, stop: nothing below will work.
  2. get_project_info — confirm a project is open and know the target part (speed grade matters; constraints are part-agnostic but I/O delay budgets are not).
  3. list_source_files + set_top_module (if not set) — you must constrain the top-level ports. Constraining ports that aren't on top is a silent no-op.
  4. list_constraint_files — find what already exists. If there is an XDC, you are editing, not creating from scratch; read_file it first and do not duplicate clock definitions.
  5. Establish the design intent — from the user and/or read_file on the top module: every clock port and its frequency, which clocks are generated (PLL/ MMCM/divider), which clock domains are asynchronous to each other, and the board-level I/O timing (source-synchronous vs system-synchronous). If any of these is unknown, ask the user — do not guess clock periods.

You author constraints at elaboration/RTL level (against port and pin names), but you VERIFY them against a real synthesized netlist where pin paths (pll/CLKOUT0, reg[*]/C) actually exist. Authoring uses port names; verification needs open_synthesized_design.


Methodology

Author in dependency order: clocks → generated clocks → I/O delays → exceptions. Each later layer references names defined by an earlier one, so order is not optional. One constraint class per step, then re-measure, so you can attribute every change in the timing report to exactly one edit.

1. Inventory what exists and what's missing

  • check_timing — the single most important diagnostic. It reports unconstrained clocks, missing input/output delays, combinational loops, and (critically) appends the full port inventory so you can cross-reference which ports still lack constraints. Read it carefully; it is your worklist.
  • get_all_clocks — what clocks (if any) Vivado has already inferred. Empty = truly from scratch. Non-empty = some clocks exist (maybe from IP); do not redefine them.
  • If clocks already exist, get_clock_info(<name>) per clock to see period and source — confirm they match intent before building generated clocks on top.

If check_timing cannot run because nothing is elaborated, run run_synthesis_async once to get a netlist, poll get_run_status, then open_synthesized_design. You need a netlist to see real clock pins anyway.

2. Define PRIMARY clocks (the foundation)

For every physical clock input port (an oscillator/board clock entering the FPGA), one create_clock_constraint:

  • create_clock_constraint(clock_port="<top clock port>", period_ns=<period>, name="<clk_name>")
  • Period = 1000 / freq_MHz. Use the real board frequency from intent, not a hopeful one. A 100 MHz oscillator is period_ns=10.0.
  • Only create primary clocks on true clock inputs, never on internal PLL/MMCM outputs — those are generated clocks (Step 3). Defining a primary clock on a PLL output double-counts and corrupts analysis.
  • Give every clock an explicit name so later steps and reports are readable.

After defining all primary clocks, get_all_clocks to confirm each appears with the intended period.

3. Define GENERATED clocks (PLL/MMCM/divider outputs)

Any clock derived inside the FPGA needs an explicit generated clock so the engine knows the true frequency/phase relationship:

  • PLL/MMCM outputs: create_generated_clock(name="<clk_out>", source="<pll>/CLKIN1", target="<pll>/CLKOUT0", multiply_by=M, divide_by=D) — M/D from the IP's actual configuration (e.g. 100→200 MHz is multiply_by=2).
  • RTL clock dividers: create_generated_clock(name="clk_div2", source="<clk_in port>", target="<div_reg>/Q", divide_by=2).
  • Phase-shifted outputs: add phase=<deg> (e.g. a 90° clock for source-sync DDR).
  • Inverted/negedge-derived clocks: invert=True.
  • Use multiply_by/divide_by for simple ratios; edges/edge_shift only for genuinely non-integer or asymmetric waveforms.

Source/target are PIN paths, not ports — they only exist on the synthesized netlist. Author them, then prove them in Step 7 with report_clock_networks / get_all_clocks. If a generated clock shows up with the wrong period after synthesis, the M/D or the target pin is wrong — fix it, don't paper over it.

Many IP cores (Clocking Wizard, MIG, transceivers) ship their OWN generated clocks in their .xdc. Check report_clock_networks / the IP's constraints first; do not redefine a clock the IP already constrains, or you create a conflicting double definition.

4. Relate or separate the clocks (groups / uncertainty)

Now that all clocks exist, declare their relationships — this is where most "failing paths" on a fresh design come from, and where the biggest honesty risk lives.

  • Asynchronous domains (clocks from independent sources / unrelated PLLs): set_clock_groups(group1="<clkA>", group2="<clkB>", relationship="asynchronous"). This tells timing not to analyze paths between them. HARD GATE: only do this for a crossing that is genuinely asynchronous AND has a proper CDC synchronizer. Grouping clocks does not synchronize them — it hides the report. See the classification table and Safety rails. If unsure whether a synchronizer exists, stop and ask / run cdc-analysis; do not group blindly.
  • Mux-selected / never-simultaneous clocks: use relationship="exclusive" (or logically_exclusive/physically_exclusive) — not asynchronous.
  • Synchronous, related clocks (e.g. PLL outputs that are integer multiples of the same source): do NOT group them — they must be analyzed together. Only add set_clock_uncertainty(<jitter_ns>, clock="<clk>") to model PLL jitter/ board skew if you have a real number; otherwise leave the tool's default.

5. Budget the I/O timing (input/output delays)

For every top-level data port that crosses the chip boundary and isn't a clock, add a delay so timing knows the external budget:

  • Inputs: set_input_delay(port_name="<port>", delay_ns=<board_delay>, clock_name="<launch clk>").
  • Outputs: set_output_delay(port_name="<port>", delay_ns=<board_delay>, clock_name="<capture clk>").
  • Numbers come from the interface spec / board: trace delay + external device tco/tsu (system-synchronous), or the source-synchronous skew window. Do not fabricate these. If the user can't supply them, either (a) state an explicit placeholder assumption and flag it loudly in the Output, or (b) ask. A guessed-zero delay is a silent lie that passes in the lab and fails in the field.
  • Pure asynchronous control inputs with no real launch edge (some resets, push-buttons) are NOT given a meaningful I/O delay — they're handled as false paths in Step 6.
  • Pin location / IOSTANDARD (separate from timing) via create_io_constraint(port_name, pin, io_standard) if the board pinout isn't already in a pin XDC.

6. Add genuine exceptions — narrowly, with a stated reason

Exceptions tell the tool to relax analysis. Each one is a claim about the design's behavior. Every exception requires a one-line justification, and if the path could be real, ask the user before adding it. Prefer the narrowest tool.

  • Async reset distribution (reset asserted async, deasserted synchronously through a reset synchronizer): set_false_path(from_signal="<rst>", from_type="port"). Justification: "reset is recovered through a synchronizer; the async assert edge has no timing requirement." If there is NO reset synchronizer, this is a bug — do not false_path it; flag it.
  • Static / quasi-static config registers read by a faster clock but written rarely: set_multicycle_path(N, from_signal="<src_reg>[*]/C", to_signal="<dst_reg>[*]/D") with a justification of why N cycles are guaranteed by the protocol. Hold MCP is auto-adjusted; verify it.
  • CDC datapath that's already synchronized but you want a bounded skew instead of fully cutting it: set_max_delay(<period_ns>, from_node, to_node, datapath_only=True) (and/or set_bus_skew for a multi-bit bus through a handshake/gray-code). This is safer than set_clock_groups because it still bounds the path.
  • Single mux-exclusive clock structures: prefer set_clock_groups relationship="logically_exclusive" over a pile of false paths.

Never add an exception merely because a path is failing setup/hold. A failing real path is a timing-closure problem, not a constraints problem.

7. VERIFY against a synthesized netlist (mandatory)

Constraints written against RTL are unproven until checked against real clock networks and pins. Do this every time:

  • save_constraints(file_name="timing.xdc") then ensure it's in the project (add_constraint_file / list_constraint_files). Or author the file directly with create_constraint_file(file_name, content) for full control of ordering.
  • run_synthesis_async → poll get_run_status until done → open_synthesized_design.
  • check_timing — the unconstrained-objects count must drop to the intended set (zero, except objects you deliberately left as no-timing async). Any remaining "no clock"/"no input delay" is an unfinished constraint, not a pass.
  • get_all_clocks + report_clock_networks — confirm every clock (primary and generated) appears with the correct period and source. A generated clock at the wrong frequency = wrong M/D.
  • report_clock_interaction — the clock-pair matrix. Confirm: related clocks are Timed, intentionally-async pairs are Ignored/partial because you grouped them on purpose, and no pair you expected to be timed is silently ignored. This is the honesty check on Step 4.
  • report_timing_summary — sanity only. At this stage you care that paths are being analyzed correctly, not that WNS is positive. Do not declare success on WNS here; that's closure/signoff.

8. Finalize

  • save_constraints the verified file; confirm it's the one in the project.
  • Produce the Output report (below): every clock, every delay, every exception with its justification, and every assumption you had to make.
  • If real timing analysis is now needed → hand off to timing-closure. If any async grouping was added → hand off to cdc-analysis to verify the synchronizers. State the handoff explicitly.

Classification: symptom → likely cause → smallest correct constraint

Use this to decide which constraint, not just that a constraint is needed.

Symptom (from check_timing / clock reports) Likely cause Smallest correct fix Honesty gate
"Clock <port> has no create_clock" Primary clock undefined create_clock_constraint with real period Period must be the real board freq, not a wish
PLL/MMCM output clock missing or at source period Generated clock undefined create_generated_clock(multiply_by/divide_by from IP config) M/D must match the IP's actual config
N input/output ports "no input/output delay" I/O unbudgeted set_input_delay/set_output_delay from interface spec Numbers from datasheet/board — never fabricated
Inter-clock paths failing between unrelated clocks Async domains analyzed as if related set_clock_groups(... asynchronous) ONLY if a real CDC synchronizer exists — else CDC bug, do not group
Mux-selected clocks both timed against each other Tool can't infer mutual exclusivity set_clock_groups(... logically_exclusive) The mux must truly make them mutually exclusive
Path from slow config reg to fast logic fails setup Real multicycle relationship set_multicycle_path(N, ...) N must be guaranteed by the protocol, justified
Async reset/async input path fails Async assert with sync recovery set_false_path(from_signal=rst, from_type=port) A reset synchronizer must exist; else flag bug
Generated clock at wrong frequency after synth Wrong M/D or wrong target pin Fix create_generated_clock args Re-verify with report_clock_networks — do not waive
Combinational loop reported RTL bug, not a constraint Do not constrain around it Report to user — this is an RTL fix

Iterate: author → synthesize → verify → refine (LOOP)

Constraints are correct only when proven on a netlist. Loop one constraint class at a time:

  1. MEASURE — check_timing + report_clock_interaction + report_clock_networks. Record: unconstrained-object count, which clock pairs are Timed/Ignored, generated-clock periods.
  2. CLASSIFY — use the table to pick the single smallest correct constraint for the top remaining gap.
  3. APPLY ONE CLASS — add only that class of constraint this iteration (all primary clocks, OR all generated clocks, OR I/O delays, OR one exception family). Save.
  4. RE-SYNTHESIZE & RE-MEASURE — run_synthesis_async → get_run_status → open_synthesized_design → repeat Step 1's measurements.
  5. SHOW YOUR WORK — before/after for this iteration: unconstrained count went X→Y, clock pair Z went Ignored→Timed (or the reverse, on purpose).

STOP conditions (stop looping and report)

  • DONE: check_timing shows zero unintended unconstrained objects; every clock present at correct period/source in report_clock_networks; every clock pair in report_clock_interaction is Timed-or-intentionally-Ignored with a stated reason. → Finalize (Step 8). Note: this is constraints complete, not timing met.
  • NEEDS HUMAN INPUT: a clock period, an I/O delay number, or an async/CDC relationship is unknown and cannot be derived. → Present what's known, list the exact unknowns, stop and ask. Do not guess.
  • NOT A CONSTRAINTS PROBLEM: the gap is an RTL/architecture issue (combinational loop, missing synchronizer, a genuinely failing real path). → Present the finding, recommend the RTL/CDC fix, hand off — do not add an exception to hide it.
  • THRASHING: two consecutive iterations don't reduce unintended-unconstrained count or you're toggling a constraint back and forth. → Stop, surface the trade-off (e.g. "this crossing is either a real CDC bug or needs a synchronizer before I can group it"), hand back to the human.

Safety rails (do not violate)

  • No fake-pass / no silencing. Never add set_false_path, set_clock_groups(asynchronous), set_multicycle_path, or waive_lint_violation to make a real path's violation disappear. An exception is a factual claim about the hardware; if it could be false, ask the user first and record the assumption. WNS going green because you cut a path is not closure — it's a hidden bug.
  • Async grouping ≠ CDC safety. set_clock_groups(asynchronous) removes the cross-domain path from analysis; it does NOT add a synchronizer. Only group domains that already have proper CDC structures. When you add any async group, you MUST flag it for cdc-analysis in the Output.
  • Real numbers only. Clock periods come from the actual oscillator/IP; I/O delays come from the interface spec/board. Never fabricate a period or a delay to make a report look good. If unknown → ask or flag a stated assumption.
  • Smallest safe change, in order. Constraints/strategy before RTL. Define clocks before exceptions. Prefer the narrowest tool: bound a CDC path with set_max_delay(datapath_only=True) rather than wholesale set_clock_groups when you only need a skew bound.
  • Primary vs generated discipline. Never create_clock_constraint on a PLL/ MMCM output or a derived clock — that double-defines and corrupts analysis. Those are always create_generated_clock.
  • Don't redefine IP-owned clocks. If a Clocking Wizard / MIG / transceiver IP already constrains a clock (check report_clock_networks / its XDC), do not add a conflicting definition.
  • Recommend RTL, don't silently rewrite it. If the correct fix is a reset synchronizer, a CDC FIFO, or a logic change, report it and let the human decide. This skill edits constraints, not source.
  • Evidence before claims. "Constraints complete" REQUIRES fresh post-synthesis check_timing + report_clock_networks + report_clock_interaction output in this session — never inferred from the RTL or from a prior run. Any signoff-level timing claim (timing met) requires post-implementation evidence and belongs to the timing-closure/signoff flow, not here.
  • One change-class per iteration so cause→effect in the timing report is attributable.

Output

Return a concise, evidence-backed report:

  1. Scope — created from scratch vs. completed an existing XDC; target part.
  2. Clocks defined — table of every clock: name, type (primary/generated), source, period (ns) / frequency (MHz), M/D or phase for generated, verified period from report_clock_networks (proves it, doesn't just assert it).
  3. Clock relationships — every set_clock_groups / uncertainty, with the relationship and a one-line reason; the relevant rows of report_clock_interaction showing the matrix is as intended.
  4. I/O delays — table of port → delay → reference clock, and the source of each number (datasheet / board / assumed (flagged)).
  5. Exceptions — every false_path / multicycle / max_delay with its one-line justification and whether the user confirmed it.
  6. Verification evidence — fresh check_timing unconstrained-object count (before → after, ideally → 0 unintended), confirmation all clocks appear correctly, and that the clock-pair matrix matches intent. Quote the numbers.
  7. Assumptions & open questions — every value you had to assume or guess, called out explicitly for the user to confirm.
  8. Handoff — explicit next step: timing-closure (to actually meet timing) and/ or cdc-analysis (mandatory if any async clock group was added) — naming the skill, not silently ending.

Honesty footer to include verbatim when relevant: "Constraints are complete and verified against the synthesized netlist. This means timing is now correctly DEFINED — it does not mean timing is MET. Any async clock groups I added assume a working CDC synchronizer exists; verify with cdc-analysis before signoff."