TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

Reading Real Reports, Part 1: The Five Questions Seniors Ask Before Reading Any Report

The Semiconductor Hub
by
0
Share Facebook X LinkedIn WhatsApp Email
Terminal-style excerpt of an OpenSTA session showing setup warnings, parasitic annotation counts and worst slack summaries.

Most engineers learn to read a timing or synthesis report by being handed one and told to fix the worst path. Nobody explains that a senior engineer usually does not start with the worst path at all. The first few minutes go to a different question: can this report be trusted? A clean report from the wrong run, the wrong corner or the wrong constraints is worse than a failing one, because it ends the conversation early.

This is Part 1 of Reading Real Reports, a daily series on how experienced engineers read tool output. Each part takes one report, shows an illustrative sample, and walks through it line by line. We start with the five questions that come before any report, whatever the tool.

Key takeaways

  • Before you read a single number, establish what ran, on what inputs, with which constraints, what changed since the last run, and what was waived.
  • Most bad sign-offs come from a report that was correct for the wrong setup, not from a misread number.
  • The tool will often tell you about the setup problems itself, in warnings and sanity checks that people scroll past.
  • Waivers and timing exceptions are part of the result. A report with no list of what was excluded is half a report.
  • Write the answers down in a short run card before the review.

In this post

What a report really is

A report is the answer to a question you did not fully write down. The tool read some files, applied some settings, assumed where information was missing, and printed a result. The result is only as good as those inputs and assumptions.

You see this in every open flow. The OpenROAD-flow-scripts tutorial shows a flow that writes one log per stage, with names such as 1_1_yosys.log, 4_1_cts.log and 6_report.log, and a separate reports folder (see Viewing Results and Logs). The report you are handed is one file out of dozens. Never read it without its log, and never read the log without the five questions below.

The five questions

1. What ran?

Which tool, which version and which script. Versions matter because commands and defaults change. The UC Santa Cruz multi-corner tutorial selects a corner with a -corner option (Reporting per corner), while the current OpenSTA command reference lists -scenes for the same job (report_checks -scenes). A script written for one release can silently mean something else on another.

Also check for setup that ran before the script. The OpenROAD documentation notes that the tool sources the file ~/.openroad at startup unless -no_init is given (Run). A personal init file can change settings without appearing in the flow script at all.

2. On what?

Which netlist, which libraries and which parasitics. A static timing run needs four kinds of input: libraries, the design, parasitics and constraints, as the UC Santa Cruz STA tutorial sets out in its single corner example. Each one can be stale. The library name tells you the corner: the same tutorial pairs a slow, hot, low-voltage library with setup checks and a fast, cold, high-voltage one with hold checks (Why multiple corners?).

Parasitics deserve their own look. The OpenROAD-flow-scripts note on endpoint slack tracking explains that a design is timed with no parasitics before placement, with placement estimates after placement, with global routing estimates later, and with extracted SPEF only at the final step (Recording). The same path can show very different slack depending on which of these was loaded.

3. With which constraints?

Which SDC file, and is it complete? A missing clock or input delay does not make a path fail. It makes the path disappear from the checks. OpenSTA has a command for exactly this, check_setup, which looks for inputs without set_input_delay, outputs without set_output_delay, unclocked register pins, endpoints with no timing check, combinational loops and generated clocks without a defined source. Units belong here too: report_units prints the time, capacitance and resistance units every number in the report is using.

4. What changed since the last run?

A result means more next to the last one. If worst slack moved from minus 50 ps to plus 10 ps, a senior wants to know why before celebrating. Was it a real fix, a relaxed constraint, a different library or a different parasitic model? The endpoint slack tracking note above makes the same point: slack changes between steps that use different parasitic models include the change of model itself (Recording), and it offers a compare mode that lists endpoints that became violating or stopped violating (Analysis).

5. What was waived or excluded?

Every flow has ways to make a check go quiet. In timing, set_false_path removes paths from analysis, and set_case_analysis ties pins to constants so no paths propagate from them. In messages, suppress_msg hides warnings by ID. In lint, Verilator can write every waiver for a run into a file with --waiver-output. None of these are wrong. All of them belong in the review, with a reason next to each one.

Common traps

  • Green summary, red log. The summary line passes, but the log above it warns about missing constraints or unread files.
  • One corner treated as all corners. Setup was checked at the slow corner and hold was also checked at the slow corner, because only one library was loaded.
  • Estimated parasitics read as final. A post-placement estimate is useful for trends, not for sign-off.
  • Units assumed. A library characterized in picoseconds and a script written in nanoseconds give numbers that look plausible and mean nothing.
  • Exceptions with no owner. An old false path still sits in the SDC, and nobody remembers why.

Deep dive: the numbers and details

This part holds the sample session, the line-by-line reading, the arithmetic and the checklists. Skip it if the overview is all you need.

The illustrative sample

The block below is an illustrative sample in the style of an OpenROAD/OpenSTA session log, written for this post. The build ID, file names, port names and values are invented. The command names come from the OpenSTA command reference, the report_units layout follows the example on that page, and the warning wording follows the OpenSTA source files CheckTiming.cc and ReportParasiticAnnotation.cc. Exact output varies by version.

OpenROAD 3c41d0a9e5b7
Features included (+) or not (-): +GUI +Python
openroad> read_liberty lib/stdcell_ss_0p72v_125c.lib
openroad> read_verilog results/top_6_final.v
openroad> link_design top
openroad> read_sdc results/top_6_final.sdc
openroad> read_spef results/top_6_final.spef
openroad> report_units
 time 1ns
 capacitance 1pF
 resistance 1kohm
 voltage 1v
 current 1A
 power 1pW
 distance 1um
openroad> check_setup -verbose
Warning: There are 2 input ports missing set_input_delay.
 irq_ext
 test_mode
Warning: There is 1 unconstrained endpoint.
 dbg_out
openroad> report_parasitic_annotation
Found 0 unannotated drivers.
Found 14 partially unannotated drivers.
openroad> report_worst_slack -max
worst slack 0.04
openroad> report_tns -max
tns 0.00
openroad> report_worst_slack -min
worst slack -0.02
openroad> report_tns -min
tns -0.11

Line-by-line walk-through

  1. OpenROAD 3c41d0a9e5b7. The build identifier. The UC Santa Cruz tutorial shows the same kind of banner when the shell starts (Tools). Record it in the run card. If two runs disagree, this is the first thing to compare.
  2. read_liberty lib/stdcell_ss_0p72v_125c.lib. Only one library is read, and its name says slow process, low voltage, hot. That is a setup corner. Nothing in this session describes a fast corner. Hold the thought for line 10.
  3. read_verilog, link_design top. The final-stage netlist. Confirm the file date matches the run you are reviewing.
  4. read_sdc results/top_6_final.sdc. The constraints came from the flow's results folder, not from the source tree. Ask whether the two still match. The OpenSTA reference also notes that read_sdc stops on errors unless a continue-on-error variable is set, so check which way this flow runs.
  5. read_spef. Extracted parasitics, which is what sign-off needs. Line 9 checks how complete they are.
  6. report_units. Time is in nanoseconds. Every slack below is in ns. A worst slack of 0.04 is 40 ps.
  7. First check_setup warning. Two inputs have no input delay, so paths from them are not timed against a clock. irq_ext sounds like an asynchronous interrupt that should be synchronized and excluded on purpose. test_mode may be meant to be a constant. Either way the answer belongs in the SDC, with a comment, not in someone's memory. Note that OpenSTA's own check skips inputs that are constant, so a set_case_analysis on test_mode would also clear it (CheckTiming.cc).
  8. Second warning, dbg_out. An output with no check. A debug port may be fine to leave untimed, but someone has to say so.
  9. report_parasitic_annotation. The command reports how complete the SPEF annotation is (report_parasitic_annotation). Zero drivers are fully missing, but 14 are partially missing. Rerun with -report_unannotated to see which nets and loads, and check whether any sit on the worst paths.
  10. Setup summary. report_worst_slack with -max gives 0.04 and TNS is zero. On its own, a pass.
  11. Hold summary. With -min, worst slack is minus 0.02 and TNS is minus 0.11. Hold fails, and it was checked only at the slow library. Hold is normally worst at the fast corner, so the real hold picture is probably worse than this.

Worked arithmetic check

Margin as a share of the clock. Suppose the SDC sets a 1.25 ns clock. Setup margin is 0.04 / 1.25 = 3.2 percent of the period. That is thin for a final run with 14 partly annotated nets still open.

How many hold violators at least? TNS is the sum of negative slack over endpoints (report_tns), and no endpoint is worse than WNS. So the number of failing endpoints is at least TNS / WNS = 0.11 / 0.02 = 5.5, which rounds up to 6. One glance at two summary numbers tells you this is not a single stray path.

Rounding in the last digit. Reports round each printed value. In one public OpenSTA regression output, a path shows 487.39 required and 117.08 arrival, and the slack line reads 370.32, while 487.39 minus 117.08 is 370.31 (report_checks_sorted.ok). A one-count difference in the last digit is rounding, not a bug. If you need to argue over the last digit, rerun with more -digits.

Red flags table

What you seeWhy it mattersWhat to ask
One library read, both setup and hold reportedHold is checked at a corner that hides most hold riskWhere is the fast-corner run?
check_setup warnings left in the logPaths are missing from the checks, not passingIs each port excluded on purpose, and where is that written?
Partially unannotated driversSome delays use estimates or no wire loadAre any on the critical or hold-critical paths?
Large slack jump since last runCould be a fix, or a changed inputWhich file changed: netlist, SDC, library or parasitics?
False paths or suppressed messages with no commentNobody can judge if they are still validWho added it, and why? OpenSTA's -comment option exists for this
No version or build ID in the logThe run cannot be reproducedWhich tool build produced this?

Sign-off checklist

  • Tool build ID and script path recorded; init files known.
  • Netlist, SDC, library and SPEF paths recorded, with dates.
  • Units printed and noted.
  • check_setup -verbose clean, or every warning explained in the SDC.
  • Parasitic annotation complete, or gaps listed and judged.
  • Setup and hold each reported at the corners meant for them.
  • Result compared with the previous run, and every big change explained.
  • All exceptions, case settings, suppressed messages and lint waivers listed, each with an owner and a reason.

Further reading

Next in the series: The STA Timing Path Report, Line by Line.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!