Skip to content

new video - #5

Merged
nanocoh merged 3 commits into
mainfrom
dev
Nov 17, 2026
Merged

nanocoh merged 3 commits into
mainfrom
dev

Conversation

@nanocoh

@nanocoh nanocoh commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

No description provided.

The GCD reference edit named each new net like the gate driving it. NajaEDA
keeps instance and net names apart and exports such a pair, and its own
reader, Kepler Formal and OpenROAD accept the file. It is not valid Verilog,
though: Verilator and Yosys reject it, and OpenROAD's SPEF read-back could
not attach the extracted parasitics of those 26 nets (253 "pin not found"
warnings). The candidate was timed without those wires and reported at
+0.045 ns; the same layout has -0.050 ns.

- reference/edit.py names new nets ppa_cla_<term>_net and checks every new
  name against instances, nets and ports before editing.
- tools/edit_validation.py finds instances named like a net or port, as a
  function and as a command for exported files. The live session refuses an
  edit that adds such a pair (set_name is allowed for the repair), and the
  reference regression runs the check after the edit.
- The regression fails a physical run whose log reports parasitics that
  were not annotated.
- The NajaEDA and OpenROAD skills state the rule and the symptom. The
  historical result is marked as unverified: it used the same names.

With legal names the 31-gate network ends at -0.025 ns in the packaged flow.
The packaged flow repairs setup timing on placement estimates and stops at
about zero slack; routing then costs several tens of picoseconds, so the
31-gate network ended at -0.025 ns.

The five majority gates are now replaced by 23 gates: the carry-in is folded
into bit 0, the generate/propagate terms form a balanced tree of four
levels, and the cells alternate polarity so that each level is a single
stage. The flow needs no setup repair and ends at +0.205 ns (+0.185 to
+0.205 ns with four other naming schemes), with less area and power.
Kepler Formal SEC proves all 18 outputs.
The video and its preview now show the packaged flow with the current
reference edit, from the baseline at -0.544 ns to +0.205 ns. Every number,
log line and layout image in it comes from that run.

The inline player in README.md still points to the earlier upload: a GitHub
attachment can only be replaced through the web interface.
@nanocoh
nanocoh merged commit c6314d9 into main Nov 17, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant