skill.md

How-to guide for this repository.

Quick Start

source setup_cadence.sh
./build.sh all

Run Spectre Simulation

cd results/<name>
spectre /path/to/netlists/<name>.scs -raw <name>.raw +log spectre.log

Check: grep "completes with 0 errors" spectre.log

Run OCEAN Verification

ocean -nograph < ocean/test_<name>.ocn
cat results/<name>_test.txt

Run SKILL Replay (Virtuoso Headless-Friendly)

./scripts/virtuoso_replay.sh skill/L5_inverter.il

Create a New Gate

1. Write the netlist

netlists/and2.scs:

simulator lang=spectre
parameters vdd_val=1.8

model nch mos1 type=n vth=0.4 kp=120u
model pch mos1 type=p vth=-0.4 kp=40u

V_VDD (vdd 0) vsource dc=vdd_val
V_A (a 0) vsource type=pulse val0=0 val1=1.8 delay=10n rise=100p fall=100p width=10n period=20n
V_B (b 0) vsource type=pulse val0=0 val1=1.8 delay=20n rise=100p fall=100p width=20n period=40n

// AND2 = NAND2 + Inverter
// NAND stage
MP0 (nand_out a vdd vdd) pch w=2u l=1u
MP1 (nand_out b vdd vdd) pch w=2u l=1u
MN0 (nand_out a mid 0) nch w=2u l=1u
MN1 (mid b 0 0) nch w=2u l=1u

// Inverter stage
MP2 (out nand_out vdd vdd) pch w=2u l=1u
MN2 (out nand_out 0 0) nch w=1u l=1u

tran_test tran stop=80n
save out a b

2. Write the test

ocean/test_and2.ocn:

out = outfile("/home/v71349/analog-gradients/results/and2_test.txt" "w")
fprintf(out "=== AND2 Verification ===\n")

simulator('spectre)
openResults("/home/v71349/analog-gradients/results/and2/and2.raw")
selectResult("tran_test-tran")

vout = v("out")
vth_low = 0.36
vth_high = 1.44
pass = t

; A=0,B=0 -> 0
v = value(vout 5n)
fprintf(out "A=0,B=0: %.3f " v)
if(v < vth_low then fprintf(out "[PASS]\n") else fprintf(out "[FAIL]\n") pass=nil)

; A=1,B=0 -> 0
v = value(vout 15n)
fprintf(out "A=1,B=0: %.3f " v)
if(v < vth_low then fprintf(out "[PASS]\n") else fprintf(out "[FAIL]\n") pass=nil)

; A=0,B=1 -> 0
v = value(vout 25n)
fprintf(out "A=0,B=1: %.3f " v)
if(v < vth_low then fprintf(out "[PASS]\n") else fprintf(out "[FAIL]\n") pass=nil)

; A=1,B=1 -> 1
v = value(vout 35n)
fprintf(out "A=1,B=1: %.3f " v)
if(v > vth_high then fprintf(out "[PASS]\n") else fprintf(out "[FAIL]\n") pass=nil)

if(pass then fprintf(out "\n=== PASS ===\n") else fprintf(out "\n=== FAIL ===\n"))
close(out)
exit()

3. Run it

mkdir -p results/and2
cd results/and2
spectre ../../netlists/and2.scs -raw and2.raw +log spectre.log
cd ../..
ocean -nograph < ocean/test_and2.ocn
cat results/and2_test.txt

Compose Circuits

Instantiate subcircuits in netlists:

// Include subcircuit definition or define inline
subckt inverter (in out vdd)
  MP0 (out in vdd vdd) pch w=2u l=1u
  MN0 (out in 0 0) nch w=1u l=1u
ends inverter

// Instantiate
I0 (a a_inv vdd) inverter
I1 (b b_inv vdd) inverter

Debug Tips

• Check spectre.log for simulation errors
• PSF results named tran_test-tran (with hyphen)
• OCEAN needs: selectResult("tran_test-tran") (quoted string)
• Voltage thresholds: LOW < 0.36V, HIGH > 1.44V (for VDD=1.8V)

CMC Cloud Access

ssh -Y -p 31487 [email protected]
source /CMC/scripts/cadence.ic23.10.140.csh  # if using tcsh
# or
source setup_cadence.sh  # if using bash

Field Notes (OCEAN / Spectre / SKILL)

Short, forum-ready answers distilled from real runs.

Licensing (LMC-01902)

If Spectre says the license search path is <none>, set a license variable:

export [email protected]   # CMC host default observed
# or
export [email protected]

Then re-run:

source setup_cadence.sh
./build.sh inverter

OCEAN in batch mode

• Always write output to a file (stdout is buffered).
• selectResult("tran_test-tran") must be a string because of the hyphen.
• Avoid fancy SKILL constructs in batch (e.g., lambdas). Explicit checks are more reliable.

Spectre netlist gotchas

• Keep subckt pin lists on one line. Multi-line pin lists caused parse errors.
• Use simple mos1 models (per repo templates) to avoid PDK dependencies.
• Use save statements to ensure OCEAN can access signals.

Virtuoso automation

• Headless replay helper: ./scripts/virtuoso_replay.sh
• If DISPLAY is unset, the runner will try xvfb-run if available.
• SKILL-based schematic creation still needs X11 on CMC; netlist flow is the stable path.

Full-flow demo operations (DC -> Innovus -> Calibre)

• Use scripts/run_fullflow_smoke.sh for the standard replay path.
• If DC is license-blocked (DCSH-1), the flow can continue using: implementation/fullflow_demo/rtl/alu4_flow_demo_fallback_mapped.v.
• Calibre license failures are recorded explicitly in: implementation/fullflow_demo/work/calibre/alu4_flow_demo_calibre_license.warn.
• For CI or strict gating, use: env FULLFLOW_STRICT=1 scripts/run_fullflow_smoke.sh (returns non-zero when any stage is blocked).

Demo asset generation

• One-command visual refresh: scripts/run_competition_visuals.sh
• Coupled-tile waveform assets are generated under:
  • competition/plots/neuro_tile4_coupled_spikes.svg
  • competition/plots/neuro_tile4_coupled_mems.svg
• Before recording, pre-run long flows and verify artifact existence so the live capture can focus on the key moments (PASS cascade, waveform evidence, GDS output).

Recording workflow

• Build scene-organized assets with: scripts/build_recording_pack.sh
• Guided pacing for one-take capture: scripts/demo_narrator.sh
• Fast rehearsal without long tool runtime: DEMO_SKIP_LONG=1 scripts/demo_narrator.sh

LaTeX paper workflow

• Prepare parsed paper datasets: python3 scripts/prepare_paper_data.py
• Build paper (if LaTeX engine is installed): scripts/build_paper.sh
• Main source: competition/paper/neurocore_workthrough.tex
• Plot style choice for evidence: dark background + unsmoothed raw data points.

Founder-thesis tracking

• Canonical thesis reference: competition/founder-thesis.md
• Keep claims tied to reproducible evidence:
  • ODE fit
  • timing sensitivity (dt_spike/dparameter)
  • energy per event
• Current analysis commands:
  • python3 scripts/analyze_lif_ode_fit.py
  • python3 scripts/analyze_temporal_sensitivity.py
  • scripts/analyze_lif_energy.sh
• Mixed-signal smoke command:
  • ./build.sh neuro_tile4_mixed_signal