10 hardening on harder inputs¶
Before the tool meets real boards, the three places where real boards differ most from what had been tested.
Hierarchical schematic (dogfood/existing/hier/, KiCad's complex_hierarchy demo)¶
One sub-sheet used twice, so each of its symbols is two parts.
- parts at first listed none of them: it read the "Reference" property,
which holds one of the two references. References are now read from the
symbol's instance list.
- set R203 Value=33K says it also changes R303; KiCad's netlist shows both
changed. remove R213 refuses, because it would remove R313 too.
- set D101 and add C199 (copy of C104, VCC to GND) on the root sheet.
netdiff between KiCad's netlists: C199 added on VCC and GND; D101, R203 and
R303 changed. Nothing else.
Router on a 6-layer board (KiCad's cm5_minima demo, read only)¶
Six two-pad nets that change layer, re-routed over F.Cu, In2.Cu, In3.Cu and B.Cu with each net's own class rules:
| Net | Result | Original |
|---|---|---|
| nLED_ACTIVITY | 8.5 mm, 1 via | 10.2 mm |
| VBUS_EN | 22.5 mm, 1 via | 43.1 mm |
| nLED_SPEED_100_1000 | failed after 84 s | |
| ETH_nLED_ACTIVITY | 30.5 mm, 2 vias, uses In2 | 31.3 mm |
| Net-(D702-K) | 4.0 mm, 2 vias, uses In2 | 5.5 mm |
| PWM | 32.3 mm, 2 vias, uses In2 | 33.2 mm |
Every routed result keeps at least its class clearance (0.13 mm) from other copper by this tool's own geometry. Not checked by KiCad's DRC: the results were not written to a board. A failure now gives up in a bounded search instead of 84 seconds.
Per-net rules¶
The router read only the Default net class. It now takes each net's class
from the project file, by explicit assignment (the Jetson demo) or by pattern
(cm5_minima: /*USB_* -> 90ohm, 0.147 mm, 0.2 mm clearance). Rules that live
in a .kicad_dru file are not read.
Also¶
A 9-layer, 22,151-track demo board loads and reports its net classes.
Left open¶
- The 6-layer routes were not judged by KiCad.
.kicad_drurules are invisible to the router and the review.addon a sub-sheet; multi-unit parts.