LEF pins causing misaligned Horizonal/Verticals and power via generation

[gatecat]

Test case, hopefully complete: sram_minimal.zip

(run run_pnr.py)

There are a few issues going on here, but one of the first I notice is around offgrid parts of the power pins in the LEF.

For example:

    PORT
      LAYER Metal3 ;
        RECT 0 40.76 5.07 47.575 ;
    END

All of the points of the rectangle are on the 0.005um grid, but as the centerline isn't, Coriolis produces this error:

[ERROR] LefParser::fromUnitsMicrons(): Offgrid value 44.1675um (DbU=883350), grid 0.005um (DbU=100).

As this is a power port, it seems like this also leads to offgrid power vias further down the line, although it may be that there's a better way to connect macros to power than just creating metal3 for all the power pin areas and relying on the usual via generation down to it.

[jpc-lip6]

Thanks for making the right analysis. This error was nagging at me for some times. The parser transforms rectangles into segments, so they use the center coordinate. Most likely I will slightly shrink the rectangle, but I need to check if it does not create other problems. The other fix would be to use Hurricane::Pad instead of segments, but this would have a lot of implication in the code (I heavily rely on shapes being segments).

I will take a serious look by the end of the week.

[jpc-lip6]

There was in fact, three different causes for the offgrid segments:

1. In your routing gauge configuration, the Metal5 layer was offgrid : 0.876 ? (replaced it by 0.88).
2. In LefImport, the translation of rectilinear polygons into Horizontal / Vertical was sometime failing because the axis may be offgrid. I replace them by two overlapping on grid.
3. The GDS parser, when reading PATH records, translate them as series of articulated Contact & Horizontal / Vertical, and it suffers the same limitation as LefImport. For now, I force them on grid while writing back with GdsDriver, but still display an error message. This may be a bit dangerous.
4. Unrouted signals like io_out_to_pad(23) may have segments offgrid because the router fails on them and the segments are never placed ongrid. Not really a problem because the layout is wrong anyway...