Are there caveats to using off-grid vertices in UDMF?

[SuaveSteve]

What it says on the tin.

[Graf Zahl]

In some extreme circumstances it can happen that there's a loss of precision in the node builder, but that normally only happens if too many off-grid vertices are too close to each other.
It's really too bad that there's no 128 bit floating point format, which would solve most of these issues. Note that the issue is solely inside the node builder, not inside the engine itself.

[Xaser]

If you ever hit such issues and you're using GZDB, you can set the grid to a fractional size (e.g. 0.25 units) and use the "snap selection to grid" feature to get things to behave better. This has saved me a few headaches.

[Nevander]

One time this was causing me to get an Unknown ZNODES lump in SLADE. I went around the map by hand moving any off-grid vertex (on the 8 unit grid) onto the grid. Eventually the Nodes Viewer showed a valid format, and then I had my precise ZNODES back.

[Zeberpal]

If you ever hit such issues and you're using GZDB, you can set the grid to a fractional size (e.g. 0.25 units) and use the "snap selection to grid" feature to get things to behave better. This has saved me a few headaches.

I have a task to do a multi-linedefs walls, so if I manage to do it with 1 mp, 0.5 mp and 0.25 mp distance between them, it will be easier for nodebuilder to "build" and fortunately would not provoke any glitch problems in the future?

From my point of noob view, the problems will occur if I use some "unsnaped to grid" distance between linedefs/vertexes like 0.666 or 0.1337

[Xaser]

In an ideal world it wouldn't matter, so you're probably fine proceeding normally and using the snap-to-grid workaround only if you get issues. I've just gotten into the habit of snapping things proactively for paranoia's sake if nothing else.

If you do hit a trouble-spot, it's worth posting the misbehaving map in this bugthread so there's more examples in the off chance someone decides to tackle the root issue.

[Graf Zahl]

The root issue is that even 64 bit floating point math can become imprecise when extremely small distances need to be calculated. Since this is a hardware limitation it is currently not fixable.