05:57 <dnkl> ahesford: there's one more thing we should test. That's dragging a floating windows between monitors with either different DPIs (dpi-aware=yes), or different scale factors (dpi-aware=no). It's important the window doesn't keep shrinking, or growing, when moving it back and forth

05:58 <dnkl> sterni: pretty sure it's impossible to guarantee reproducibility. Especially across different systems.

06:02 <dnkl> ahesford: tested on river

06:02 <dnkl> it's slowly shrinking

06:03 <dnkl> (different DPIs and scaling factors in my case, but only DPI should matter since I use dpi-aware=yes)

06:07 <dnkl> with resize-by-cells=no it works as expected: the grid size is different on the two monitors, but it's correctly "restored" when moving the window back

13:23 <ahesford> dnkl: I was not able to reproduce this with labwc by changing the resolution of one display, even though foot sees a different DPI

13:24 <ahesford> the window changes size as expected, but dragging back and forth repeatedly always keeps the grid the same size and the window has the same size on the original display

13:24 <ahesford> can you see if rounding instead of truncated fixes the problem? https://somebits.link/u/p/5b24f7a993/stream

13:27 <ahesford> another possible fix is to only apply the cell-size truncation if render_resize is given a nonzero width and height and is not given RESIZE_KEEP_GRID, i.e., when the compositor is trying to resize; we can just live with whatever rounding errors accumulate for dpi or scale factors

13:56 <sterni> dnkl: partial PGO is reproducible across multiple machines according to my testing, so it is possible at least in some cases

14:46 <dnkl> sterni: partial pgo should be completely deterministic. Or doesn't start any threads, or depend on I/O.

15:01 <sterni> that's my experience, too!

15:02 <sterni> I guess that is the recipe for deterministic PGO, just add a fake executable that tries to hit as many of the big cost centers without introducing nondeterminism somehow

15:02 <sterni> this is probably not very feasible for e.g. firefox, but I'd hope you can do so for python, gcc, …

16:25 <ahesford> hrm, I can reproduce the size shrink on my work laptop with external displays

16:27 <ahesford> but not with a non-zero pad, hrm

16:43 <ahesford> alright, I understand the problem

16:45 <ahesford> it's not resize-by-cells; I can reproduce it even with that option disabled

16:49 <ahesford> seems like the grid-preservation logic applies on display change before the scale has been updatedj

18:14 <ahesford> fixed, with an explanator comment left in the PR

18:14 <ahesford> explanatory*