Godot dog walk poor performance compared to fex and native on rk3588

[JSaurusRex]

Not an issue perse, but since godot is opensource, this demo is free and you can easily compare to native on arm64 might be a nice case to optimize.

I am running an orange pi 5 with an rx 470 over pcie gen 2 x1, thought my weird setup was causing the lower performance but box64 specifically is significantly lower than both FEX and native.

game / demo: https://godotengine.org/showcase/dogwalk/

you can download the export templates from here: https://godotengine.org/download/archive/4.4.1-stable/
this gives you a zip file, you can take the arm64 linux build and run the pck file with the argument --main-pack dogwalk.pck

here is what i got:
box64 - 5 ~ 20 fps
fex - 12 ~ 37 fps
native - 26 ~ 46 fps

using the latest git of box64 i could get at 25 feb 2026: fafdb6e
previous versions i tried a month prior had the same performance.

Godot also doesn't seem to use a JIT for its gdscript. This is also the biggest difference I've seen between FEX and box64. Ran with profile fastest, default makes the fps even worse.

[ptitSeb]

I can reproduce the slowyness on my side. Well, at least I can see something is wrong. Tryting some of the classic "speedup" switch like BOX64_DYNAREC_CALLRET=1 or BOX64_DYNAREC_DIRTY=1 leads to graphic artifacts (blinking in main menu) and crash at exit. There seems to be heavy use of JIT (I guess for the Godot Scripting), I need to analyze what is happening and why it would not accept optimisation switches.

[ptitSeb]

You are right, after some more analysis, JIT doesn't seems to be used. I need to understand why using dirty=1 brings artifact in that case!

[ptitSeb]

I have identified some regression. If you use Box64 v0.3.8, the speed should be more in par with what is expected. I'll now work on fixing this speed regression.

[ptitSeb]

Note that with 0c76a09 I have fixed the issue with BOX64_DYNAREC_CALLRET=1, but I'm still looking for the speed regression.

[ptitSeb]

So, I have identified some issue. If you can build the latest version of Box64 and try again, see how the perfs have evolved?

Also, note that you can use BOX64_DYNAREC_CALLRET=1 to launch the game, that might further improved performances, or even BOX64_PROFILE=fastest

[JSaurusRex]

I ran my test scenario with BOX64_DYNAREC_CALLRET=1 or BOX64_PROFILE=fastest
2ef63f2 box64 - max 29 min 5
fafdb6e box64 - max 15 min 5
0.3.8 box64 - max 23 min 5
0.3.6 box64 - does not run
fex - max 46 min 11
native - max 47 min 24

Fps certainly improved
I also tested in pseudoregalia with it, that game also runs way worse compared to fex (30 vs 70) and no difference sadly.
Would more performance reports/issues help? I have a bunch more, mainly pseudoregalia (UE4), sword with sauce (UE3), Teardown

[ptitSeb]

Sure, performance report can help. I have some other things to analyse as I have noticed another speed regression in dishonnered between v0.4.0 and current main.
Issue is: I don't have all the games, so for some games, I cannot test on my side (pseudoragalia & sword with sauce for example)

But it seems I'm not done with this one already.

[ptitSeb]

Just to be clear: if in your testing some games are slower, and some are faster, then yes, more example are useful. but if you find all your games slower, then I don't need more examples (as I can use any game anyway).