Emulator

If i had to guess, i'd say its fpu drifting errors accumulating over time

yeah, I've been getting that issue for the past 1000 or so revisions, I just always forgot to post this issue on google code.

issue is still valid...I've got a feeling the ghosts have gotten a tidbit more intelligent, but they still do crazy shit xD

I just tested this to see if it's still valid, since it's been over two years. It is, and really without any changes. It's probably the same thing as the Brawl Replays.

The guess of accumulating errors may beright, but one thing I noticed is that after a jump the Wii ghosts seem to be basically screwed. That could be because small differences in position before hand causes the jump to go differently, or something else in Dolphin may be wrong in how it handles something. It does seem pretty sudden when things do go wrong.

Dualcore doesn't make a difference, as well as idle skipping (I dunno, I was just trying stuff at that point.) or HLE/LLE Audio. It's not affected by any of those, and the ghosts always fail the same way, so at least it's consistent.

Do You run it 100% speed and 60fps?

Issue 6290 has been merged into this issue.

eh...
Mario Circuit is okay-ish sometimes because it's mostly flat?
SNES Mario Circuit is OK because it's completely flat.
Correct me if I'm wrong.

This might be a rounding issue related to going downhill.
Disclaimer: I know nothing about the Wii architecture other than it having two processors and running on IOSes or something.

Yes, I confirmed that mario circuit is not affected, as far as I can tell, no stupid errors. And the bug exists on r6515.

I recorded a quick demonstration of this glitch.

http://www.youtube.com/watch?v=SQGtSTx-UzQ

This video shows side by side what happens and may help with debugging exactly what's happening. I also have the source file for going frame by frame.

Two things I observed:

Ghosts that stay on flat surfaces work in Mario Kart Wii. In Double Dash, I don't think that's the case.

The landing, even though it seems to happen at the same point, gets less steering input on Dolphin.

I watched the video and saw something interesting: before the big jump at the end (when things really start going downhill), the kart gets a small bump when it goes offroad (check video at minute 0:58).

When landing from that bump, both karts are already in different places and going at different speeds.

Then, when the original Wii kart lands from the big jump, it immediately turns slightly to the left. However, since at that point in time the Dolphin kart has not yet landed, it also doesn't turn and ends up going offroad when it shouldn't.

So perhaps this is a CPU bug which results in wrong game physics? It can also be wrong collision detection...

That's what I assumed, but we really have no leads on what. This even happens in interpreter.

Dolphin's floating point emulation is known to be inaccurate in a number of ways.

The two instructions which I'm really suspicious about are fres and frsqrte.

These instructions provide estimates and are not even guaranteed to provide the same result across Intel and AMD processors, or even different generations of the same family of processors.

Is there a way to make them more accurate or simply different, if only in a branch or some kind of experiment to see if it fixes or effects this issue?

Unfortunately, the exact algorithm used for these instructions on the gecko isn't documented, so someone would have to spend some time with a real wii to RE these instructions, which would take a lot of time.

And we don't even know if it's these instructions that are at fault, they are just the ones which stick out the most when I read through the gecko manual.

This should be fixed in JIT with this build: https://dl.dolphin-emu.org/prs/pr-834-dolphin-latest-x64.7z