Emulator

4.0-548 -> https://github.com/dolphin-emu/dolphin/commit/42217fbd755d88b461db9f9eecfb568e50167d54

However I doubt this revision broke the game. Could you try using versions inbetween to find out what revision really broke it?

Ok,I download all build of dolphin and it work correctly up to version 4.0 618.
After with version 4.0-619(in the description "VertexLoader: don't check for possible range") and all other there is this issue

Sounds like the magic -1 attribute index. But I doubt the old implementation is better. It's more likely that the complete vertex will be skipped.

Galop1n's builds maybe fix this?

Yeah, it did. The old code just hide this issue and was only working because it's the last vertex in mario tennis. Honestly, vertex skipping is quite hard with out current vertex loader design :/

As a small summary of what's going on: The issue in this game seems to be that the index passed for the position vertex attribute is -1, which triggers special hardware behavior for skipping the current vertex and fetching the next one instead for the current primitive.

We currently ignore this feature and (I guess) grab invalid data instead of skipping the vertex.

I don't know what's so hard to implement about it in the current vertex loader design, but otoh I also haven't looked into the issue in much detail.

By the way, chances are I got the expected behavior more or less wrong. In any case a proper fix should also provide a hwtest to show that the implementation indeed reproduces hardware behavior.

I think this is easy enough to do that we can keep this milestone current. Is anyone willing to tackle it? IF not, I'll move it to next.

but it's easy to hack in a fix:
if(index == -1) output = NaN;

Can we do some kind of simple hardware test to at least verify this?

A revert of this regression is on https://github.com/dolphin-emu/dolphin/pull/1269

fixed in 4.0-4796