Replace pixel rounding with floor(x + 0.5)

[markdibarry]

Fixes: #93731

We should do some more testing to be sure this has no side effects in pixel perfect art projects.

[alvinhochun]

Thanks for making the PR. I was also looking at this more in-depth locally, and noticed

godot/scene/gui/rich_text_label.cpp

Line 1252 in 25de53e

fx_offset = fx_offset.round();

might also need the same rounding change, since it checks snap_2d_transforms_to_pixel.

I do also have to note that, I realized there is a very specific edge case: the value 0.49999997 will be rounded incorrectly because 0.49999997f + 0.5f == 1.0f. But I think it is acceptable to ignore this edge case, because it is very rare to end up with this value, compared to multiples of -0.5. (For double the problematic value is 0.49999999999999994.)

Finally, there is the question whether we should do the same for GUI snap_controls_to_pixels, but that should be a separate PR.

[markdibarry]

Ah, yeah. I think that has always been rounded, which is why we never changed it.

[adamscott]

Did you try if it works with @KeyboardDanni's test scene found in #84380?

[markdibarry]

@adamscott Yep. Same result.

[adamscott]

Testing some games, all seems right. Tested out the issue MRP and it fixes it as well. Let's go, then?

[akien-mga]

Thanks!

[Gnumaru]

Instead of replacing round(x) with floor(x + 0.5), wouldn't it be better to replace round(x) with round(x + FLOAT_EPSILON)? (being FLOAT_EPSILON std::numeric_limits<float>::epsilon() or any equivalent constant), or any really small value like 1E-6. This way rounding -0.1 and +0.1 would both result 0 and the code would be more explict about the intention of rounding the value.

(I don't know it if is allowed to comment closed pull requests or issues. If it is not allowed, please forgive me.)

[KeyboardDanni]

I'm not quite sure what epsilon would do for us here. Epsilon is generally used when comparing equality between two floating point numbers, not flooring or rounding. Even there, epsilon is a bit of a flawed concept, because the "minimum distance between two numbers" is different based on the current exponent.

[Chaosus]

Seems like it's cause a serious regression: #94298