Add content scale stretch modes, implement integer scaling

[deralmas]

Supersedes (and based on) #63206
Fixes godotengine/godot-proposals#1666

Integer scaling is achieved (after aspect expansion) by "lying" to the stretching code about the window's size, telling it that it's always an integer multiple of the viewport so that it only gets stretched to an integer factor.

This approach works with all stretch and aspect modes and doesn't require handling for each, only requiring to "loosen up" some self-excluding conditions (in other words, replacing some else ifs with just ifs) regarding viewport offset and margin calculation (black bars).

Includes a tiny usability change that adds a range hint for the content scale factor between 0.5 to 8.0.

Includes stubs for an unimplemented "hybrid" mode (see below), which would allow to fill the whole screen while keeping square pixels with minimal blurring. It might have to be implemented later until core supports more fine-grained window filtering controls.

For hybrid mode see: https://gamingprojects.wordpress.com/2017/12/03/reducing-pixel-blur-and-distortion/ Update: removed.

Notes:

• I don't really like the current name of this property nor its path (.../window/stretch/stretch), but I can't really think of a good alternative either if not something like "scaling mode", which I'm not sure of either. Update: went for scale_mode.

• Currently sub-one scales are clamped to one. This means that the window will clip with the aspect ratio calculation still going on, thus making games (especially ones with expanding aspect ratios) pretty unusable if for some bizarre reason the window must stay very small. I'm not sure if I should let the window stretch in these cases. Update: resolved.

• I feel like this code might be commented/cleaned a bit but I haven't done any since I have no idea what to do with hybrid scaling. I feel like I could seriously remove its stub since it, while being a great feature, would be too much of an hassle right now. Update: resolved.

• I have tested it both atop of my Wayland branch on my Wayland/AMD laptop and on my Xorg/NVIDIA desktop. Both seemed to work fine (although the desktop was super slow at resizing but I recall that it was normal, dunno, I haven't used it for serious godot development in ages). That said, more serious and thorough testing is always appreciated.

[deralmas]

CC @Calinou (I know that you don't have time to work on this, but I think that you'd like just seeing this :) )

[Calinou]

• I don't really like the current name of this property nor its path (.../window/stretch/stretch), but I can't really think of a good alternative either if not something like "scaling mode", which I'm not sure of either.

"scaling mode" may be confused with the 3D scaling project settings.

• Currently sub-one scales are clamped to one. This means that the window will clip with the aspect ratio calculation still going on, thus making games (especially ones with expanding aspect ratios) pretty unusable if for some bizarre reason the window must stay very small. I'm not sure if I should let the window stretch in these cases.

I think it's fine this way, as long as it's documented in the class reference. We can recommend setting the minimum window size to the base window size if you want to avoid display issues.

• I feel like this code might be commented/cleaned a bit but I haven't done any since I have no idea what to do with hybrid scaling. I feel like I could seriously remove its stub since it, while being a great feature, would be too much of an hassle right now.

I agree with removing hybrid mode from this PR, but keep the property as an enum so that more options can be added without breaking compatibility.

[deralmas]

@Calinou

"scaling mode" may be confused with the 3D scaling project settings.

Is that an issue with 2D?

I think it's fine this way, as long as it's documented in the class reference. We can recommend setting the minimum window size to the base window size if you want to avoid display issues.

That would probably be the best option. Is there a project setting for setting the minimum window's size? I recall not being able to find it before.

I agree with removing hybrid mode from this PR, but keep the property as an enum so that more options can be added without breaking compatibility.

I think I'll move in this direction then if no one else is against. I'll update the PR later.

[Calinou]

Is that an issue with 2D?

Yes, as many people will attempt to tweak the wrong setting 🙂

That would probably be the best option. Is there a project setting for setting the minimum window's size? I recall not being able to find it before.

Not yet, as this has been attempted and it caused crashes in the past. You have to call DisplayServer.window_set_min_size() in an autoload's _ready() function instead.

It's possible that this setting can be reintroduced just fine with 4.0's DisplayServer still.

[deralmas]

@Calinou

Not yet, as this has been attempted and it caused crashes in the past. You have to call DisplayServer.window_set_min_size() in an autoload's _ready() function instead.

It's possible that this setting can be reintroduced just fine with 4.0's DisplayServer still.

I'm pretty sure that Godot already sets a minimum window size, so I suppose that whatever issue caused these crashes isn't there anymore. Anyways we can see in another time.

[deralmas]

@Calinou I see that in the original PR you added a range hint to the content scale factor property. Why's that? Should I remove/keep it?

[Calinou]

@Calinou I see that in the original PR you added a range hint to the content scale factor property. Why's that? Should I remove/keep it?

It was an additional usability improvement, as the setting previously lacked any range hint. Only values greater than 0.5 really make sense to use, and values above 8.0 are probably too large for any project. I suggest keeping it in this PR, but it could be moved to another PR in theory 🙂

[deralmas]

@Calinou mh, I'd be tempted to make this a separate tiny PR, but I guess that since this is a content scale related PR it might make sense to make this tiny change.

I'll update the commit and the PR to document this change.

[deralmas]

Well, I think we're pretty close to mark this as ready for review. I've tried to add some documentation but I'm not sure whether it's good enough. Other than that we should be done.

I've also taken a look around the scaling code to see if I could clean up something, but everything looks both pretty intertwined and important, so I'd avoid changing things willy-nilly without being extremely sure to not have broken something. This also has the advantage of a smaller and cleaner diff, which never hurts.

Edit: wait, if the only thing left is making sure that documentation is good that means that this is ready for review! Silly me. Marking it now as such.

[Calinou]

Tested this locally (rebased against master as of today). It works as I anticipated, with one exception:

The only issue is that using the 2d stretch mode with integer scaling does not fully stretch the viewport to match the window size. What I expect is that 2D scaling will always be an integer factor, but the viewport size should match the window size (so that there are no black borders, and 3D rendering is visible across the entire window). This requires project developers to configure UI anchors correctly, but this is already an expectation when using the 2d stretch mode in general (or even disabled).

This is generally what 3D games do when they have a pixel art aesthetic but want to keep rendering at native resoluton, such as Minecraft. Fortunately, you can emulate this by using the disabled stretch mode, integer scale mode and setting the content scale factor based on float(window_height) / base_window_height. Godot will then floor() the resulting scale factor, ensuring it's always integer.¹

Nonetheless, I think this is mergeable already after applying suggestions. We can improve the 2d stretch mode later as it can be manually emulated as I explained. Great work 🙂

Testing project: multiple_resolutions.zip

Screenshots (648×648 base window size, 3840×2160 window size)

2d, fractional

2d, integer

viewport, fractional

viewport, integer

disabled, fractional, scale factor set to 200%

The result looks identical regardless of whether you set the scale mode to fractional or integer in this particular case, since 200% is already an integer scale factor.

Footnotes

1. This can be done even without this PR, but you'll have to floor() the resulting scale factor yourself. ↩

[Calinou]

I disagree, that the 2d stretch mode with integer scaling should fully stretch the viewport to match the window.

For a crisp pixel-perfect scaling mode it should just scale the base resolution of the game with the highest possible integer factor and keep it as is - so black borders are expected, if the window size is not a perfect multiple.

Unfortunately, this is a bad approach for 3D games as you don't want black borders to be present on all sides. It unnecessarily limits the visible screen real estate. We don't have a choice when using the viewport stretch mode, but we do when using the canvas_items stretch mode. If you want pixel-perfect behavior for 2D games (with everything aligning to a pixel grid), you'll want to use the viewport stretch mode anyway.

There could be an additional but slightly different stretch mode, where it uses the highest possible integer factor (let's say 3x), but then stretches the rendered image to the window size, making it slightly blurry, but retaining the initial 3x scale. If I recall correctly, some pixel-art games offered this option, so it might be interesting to test out. But I think this would require a different approach.

This is pretty much the hybrid approach I originally mentioned in my PR. (It can also be done by rendering at an integer scale, using bilinear filtering and a sharp-bilinear shader on the final native-resolution viewport.)

I also noticed a slow down / FPS drop when resizing the window with Vulkan, but that also happens on 4.0.3 release with viewport stretch / keep aspect ratio.

Recreating buffers is quite slow, especially in Vulkan. Even if the viewport size remains constant, the window buffer (which must always match the window size) has to be recreated every time you resize the window.

[georgwacker]

Unfortunately, this is a bad approach for 3D games as you don't want black borders to be present on all sides. It unnecessarily limits the visible screen real estate. We don't have a choice when using the viewport stretch mode, but we do when using the canvas_items stretch mode. If you want pixel-perfect behavior for 2D games (with everything aligning to a pixel grid), you'll want to use the viewport stretch mode anyway.

After reading your previous post again, I think I got initially confused by the terminology that the canvas_items stretch mode was called 2D.

I think it would make the most sense, to have the additional hybrid option with integer scaling + filter stretch to work similar for both canvas_items and viewport. So that use-cases for integer scaling and 3D games are covered. That means rendering with integer scale and then stretching the resulting image.

And not have canvas_items + integer work differently than viewport + integer. That means the integer stretch mode from this PR wouldn't make much sense for 3D games as is, but at least it stays consistent on what you would expect an integer scaler would do (including resulting in black bars). I would imagine most use-cases for the pure integer scale vs hybrid are 2D pixel-art games anyways.

I could look into implementing the hybrid mode.

Perhaps "Scaling Mode" would be a better name than "Stretch"?

[deralmas]

@georgwacker

Perhaps "Scaling Mode" would be a better name than "Stretch"?

IIRC it's already used.

[georgwacker]

Perhaps "Scaling Mode" would be a better name than "Stretch"?

IIRC it's already used.

I've built your branch int-scale two days ago and it's still showing up as "Stretch" under the category Stretch.

[deralmas]

@georgwacker

I've built your branch int-scale two days ago and it's still showing up as "Stretch" under the category Stretch

I meant that "stretch mode" is already used elsewhere:

godot/core/config/project_settings.cpp

Line 1318 in 2210111

GLOBAL_DEF_BASIC(PropertyInfo(Variant::STRING, "display/window/stretch/mode", PROPERTY_HINT_ENUM, "disabled,canvas_items,viewport"), "disabled");

[georgwacker]

I meant that "stretch mode" is already used elsewhere:

Ah, gotcha. Well, it could still be possible to have this option as "display/window/stretch/scaling_mode" in ProjectSettings, but if we want to avoid another mode named setting, maybe "Scaling method"?

[deralmas]

maybe "Scaling method"?

@georgwacker that's a bit confusing too.

Personally I think that the whole category should be renamed from scratch but we can't as that would break compatibility so I'd just not worry too much... It's all a mess honestly.

[Calinou]

I've retested this PR (rebased on top of master f8dbed4), it works as expected.

I think we can go with this solution. Even if it's not perfect (especially for 3D games), it's much better than the status quo. For 3D games, I can document the currently possible solution with the disabled stretch mode + a method called on resize that changes the content scale factor. It's made a bit easier by this PR as you no longer have to snap the value to integers yourself. For 3D rendering itself, godotengine/godot-proposals#4697 is a more flexible solution.

[akien-mga]

Asked reduz for opinions, he said "makes sense I guess".

[akien-mga]

Thanks!