Viewport out-of-range when first constructing certain UIs rendered with egui-wgpu callbacks

[msparkles]

Describe the bug

Viewport out-of-range, causing a validation error, when first constructing certain UIs rendered with egui-wgpu callbacks.
This only happens on the first frame, the fix would simply be skipping the callback and set viewport command when the viewport rect is outside of the display.

To Reproduce
Steps to reproduce the behavior:

1. Make a Bottom panel with a scroll area inside
2. Add egui-wgpu Callbacks to it

Expected behavior
The item to be skipped if the viewport is outside of display.

Additional context
https://github.com/automancy/automancy/blob/main/src/bin/automancy/gui/tile_selection.rs

(Remove the following from the code to actually get the panic:

        if !ui.ctx().screen_rect().contains_rect(rect) {
            continue;
        }

)

[msparkles]

This needs an egui fix because the scissor could be smaller than the viewport.

[emilk]

Is it the PaintCallback::rect that is wrong, or the clip rect, or both? What are their values compared to the viewport rect?

Why does it only happen on the first frame?

the fix would simply be skipping the callback and set viewport command when the viewport rect is outside of the display.

what about when it is partially outside the frame?

[msparkles]

Actually, we just thought about it for a bit- the problem should be that the scissor isn't being correctly set..

[msparkles]

Why does it only happen on the first frame?

Regarding this, we can only assume it's some placement negotiation oddities that's happening, and that's beyond our knowledge to really understand..

Here's a screenshot of that first frame

[msparkles]

Gonna work on this right now and implement a fix. \o/

[msparkles]

Okay, to recap everything:

• Apparently there's a discussion? on this-- We came from Vulkan, we used to render outside of the render attachment and clip stuff with scissor, so that we can render half an object-- Viewport rect must lie entirely within the current attachment size? gpuweb/gpuweb#373 apparently this is an.... issue in wgpu?
• Apparently there's still a minor bug in egui where it would try to set a viewport of width and/or height of 0, just plug this before the set viewport command- this is unrelated to this issue

if viewport_px.width_px <= 0 || viewport_px.height_px <= 0 {
    continue;
}

• Part of this is an user error on our part- with that piece of code patched onto egui, the issue now is this:

Aka. The Thing Is Not Supposed to be Squished- but this is apparently wgpu's issue now?

[msparkles]

CC: @cwfitzgerald You're the wgpu person right? :P

tldr: apparently WebGPU? says that, you should be able to set viewport to be outside of attachment, but wgpu doesn't allow that

[emilk]

I think @Wumpf has thought about this before when we implemented a similar thing in Rerun

[Wumpf]

Today, WebGPU specification requires to fail validation on the device timeline if the viewport is outside of the render targets, see https://www.w3.org/TR/webgpu/#dom-gpurenderpassencoder-setviewport so wgpu behaves correctly here.

With that am I correct that the only thing wrong on the egui side is that it should discard zero area viewports?
Not sure actually whether it should still invoke egui render callbacks 🤔, I guess it should.

[msparkles]

Today, WebGPU specification requires to fail validation on the device timeline if the viewport is outside of the render targets

That is, rather annoying.. is there any other way to achieve this rendering then? Partially out of range and then clipped into the screen.

With that am I correct that the only thing wrong on the egui side is that it should discard zero area viewports?

Yeah.

Imo it should skip the whole callback, because if not updated it'll just use the older viewport, which is gonna override other stuff.

[cwfitzgerald]

That is, rather annoying.. is there any other way to achieve this rendering then? Partially out of range and then clipped into the screen.

Intersection of your desired rect and the screen viewport rect.

[emilk]

Shapes with empty clip rects are already pruned by the epaint tessellator before the results are handed to the rendering backend:

egui/crates/epaint/src/tessellator.rs

Lines 1140 to 1142 in 34c3414

	if !new_clip_rect.is_positive() {
	return; // skip empty clip rectangles
	}

[msparkles]

In here..

egui/crates/egui-wgpu/src/renderer.rs

Line 453 in 34c3414

if callback.rect.is_positive() {

The rect is checked during callback...

...but that rect isn't clamped to screen size yet.