Workaround for feathering of sharp edges

[podusowski]

I am working on rendering vector maps in Walkers (which looks promising!) and I noticed weird artifacts:

It appears that the issue is caused by edge normals, which kind of explode when the shape is sharp enough. This PR adds a limiter, so that the normal is clamped when it crosses certain length. It's not perfect and most likely can be fixed in more elegant way, but it does seem like an improvement:

[github-actions]

Preview available at https://egui-pr-preview.github.io/pr/7554-feathering-fix
Note that it might take a couple seconds for the update to show up after the preview_build workflow has completed.

View snapshot changes at kitdiff

[emilk]

This looks very promising! Thanks for working on this.

There is some related code here:

egui/crates/epaint/src/tessellator.rs

Lines 404 to 418 in b36a259

	let normal = (n0 + n1) / 2.0;
	let length_sq = normal.length_sq();
	let right_angle_length_sq = 0.5;
	let sharper_than_a_right_angle = length_sq < right_angle_length_sq;
	if sharper_than_a_right_angle {
	// cut off the sharp corner
	let center_normal = normal.normalized();
	let n0c = (n0 + center_normal) / 2.0;
	let n1c = (n1 + center_normal) / 2.0;
	self.add_point(points[i], n0c / n0c.length_sq());
	self.add_point(points[i], n1c / n1c.length_sq());
	} else {
	// miter join
	self.add_point(points[i], normal / length_sq);
	}

and here:

egui/crates/epaint/src/tessellator.rs

Lines 450 to 472 in b36a259

	// We can't just cut off corners for filled shapes like this,
	// because the feather will both expand and contract the corner along the provided normals
	// to make sure it doesn't grow, and the shrinking will make the inner points cross each other.
	//
	// A better approach is to shrink the vertices in by half the feather-width here
	// and then only expand during feathering.
	//
	// See https://github.com/emilk/egui/issues/1226
	const CUT_OFF_SHARP_CORNERS: bool = false;
	let right_angle_length_sq = 0.5;
	let sharper_than_a_right_angle = length_sq < right_angle_length_sq;
	if CUT_OFF_SHARP_CORNERS && sharper_than_a_right_angle {
	// cut off the sharp corner
	let center_normal = normal.normalized();
	let n0c = (n0 + center_normal) / 2.0;
	let n1c = (n1 + center_normal) / 2.0;
	self.add_point(points[i], n0c / n0c.length_sq());
	self.add_point(points[i], n1c / n1c.length_sq());
	} else {
	// miter join
	self.add_point(points[i], normal / length_sq);
	}

I wonder if that is still needed 🤔

Also worth checking that we don't regress on #1226

thanks for adding a test btw 🙏

[podusowski]

I did couple of changes to the test, namely to be more like #1226. CUT_OFF_SHARP_CORNERS is not affected by this change, I mean, it still needs to be turned off, otherwise jagged edges come back.

I've put the limit into TessellationOptions, so now it's easier to play with it.

Unfortunately, although IMO it's still better to have this, but it seems that there are no golden value here.

Baseline:

[emilk]

I think this is incorrect.

if normal = [10, 0] for instance, and max_feathering_normal_length = 2, so max_feathering_normal_length_sq = 4, then we would divide by 100/4 instead of 10/2.

Suggested change

	if normal.length_sq() > max_feathering_normal_length_sq {
	normal * (max_feathering_normal_length_sq / normal.length_sq())
	if normal.length_sq() > sqr(max_feathering_normal_length) {
	max_feathering_normal_length * normal.normalize()

[podusowski]

I had some issues when I did literal clamping, not sure why because now it looks fine (more or less).

Baseline:

Default (40.0):