Use less blur for distant directional shadow splits by Calinou · Pull Request #48776 · godotengine/godot

Calinou · 2021-05-16T21:48:07Z

master version of #60186.

This makes the transition between shadow splits less noticeable, specially when the expensive Blend Splits property is disabled. In many scenarios, you no longer need to enable Blend Splits to get good directional shadow appearance.

From empirical benchmarking, this seems to have a small performance cost (average 160.6 FPS over 5 screenshots instead of 163.8 FPS). I'm not sure if this is introduced by benchmarking. It's not very scientific since I carried this out from the editor and its fluctuating frame time information.
Edit: Further testing didn't reveal any performance difference.

This closes godotengine/godot-proposals#2729.

Note: Not cherry-pickable to the 3.x branch, since this relies on the per-light shadow blur feature found in the new Vulkan renderer.

3.x equivalent of this PR (doesn't work with Blend Splits enabled yet): https://github.com/Calinou/godot/tree/directional-shadow-distant-split-lower-blur-3.x

TODO

Don't hardcode the blur factors. Instead, calculate them based on the effective "share".
- By default, the multipliers would be:
  - 1.0 for the first split.
  - 0.5 for the second split.
  - 0.2 for the third split.
  - 0.1 for the fourth (and last) split.
- These multipliers should change automatically when the split offsets are adjusted in the DirectionalLight.
Adapt the mobile renderer shader code as well.

Preview

Note: GI and ambient lighting was disabled to make the difference in shadow rendering more obvious.

Before

Distant shadows have less detail. There is a noticeable "cut" between splits near the lamp posts in the middle of the image.

After

Distant shadows keep a similar level of detail. Despite Blend Splits being disabled, the cut between splits is now much less noticeable.

jcostello · 2021-08-30T12:56:23Z

I was about to create a proposal for this. Is this and the blend splits at draft?

Calinou · 2021-08-30T13:20:07Z

I was about to create a proposal for this. Is this and the blend splits at draft?

Yes, this PR needs to be finished before it can be merged. Note that there's already a proposal open: godotengine/godot-proposals#2729

Calinou · 2021-10-22T22:13:36Z

Rebased and tested again. Some changes I did:

Refactored to avoid repetition.
Bias splits' share is now calculated automatically.
Copied over the changes to the mobile renderer. Note that light rendering is currently broken there, but it's unrelated to this PR.

The performance impact seems to be unnoticeable in practice.

Before

After

mrjustaguy · 2022-04-05T12:57:24Z

Any reason why this PR has been dormant for so long, or just faded into the backlog of PRs to check?

Calinou · 2022-04-05T14:32:04Z

Any reason why this PR has been dormant for so long, or just faded into the backlog of PRs to check?

It needs a review by a rendering contributor (other than me) 🙂

Calinou · 2022-04-10T18:22:56Z

I applied @clayjohn's suggestions and tested again on both clustered and mobile backends (with and without Blend Splits). It looks good, even when split distances are customized.

PS: I found a way to further improve shadow appearance by scaling the normal bias depending on the split's share. This is in a separate branch for now, but I can include it in this PR if you want: https://github.com/Calinou/godot/tree/directional-shadow-normal-bias-split
With both changes, I'd say Blend Splits becomes unnecessary 99% of the time 🙂

mrjustaguy · 2022-04-10T19:18:03Z

Blend Splits with consistent blur levels alone probably becomes unnecessary for plenty of scenes. If Normal bias change adds to that.. especially if and when godotengine/godot-proposals#3908 (spherical Splits) become a thing... Blend Splits will really be a nearly pointless feature for most use cases.. 😃

mrjustaguy · 2022-04-10T19:34:35Z

Eh, I've noticed how Shadow Blur quality setting is now very inconsistent in what it's doing...

For example, Soft Shadows Low look the softest, while going towards Ultra generally slowly decreases how soft they are, and the level of graininess is about the same.

Also, I didn't test out the Normal tweaked branch, but this seems to negate the need for #55757 Normal bias setting changes

Calinou · 2022-04-10T20:00:56Z

For example, Soft Shadows Low look the softest, while going towards Ultra generally slowly decreases how soft they are, and the level of graininess is about the same.

I can reproduce this on the Vulkan Clustered backend, but not the Vulkan Mobile backend. Can you reproduce this on a vanilla master branch and previous 4.0 alphas?

Also, I didn't test out the Normal tweaked branch, but this seems to negate the need for #55757 Normal bias setting changes

The shadow bias still needs to be increased up close, so I think #55757 is still relevant.

clayjohn · 2022-04-11T05:00:07Z

For example, Soft Shadows Low look the softest, while going towards Ultra generally slowly decreases how soft they are, and the level of graininess is about the same.

This appears to be a regression in master. It looks like the enums are misaligned or something.

clayjohn

Looks great! Once we resolve that shadow regression in master this will look even better!

mrjustaguy · 2022-04-11T05:29:17Z

Alpha 6 has no issues, only Clustered has this issue as stated above.

On another note, the Normal Bias PR increases shadow acne considerably in some scenarios, such as on a flat plane when the sun is just rising, however in other cases it massively improves quality. (16k shadow map can easily do 10-20km view distance with it)

And I found why Normal Bias at 2 is a good idea, highly curved objects clear out micro-shadow acne quite well... and the issue that I've reported for Directional Shadows (#55757 (comment)) is no longer effectively present, well in the Normal bias split branch.

mrjustaguy · 2022-04-11T15:05:27Z

I think I may have figured out why there is such shadow acne appearing on a Plane when the plane is viewed (by the Directional Shadow) from such an angle (85-90 degrees angle from Plane's Normal)
The Shadow Map for that Plane is turning into essentially a Pixel Line, which doesn't have enough information for Shadows to be Acne-free, and there is an extremely a Sharp change in the Depth.

There might be a solution for this, which would be to detect Large differences in Depth on the Shadow Map and based on that increase Bias for the areas of the Shadow Map that are affected..

Here's a Sketch of the base Idea:

White - Depth Far
Black - Depth Close
Red - Huge difference between Depths of two neighboring Pixels - Increased Bias for any shadows rendered using those, proportional to the difference between Depths, the Higher the Depth Difference, the More Bias (Not Normal Bias, regular Bias) is added

Note: This idea could also be applied to Spot and Omni Lights (especially Omni Lights) to Reduce their Shadow Acne in cases like #55757 (comment)

This makes the transition between shadow splits less noticeable, specially when the expensive Blend Splits property is disabled.

Calinou · 2022-04-11T17:49:56Z

Tested again after rebasing on latest master. This should be good to merge now.

Before

After

Performance is pretty much identical.

Testing project: test_shadow_b.zip

clayjohn

Yep, looks great rebased on master. Great work!

akien-mga · 2022-04-12T06:37:23Z

Thanks!

Calinou added enhancement topic:3d topic:rendering labels May 16, 2021

Calinou added this to the 4.0 milestone May 16, 2021

Calinou mentioned this pull request May 23, 2021

Enable Blend Splits in DirectionalLight3D by default #40971

Closed

Calinou mentioned this pull request Jul 24, 2021

Use dithering to fade DirectionalLight shadow splits between each other #50798

Draft

Calinou force-pushed the directional-shadow-distant-split-lower-blur branch from 5af4c62 to f8c168e Compare October 22, 2021 22:11

Calinou marked this pull request as ready for review October 22, 2021 22:13

Calinou requested a review from a team as a code owner October 22, 2021 22:13

Calinou mentioned this pull request Oct 25, 2021

Use lower shadow blur for distant directional shadow splits godotengine/godot-proposals#2729

Closed

Calinou mentioned this pull request Dec 9, 2021

Automatically increase Light3D shadow bias when shadow blur is above 1.0 #55758

Closed

1 task

clayjohn reviewed Apr 6, 2022

View reviewed changes

Comment thread servers/rendering/renderer_rd/shaders/scene_forward_clustered.glsl Outdated

Calinou force-pushed the directional-shadow-distant-split-lower-blur branch from f8c168e to 825f693 Compare April 10, 2022 18:22

Calinou force-pushed the directional-shadow-distant-split-lower-blur branch from 825f693 to e093668 Compare April 10, 2022 18:30

clayjohn approved these changes Apr 11, 2022

View reviewed changes

clayjohn mentioned this pull request Apr 11, 2022

Specialization constants not updating #60132

Closed

Use less blur for distant directional shadow splits

1a41a17

This makes the transition between shadow splits less noticeable, specially when the expensive Blend Splits property is disabled.

Calinou force-pushed the directional-shadow-distant-split-lower-blur branch from e093668 to 1a41a17 Compare April 11, 2022 17:47

clayjohn approved these changes Apr 12, 2022

View reviewed changes

akien-mga merged commit 35d2efc into godotengine:master Apr 12, 2022

Calinou deleted the directional-shadow-distant-split-lower-blur branch April 12, 2022 14:27

This was referenced Apr 12, 2022

Use lower shadow normal bias for distant directional shadow splits (reverted) #60178

Merged

[3.x] Use less blur for distant directional shadow splits #60186

Open

Calinou mentioned this pull request May 8, 2022

Change Directional Shadow Splits from Percentage of Max Distance to Distance godotengine/godot-proposals#4512

Open

mrjustaguy mentioned this pull request Aug 1, 2022

Allow non-power-of-two directional shadow size in 3D #54041

Open

mrjustaguy mentioned this pull request Feb 21, 2023

Distance Directional Shadows Pixelize as you get further away #73677

Closed

mrjustaguy mentioned this pull request Dec 13, 2024

Distant Shadow Splits have Wrong Bias #100375

Open

Uh oh!

Conversation

Calinou commented May 16, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

TODO

Preview

Before

After

Uh oh!

jcostello commented Aug 30, 2021

Uh oh!

Calinou commented Aug 30, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Calinou commented Oct 22, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Before

After

Uh oh!

mrjustaguy commented Apr 5, 2022

Uh oh!

Calinou commented Apr 5, 2022

Uh oh!

Uh oh!

Calinou commented Apr 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mrjustaguy commented Apr 10, 2022

Uh oh!

mrjustaguy commented Apr 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Calinou commented Apr 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

clayjohn commented Apr 11, 2022

Uh oh!

clayjohn left a comment

Choose a reason for hiding this comment

Uh oh!

mrjustaguy commented Apr 11, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mrjustaguy commented Apr 11, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Calinou commented Apr 11, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Before

After

Uh oh!

clayjohn left a comment

Choose a reason for hiding this comment

Uh oh!

akien-mga commented Apr 12, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Calinou commented May 16, 2021 •

edited

Loading

Calinou commented Aug 30, 2021 •

edited

Loading

Calinou commented Oct 22, 2021 •

edited

Loading

Calinou commented Apr 10, 2022 •

edited

Loading

mrjustaguy commented Apr 10, 2022 •

edited

Loading

Calinou commented Apr 10, 2022 •

edited

Loading

mrjustaguy commented Apr 11, 2022 •

edited

Loading

mrjustaguy commented Apr 11, 2022 •

edited

Loading

Calinou commented Apr 11, 2022 •

edited

Loading