Shadow performance and quality -- will it be improved? Will we ever get better soft shadows?

Pretty sure Unity can't use any of these unless you make PRs to their Graphics repo with these changes

 

This is trivial to achieve in URP, but you'd have to use unlit and roll your own lighting, or make use of emission output. It definitely is doable but cumbersome.

 

Fair enough; I didn't look hard enough at it; looks like MadWorld is basically just using unlit for everything. Bit trickier with cel-shaded games that need to take lighting info from the actual lights (realtime, probes, baked, etc) and do something with it other than PBR, although given the number of cel shading shader graphs on the asset store, certainly not an unsolvable problem.

 

Hi @nigeljw_unity,

I just tested conservativeEnclosingSphere and numIterationsEnclosingSphere in 2021.2.15f1. There are worrysome things that I noticed while enabling the setting and tweaking the values, both from numIterationsEnclosingSphere and cascade Splits:

• The shadow drawcall count skyrockets between +30% and +200%, depending on FOV, far plane, shadow distance, cascade count and cascade splits. This can be a huge performance regression for CPU.
• It also seems like a lot of texel density (shadow quality) is lost from enabling conservativeEnclosingSphere. You can clearly see this when looking at the shadowmap in the frame debugger. If a game uses only 1 shadow cascade, there is no possibility to tweak the cascade splits, so no possibility to get back that lost texel density after enabling conservativeEnclosingSphere. This also seems like a big visual quality regression.
• Those problems are very apparent on a perspective camera. Orthographic camera seems little to not affected by conservativeEnclosingSphere.

I'd like you guys to highlight those potential problems if those don't get fixed. And keep the conservativeEnclosingSphere toggle exposed. I'm not even sure if conservativeEnclosingSphere should be enabled by default honestly, as shadows can be a huge bottleneck.

On the matter of directional shadowmap texel density, I've always struggled with understanding how the orthographic size of the shadow projection is generated. I've noticed that a lot of the times it seems the projection is too large, like a lot of texture space is wasted. And as I said above, this is very problematic with conservativeEnclosingSphere. Is there a way you could expose the size of the shadow projections to users?

In BiRP, there was a setting ShadowProjection.CloseFit which reeeally tightened the projection and texel density. Are there any plans to bring CloseFit back? If not, is there a way that we can reproduce its result?

About ScriptableCullingParameters.shadowNearPlaneOffset: I am grateful for this addition. Sounds like less fighting with the SetShadowCullingPlane API. I noticed that pushing a negative value will reduce shadow drawcalls and result in better texel density, which is good. On the contrary, pushing a positive value results in lower texel density, BUT also reduces shadow drawcalls, which doesn't sound expected. Why does it reduce drawcalls?

For the record, there is also a "shadowNearPlaneOffset" setting in QualitySettings when using BIRP, which does something completely different and does not influence culling. Could be good to clarify this up.

Cheers!

 

@nigeljw_unity
Any ideas on what might be coming to URP in regards to shadows for 2022.2?
Any noticeable new features/improvements?

 

Shadow culling in Unity has always been lackluster, and URP inherited most of that. Here are two common and relatively simple techniques Unity could use which can drastically reduce the number of rendered shadow casters:

• Exclusive cascade culling: if an object bounding volume is completely inside a cascade, exclude it from the next cascades.
  
• Extruded frustum shadow culling: to cull shadows which would fall outside the camera frustum, create a convex by extruding the frustum volume in the inverse light direction and cull casters which don't overlap it. The extruded convex has at most 12 planes compared to a frustum 6 planes.

 

Will URP get capsule shadows in 2022.2, it's coming for HDRP.
Curious since I saw the author experimenting with URP for a bit as well.

 

Not sure how notifications work when a thread is moved but for anyone interested (mainly thinking of those Nigel directed to this thread for response) the topic I started is now here:

https://forum.unity.com/threads/1169018-shadow-length-is-reverting-back-every-few-frames.1261622/

 

From 2022.2.0a9 release notes:

 

Nice!
Although I heard from the next-gen shadows author that this is as good as it's gonna get, Tent 7x7. They don't plan for anything better after that. This was a week or two ago.

 

Thanks for the much needed explanation, it helps a ton.

 

"Medium 5x5 soft shadows": (previous default)


"High 7x7 soft shadows": (only 2022.2)



next-gen shadows:


next-gen lower smoothness value & higher res (pretty low on previous screenshot):

I have to say, 7x7 is better than 5x5 so no complaints from me. Great stuff
But, next-gen is on a whole new level, and completely adjustable including smoothness value.. a bit disheartening
It almost feels raytraced.

 

@PutridEx,
Side NGSS note, using a PCF with min softness at min and max at max would provide pseudo raytraced like shadows where shadows are harder at close distance between caster/receiver and softer if caster/receiver distance is large.
The reason why Unity opted for a fixed kernel is because is more predictable for drivers to optimize and tend to miss cache less than variable kernel size algorithms such as the ones in NGSS. However this is now less true as even mobile drivers start to get a lot better at optimization and cache miss.

 

Hi @nigeljw_unity

may I ask if there is any ETA on the shadow caching for URP yet? Is this still planned?

"Cached static shadows is very high in priority on our raster shadow roadmap. The only thing we actually have as a higher priority at the moment is moving the shadow depth bias from the caster to the receiver"

This was almost one year ago but I can not find any info regarding this anywhere. With no mention of this feature in the official Unity URP Roadmap.

Even a very rough ETA would be immensely helpful.
Especially considering that Unreal implemented Shadow Caching more than 4 years ago (see Fortnite mobile)...
This is somewhat discouraging for every Unity Developer...

 

Where do you find the options for the 5x5 and 7x7? Using 2022.2 a11 rn and its nowhere to be seen for me

 

I believe it's configurable per light, I did my test with the directional light.
Click on the directional light and in the inspector, the option should be somewhere around the shadow settings.

 

Ah I see yeah.

I tried to compare it to 2021 LETS but the shadow quality is still the same no matter what, i wish NGSS would support URP already, unitys default shadows are just so bad

 

The extensibility NGSS uses to work on built-in doesn't exist in URP. The only viable way, AFAIK, would be to make a fork of URP since those changes would need to go directly into its source code.

 

@nigeljw_unity it is not as simple as re-authoring cascades to fix everything. Metric working unit is plain wrong when conservative enclosing sphere is enabled. (Last Border is also wrong in all cases, it controls 0-100% Max Distance fade range no matter what the inspector and documentation says).

Unity 2021 LTS URP 12.1.7
Cascade splits 40, 60, 100, max 150 Metric without conservative, fairly accurate and goes a bit beyond max distance:



And with conservative, same settings, doesn't match metric at all and goes way beyond max distance, note that the last cascade is not visible at all:


Shadow map without conservative:



Shadow map with conservative:



With conservative mode the metric working unit is plain wrong. Percent works, but you have no idea how far the shadows extend because they go way beyond the max distance, rendering useless shadows. Re-authoring cannot avoid rendering many unneeded shadows here.

I rather use the old approach, plus improved shadow fading that also takes the outer cascade sphere into account.

 

@nigeljw_unity Thank you for your reply!

It would be so great if this was actually documented! Whatever the non-conservative mode does is way more intuitive as it better matches the suggested distances and seems to have a sensible relationship with the max distance.

The last cascade is practically useless because it is not visible as it lies beyond the max shadow distance. Note where the shadows get cut off, about halfway the yellow band in the second image. Thus it is a performance problem because way too many shadows casters get rendered. The max distance has become an arbitrary cut-off line for showing shadows, but they still get rendered way beyond that distance.

Neither in the second image. First is missing cascades 0 and 1, while second is missing cascade 0. This is because those cascades are used for close-ups while the current view is high-altitude. In the actual project I only use two cascades, which have the same problem as shown here - most of the last cascade is wasted - it's just more visually obvious with four cascades.

Yes: IN-21329 - URP incorrect shadow cascades and fading. In there I also show the max shadow distance on the terrain to make it easy to notice how much shadows get rendered for nothing.

I suggest eliminating the Metric mode entirely because it is plain wrong.

 

@nigeljw_unity Thanks for the explanation!

Now that I better know how it works the results and differences make sense. Although I'm still fuzzy about the exact definition of the cascade splits. What kind of frustum size? Without feedback to visualize it it's a fairly arbitrary value. Likewise with percentage. It gets tuned visually without any idea what the numbers actually represent.

Currently conservative doesn't get better for ultra wide displays either. Here using two cascades (debug view cannot deal with a single cascade; it also bugs out if everything is beyond max distance by the way).

Nonconservative:


Conservative:


Nonconservative shadow map:


Conservative shadow map:


I'd actually be interested in the old approach taken to the extreme: place cascades based on Max Distance only, don't demand the entire frustum be filled, so the squiggly filled missed regions in your second diagram are fine. Then fade shadows based on the cascade sphere, just like fading between cascades.

In my case this makes sense because there is usually nothing behind the camera, so overlapping culling spheres don't lead to many useless shadow casters being rendered.

Basically it's the current max distance fade that makes the conservative approach silly. Including that line in your first diagram - if I understand it correctly - shows the problem with the last cascade:


(Not sure about the placement of the distance sphere, but it illustrates the problem nonetheless.)

This suggests to toss the current distance-based fade and instead fade based on the cascade sphere and suddenly many more of the rendered shadows become useful. Then max distance can be tuned to an acceptable level (in the diagram it got basically doubled). But then the old approach makes more sense again, as I described in the previous paragraph, or something that combines the best of both.

Another potential issue I see with the conservative approach is that if cascade blending regions are too large - if smaller cascades aren't perfectly contained by larger ones - holes might appear near the edge of the frustum.

 

What about adding a configurable minimum renderer-bounds-to-cascade-size ratio to cull casters? It would be nice to remove small props from the larger cascades. It makes no sense that hundreds of props that won't even be a pixel wide to be rendered to the shadowmap.

Another really nice thing would be to have a fully configurable cascade count. Huge for space games: Elite Dangerous uses 6 cascades, allowing for real time shadows across huge asteroid fields AND your own ship, which would be unworkable with just 4 cascades.

 

This has probably already been proposed, but who knows: what about rendering the shadows at half resolution? And then apply a median filter to reduce the blocky look.
Would it give a good boost or not so much?