Bindless Deferred Decals in The Surge 2

Bindless Deferred Decals in
Philip Hammer
(DECK13 Interactive GmbH)
Digital Dragons 2019, Krakow

Philip Hammer (Deck13 Interactive)
This talk is ..
● .. about
○ our own implementation of bindless deferred decals in D3D12 and Vulkan
○ overcoming problems of past approaches
○ practical bread & butter techniques
○ how to eﬃciently render many decals on the GPU
○ tackling common issues and glitches
● .. but unfortunately not
○ super-fancy new research
○ a comprehensive view on our new renderer
● .. and has
○ probably more too much source code
2
2

Introduction
● Brief overview of the latest iteration of our inhouse FLEDGE engine
● 4th generation: re-designed renderer module from ground up:
○ Vulkan + D3D12 support only
○ Dropped legacy code (D3D11)
○ No more OOP code - all in for DOD / SoA
○ New possibilities for high-level renderer design (e.g. bindless texturing)
● Improved some existing features e.g.
○ Clustered Deferred + Clustered Forward Rendering
○ Physical-based Rendering
● Some fancy new rendering features, e.g.
○ Uniﬁed Volumetric Lighting, Irradiance Probe Grid, Simulated interactive grass, Temporal Antialiasing, etc.
○ Unfortunately not part of this talk :-)
3
3

Motivation
● Decals = one key beneﬁt of deferred rendering
● Used in environment art + VFX
○ Static decals to detail the world
○ Dynamic decals spawned by the effect system
○ Gameplay-driven decals like footsteps
● High demands on decal count, usability and performance
8
8

Decal usage & examples
99
9

Rasterization-based deferred decals
● Had deferred decals for quite a while [LotF14]
○ Be aware of decal count, overdraw, etc.
○ Limitations led to sparse usage
● “Classic” deferred decal rendering
○ analogous to deferred light volume rendering
● Improved that a lot over the years, but basic principle stayed the same
16
16

Rasterization-based deferred decals
● Advantages
○ Easy to implement and possible with most
(even older) graphics APIs
○ Near pixel-perfect culling and shader segmentation
● Disadvantages
○ CPU-Performance
■ one drawcall per decal
■ Lots of state setup per decal
○ GPU-Performance
■ Double work: fetching gbuffers,
rasterizing fragments, etc.
■ Need to update the G-Buffer as
intermediate blend buffer
17
17
Transformed box geometry rasterizing a decal

Bindless Deferred Decals
● How about rendering all decals at once?
18
18

● Problem: Too many resource bindings
○ Each decal references a potentially unique set of textures
○ Not possible to bind arbitrary amount of texture resources
19
19

● Problem: Too many resource bindings
○ Each decal references a potentially unique set of textures
○ Not possible to bind arbitrary amount of texture resources
● Solution(s):
○ Texture atlases or arrays
-> limited and inﬂexible
○ is around since 2013 [Everitt14]
-> vendor speciﬁc and what about consoles?
20
20

Bindless Texturing
● D3D12 + Vulkan to the rescue!
○ Built-in support for bindless texturing
○ API speciﬁc descriptor arrays
○ Create a descriptor table for all textures in the game
○ Descriptors are globally available and bound to any draw/dispatch in need
21
21

Bindless Texturing
● D3D12 + Vulkan to the rescue!
○ Built-in support for bindless texturing
○ Create an (API speciﬁc) descriptor array for all textures in the game
○ Descriptors are globally available and bound to any draw/dispatch in need
● Idea of bindless texturing is around for a while!
○ e.g. [Reed14], [Pettineo16], [Sousa16], [Drobot17]
22
22

Bindless Texturing
D3D12 / HLSL Vulkan / GLSL
23
23

Non-uniform resource indices
● Texture index must be uniform
○ In GCN terms: loaded into an SGPR
○ It’s not by default if fetched from a GPU buffer
○ Non-uniform indices lead to unexpected results
and graphical corruptions
● Need to explicitly make the index uniform!
24
24

Non-uniform resource indices
● D3D12 / HLSL:
● Vulkan / GLSL:
○ Check [GpuInfo] for availability
● Alternative: use wave intrinsics
○ D3D12: Shader Model 6 wave intrinsics
○ Vulkan: ballot extensions
● Heavily driver- and vendor-dependent
○ Availability of extensions and performance
25
25

Clustered Deferred Decals
● We already have a clustered deferred renderer for lighting
○ Similar approach to [Persson], [Sousa16], etc.
○ Cull lights, specular cubemaps, etc. into subdivided view frustum bins
○ Render everything in one go (single compute dispatch / fullscreen quad)
○ Easy to query clustered data also for forward-rendered objects
○ Already a good coverage on the topic, so we won’t go much deeper here
26
26

● We already have a clustered deferred renderer for lighting
○ Similar approach to [Persson], [Sousa16], etc.
○ Cull lights, specular cubemaps, etc. into subdivided view frustum bins
○ Render everything in one go (single compute dispatch / fullscreen quad)
○ Easy to query clustered data also for forward-rendered objects
○ Already a good coverage on the topic, so we won’t go much deeper here
● Good ﬁt also for decals!
27
27
Output
scene color
Fetch
G-Buffer
Compute
decals
Compute
direct
lighting
Lighting shader
Compute
indirect
specular
Compute
indirect
diffuse

● Decal blending only in registers, no double fetching or updating of G-Buffer
● Very eﬃcient, but be aware of register pressure / low CU occupancy
28
28

Bindless Decals Optimization
● High-level art optimizations
○ Cull by distance with soft fade-out
○ Switch to simpler feature set over distance
29
29
Decal overlay icons for all decals

● High-level art optimizations
○ Cull by distance with soft fade-out
○ Switch to simpler feature set over distance
30
30
Decal overlay icons for all decals surviving distance culling

● Low-level optimizations
● Biggest bottleneck in shader: VGPRs / CU Occupation
● GCN Loop Scalarization works great with decals
○ Load decal and material data in SGPRs and loop over necessary lanes
○ Note: not necessary anymore!
● Splitting decal and material features into isolated blocks
○ Free up registers by grouping features
○ Register lifetime is not always obvious - measure and proﬁle
○ Reorganization of large shaders with a lot of resource access is always worth a try!
31
31

GCN loop scalarization
● Great in-depth coverage available [Drobot17]
● Make cluster index uniform / scalar
○ Load all related data (DecalData, MaterialData, etc.) into SGPRs
○ Index is actually not uniform, but most pixel in a wave likely hit
the same index
○ Double work can happen but higher occupancy will cancel it out
○ Only doable with wave intrinsics
32
32

Bindless Decals Limitations
● Downsides of blending decals in registers
● Decals not visible for other (deferred) renderpasses
○ normals + albedo won’t see decal modiﬁcations (e.g. SSAO, SSR, separate ambient pass)
○ Material-ID not updated (TAA Ghosting! [Xu16])
● Could render decals before lighting and update G-Buffer
○ Additional costs for e.g. copying G-Buffer Normals
33
33

Common decal issues
34

Common decal issues #1: Normal blending
● Linear blending is trivial and useful for e.g.
snow accumulation
○ Perfect for thick covers like layers of sand or snow
○ Tends to “ﬂatten” normals in gradual blending:
normal directions cancel out each other
35
35
Decal with linear normal blending

Common decal issues #1: Normal blending
● Re-orientated normal mapping [Hill12]
○ Our default mode for blending, works best for most decals
○ Decal normal will always follow surface normal
○ Fits mostly better into the scene
36
36
Decal with re-oriented normal blending

Common decal issues #2: screenspace UV gradients
● UV- and Z-gradients cannot be derived correctly in screenspace
○ No hardware gradient data (ddx/ddy) available in shader loops
○ No mip-mapping and wrong UV-gradients around object boundaries
● Solution:
○ Sample a ﬁxed Mip Level :-(
○ Compute gradients in shader
■ Slightly higher runtime cost
■ Need to compute 2 addition decal-space positions with 1-pixel-offset in X and Y direction
■ Be aware of decal & uv scaling!
37
37

Common decal issues #2: screenspace UV gradients
● Computation of UV gradients in the shader
● Be aware that hardware gradients of position is not correct
38
38

Common decal issues #3: screenspace Z gradients
● UV gradients = almost good
● But can get huge at depth discontinuities
if gradient is computed in the shader
● Two possibilities
○ Store z-gradients in G-Buffer and compute
with these
○ Or check for depth discontinuities
in shader + ﬁx-up [Hammer18]
39
39

● Checkerboard is a prime example because of
maximum contrast
○ lower mips converge to middle gray due to downscale
● Large UV gradients will fetch a very low mip
ending up in a color discontinuity
40
40

● Find the edges causing the trouble
● Sample Lod0 if fragment is on edge
● Can be done directly in the decal fullscreen
shader based on two depth quads
41
41

● Find the edges causing the trouble
● Sample Lod0 if fragment is on edge
● Can be done directly in the decal fullscreen
shader based on two depth quads
42
42

● Color discontinuities are now gone
● Fixed edges can now suffer from Mip-Flickering
● In practice a rather minor issue
○ Barely noticeable during gameplay
○ Antialiasing - especially Temporal AA - migitates as well
43
43

Common decal issues #4: Leaking and object separation
● Decals may leak into areas where they don’t belong
44
44

Common decal issues #4: Leaking and object separation
● Decals may leak into areas where they don’t belong
● Solution:
○ Decals and materials both store a “DecalGroup” bitmask
○ Bitwise AND the masks and check if at least one bit matches
○ Up to artists how to assign bits (e.g. character, vegetation,
environment, transparent, etc.)
45
45

Dirty decal tricks
4646
46

Dirty decal tricks #1 - Smoothing Edges
● Modify decal normal to “smooth” corners and edges
● Interpolate surface normal towards (uniform) decal direction
● Simple but effective and artists love the hack
● Useful for covering uneven geometry e.g. with mud
47
47
Smoothed Edges off (smoothFactor = 0.0)

Dirty decal tricks #1 - Smoothing Edges
● Modify decal normal to “smooth” corners and edges
● Interpolate surface normal towards (uniform) decal direction
● Simple but effective and artists love the hack
● Useful for covering uneven geometry e.g. with mud
● Geometry normal is otherwise hard to override
● Be careful, it’s a hack
48
48
Smoothed Edges on (smoothFactor = 1.0)

Dirty decal tricks #2 - Water puddles
● Decalling shallow water puddles
○ Faking surface porosity by blending
black decal albedo (only 10-20% opacity)
○ low roughness (100%)
○ water ripple normal (100%)
● No special code path, just special
parametrization
49
49

Dirty decal tricks #3 - Texturing the world
● Base model (no decals)
50
50

● Base model + decal with UV-free texturing
aka Triplanar mapping
○ Sample texture(s) 3x from different directions
and blend based on decal normal
○ Triplanar contrast deﬁnes the “steepness”
of the blending according to the normal
51
51

● Base model + decal with triplanar +
slope dependent blending
○ Generate a blend mask based on the slope of the
underlying surface and the decal direction
52
52

Future improvements
53

Future improvements
54
54
● Better tooling for bulk decalling
○ Brush decals with randomized parameters
We already have that for small geometry instances like plants, debris, etc.!
○ Procedural placement based on world parameters
(e.g. biomes, material types, indoor/outdoor)
○ Physically simulated placement
(e.g. rigid body / particle simulation)
● Extend system to go full bindless
○ Bindless texturing also for scene rendering
○ Prototype exists but not yet ready for production

Thank you for listening!
@philiphammer0
phammer@deck13.com
55
Also a big thanks to the amazing DECK13 team which helped
with contributions, reviews and suggestions!

We’re hiring! :)
@deck13_de
https://www.deck13.com/jobs/
56

Q & A
57

References
[Persson] “Practical Clustered Shading”, Emil Persson,
http://www.humus.name/Articles/PracticalClusteredShading.pdf
[Drobot17] “Improved Culling for Tiled and Clustered Rendering”, Michal Drobot, Siggraph 2017
http://advances.realtimerendering.com/s2017/
[Hammer18] “Dissecting the Rendering of The Surge”, Philip Hammer, Quo Vadis Berlin 2018
https://de.slideshare.net/philiphammer/dissecting-the-rendering-of-the-surge
[Xu16] "Temporal Antialiasing in Uncharted 4", Ke Xu, Siggraph 2016
http://advances.realtimerendering.com/s2016/
[Lagarde14] “Moving Frostbite to physical based rendering”, Sébastien Lagarde, Siggraph 2014
https://seblagarde.ﬁles.wordpress.com/2015/07/course_notes_moving_frostbite_to_pbr_v32.pdf
[Wronski15] “Fixing screen-space deferred decals”, Bart Wronski, 2015
https://bartwronski.com/2015/03/12/ﬁxing-screen-space-deferred-decals/
[Sawicki15] "Watch out for non-uniform resource index!", Adam Sawicki,
http://www.asawicki.info/news_1608_direct3d_12_-_watch_out_for_non-uniform_resource_index.html
[Sousa16] "The Devil is in the Details: idTech 666", Tiago Sousa, Siggraph 2016
https://de.slideshare.net/TiagoAlexSousa/siggraph2016-the-devil-is-in-the-details-idtech-666
[Pettineo16] “Bindless Texturing for Deferred Rendering and Decals”, Matt Pettineo
https://mynameismjp.wordpress.com/2016/03/25/bindless-texturing-for-deferred-rendering-and-decals/
[Reed14] “Deferred Texturing”
http://www.reedbeta.com/blog/deferred-texturing/
[LotF14] “Building your own engine, part 3 – Visual effects for the next generation”, Benjamin Glatzel, David Reinig
https://www.makinggames.biz/news/building-your-own-engine-part-3-visual-effects-for-the-next-generation,6330.html
[Hill12] “Blending in Detail”, Stephen Hill, 2012
https://blog.selfshadow.com/publications/blending-in-detail/
[Everitt14] “Approaching zero driver overhead”, Cass Everitt et. al.,, Siggraph 2014
https://de.slideshare.net/CassEveritt/approaching-zero-driver-overhead
[GpuInfo] “Vulkan Hardware Database”, Sascha Willems
https://vulkan.gpuinfo.org/listextensions.php
58
58

Bindless Deferred Decals in The Surge 2

More Related Content

What's hot (20)

Similar to Bindless Deferred Decals in The Surge 2 (20)

Recently uploaded (20)

Bindless Deferred Decals in The Surge 2