Real-time Atmospheric
Effects in Games Revisited
Carsten Wenzel
The deal
! Follow up to a talk I gave at SIGGRAPH
2006
! Covers material presented at the time
plus recent additions and improvements
Overview
! Introduction
! Scene depth based rendering
! Atmospheric effects breakdown
! Sky light rendering
! Fog approaches
! Soft particles
! Cloud rendering (updated/new)
! Volumetric lightning approximation
! River and Ocean rendering (updated/new)
! Scene depth based rendering and MSAA (new)
! Conclusions
Introduction
! Atmospheric effects are important cues
of realism (especially outdoors)
! Why…
! Create sense of depth
! Increase level of immersion
Motivation
! Atmospheric effects are mathematically
complex (so far usually coarsely
approximated if any)
! Programmability and power of today’s
GPUs allow implementation of
sophisticated models
! How to can these be mapped efficiently?
Related Work
! Deferred Shading (Hargreaves 2004)
! Atmospheric Scattering (Nishita et al
1993)
! Cloud Rendering (Wang 2003)
! Real-time Atmospheric Effects in Games
(Wenzel 2006)
Scene Depth Based Rendering:
Motivation
! Many atmospheric effects require accessing
scene depth
! Similar to Deferred Shading [Hargreaves04]
! Mixes well with traditional style rendering
! Deferred shading is not a must!
! Think of it as writing a pixel shader with scene
depth available
! Requires laying out scene depth first and
making it available to following rendering passes
Scene Depth Based Rendering:
Benefits
! Decouple rendering of opaque scene geometry
and application of other effects
! Atmospheric effects
! Post-processing
! More
! Apply complex models while keeping the
shading cost moderate
! Features are implemented in separate shaders
! Helps avoiding hardware shader limits (can support
older HW)
Scene Depth Based Rendering:
Challenges
! Alpha-transparent objects
! Only one color / depth value stored
! However, per-pixel overdraw due to alpha
transparent objects potentially unbound
! Workaround for specific effects needed (will be
mentioned later)
Scene Depth Based Rendering:
API and Hardware Challenges
! Usually cannot directly bind Z-Buffer and
reverse map
! Write linear eye-space depth to texture
instead
! Float format vs. RGBA8
! Supporting Multi-Sample Anti-Aliasing is
tricky (more on that later)
Recovering World Space Position
from Depth
! Many deferred shading implementations
transform a pixel’s homogenous clip space
coordinate back into world space
! 3 dp4 or mul/mad instructions
! There’s often a simpler / cheaper way
! For full screen effects have the distance from the
camera’s position to its four corner points at the far
clipping plane interpolated
! Scale the pixel’s normalized linear eye space depth by
the interpolated distance and add the camera position
(one mad instruction)
Sky Light Rendering
! Mixed CPU / GPU implementation of
[Nishita93]
! Goal: Best quality possible at reasonable
runtime cost
! Trading in flexibility of camera movement
! Assumptions and constraints:
! Camera is always on the ground
! Sky infinitely far away around camera
! Win: Sky update is view-independent, update
only over time
Sky Light Rendering: CPU
! Solve Mie / Rayleigh in-scattering integral
! For 128x64 sample points on the sky
hemisphere solve…
Pb
⎛ H− h (−t ( PP ,λ )−t ( PP ,λ )) ⎞
I v (λ ) = I s (λ )K (λ )F (θ , g ) ∫ ⎜ e 0 e c a ⎟ (1)
⎜
Pa ⎝
⎟
⎠
! Using the current time of day, sunlight direction,
Mie / Rayleigh scattering coefficients
! Store the result in a floating point texture
! Distribute computation over several frames
! Each update takes several seconds to compute
Sky Light Rendering: GPU
! Map float texture onto sky dome
! Problem: low-res texture produces blocky results
even when filtered
! Solution: Move application of phase function to GPU (F(θ,g)
in Eq.1)
! High frequency details (sun spot) now computed per-pixel
! SM3.0/4.0 could solve Eq.1 via pixel shader and
render to texture
! Integral is a loop of ~200 asm instructions iterating 32 times
! Final execution ~6400 instructions to compute in-scattering
for each sample point on the sky hemisphere
Global Volumetric Fog
! Nishita’s model still too expensive to
model fog/aerial perspective
! Want to provide an atmosphere model
! To apply its effects on arbitrary objects in the
scene
! Developed a simpler method to compute
height/distance based fog with
exponential fall-off
Global Volumetric Fog
f (( x , y , z ) ) = be − cz
T
v v v
v (t ) = o + t d
∫ f ((o ) v
+ td x , o y + td y , o z + td z )
1
v
∫ f (v (t ))dt =
T
x d dt
0
− co z ⎡ 1 − e − cd z ⎤
= be dx + dy + dz
2 2 2
⎢ ⎥
⎣ cd z ⎦
v
F (v (t )) = e
v
− ∫ f (v (t ))dt (2)
f – fog density distribution b – global density
c – height fall-off F – fog density along v
v – view ray from camera (o) to target pos (o+d), t=1
Global Volumetric Fog:
Shader Implementation
Eq.2 translated into HLSL…
float ComputeVolumetricFog( in float3 cameraToWorldPos )
{
// NOTE: cVolFogHeightDensityAtViewer = exp( -cHeightFalloff *
cViewPos.z );
float fogInt = length( cameraToWorldPos ) *
cVolFogHeightDensityAtViewer;
const float cSlopeThreshold = 0.01;
if( abs( cameraToWorldPos.z ) > cSlopeThreshold )
{
float t = cHeightFalloff * cameraToWorldPos.z;
fogInt *= ( 1.0 - exp( -t ) ) / t;
}
return exp( -cGlobalDensity * fogInt );
}
Combining Sky Light and Fog
! Sky is rendered along with scene
geometry
! To apply fog…
! Draw a full screen quad
! Reconstruct each pixel’s world space position
! Pass position to volumetric fog formula to
retrieve fog density along view ray
! What about fog color?
Combining Sky Light and Fog
! Fog color
! Average in-scattering samples along the
horizon while building texture
! Combine with per-pixel result of phase
function to yield approximate fog color
! Use fog color and density to blend
against back buffer
Combining Sky Light and Fog:
Results
*
Fog Volumes
! Fog volumes via ray-tracing in the shader
! Currently two primitives supported: Box,
Ellipsoid
! Generalized form of Global Volumetric Fog
! Exhibits same properties (additionally, direction of
height no longer restricted to world space up vector,
gradient can be shifted along height dir)
! Ray-trace in object space: Unit box, unit sphere
! Transform results back to solve fog integral
! Render bounding hull geometry
! Front faces if outside, otherwise back faces
! For each pixel…
! Determine start and end point of view ray to plug into
Eq.2
Fog Volumes
! Start point
! Either camera pos (if viewer is inside) or ray’s
entry point into fog volume (if viewer is outside)
! End point
! Either ray’s exit point out of the fog volume or
world space position of pixel depending which
one of the two is closer to the camera
! Render fog volumes back to front
! Solve fog integral and blend with back
buffer
Fog Volumes
Rendering of fog volumes: Box (top left/right), Ellipsoid (bottom left/right)
Fog and Alpha-Transparent
Objects
! Shading of actual object and application of
atmospheric effect can no longer be
decoupled
! Need to solve both and combine results in same pass
! Global Volumetric Fog
! Approximate per vertex
! Computation is purely math op based (no lookup
textures required)
! Maps well to older HW…
! Shader Models 2.x
! Shader Model 3.0 for performance reasons / due to lack
of vertex texture fetch (IHV specific)
Fog and Alpha-Transparent
Objects
! Fog Volumes
! Approximate per object, computed on CPU
! Sounds awful but it’s possible when
designers know limitation and how to work
around it
! Alpha-Transparent objects shouldn’t become too
big, fog gradient should be rather soft
! Compute weighted contribution by
processing all affecting of fog volumes back
to front w.r.t camera
Soft Particles
! Simple idea
! Instead of rendering a particle as a regular billboard,
treat it as a camera aligned volume
! Use per-pixel depth to compute view ray’s travel
distance through volume and use the result to fade out
the particle
! Hides jaggies at intersections with other geometry
! Some recent publications use a similar idea and treat
particles as spherical volumes
! We found a volume box to be sufficient (saves shader
instructions; important as particles are fill-rate hungry)
! GS can setup interpolators so point sprites are finally
feasible
Soft Particles: Results
Comparisons shots of particle rendering with soft particles
disabled (left) and enabled (right) *
Clouds Rendering Using Per-
Pixel Depth
! Follow approach similar to [Wang03],
Gradient-based lighting
! Use scene depth for soft clipping (e.g. rain
clouds around mountains) – similar to Soft
Particles
! Added rim lighting based on cloud density
Cloud Shadows
! Cloud shadows are cast in
a single full screen pass
! Use depth to transform
pixel position into shadow
map space
Distance Clouds
! Dynamic sky and pre-baked sky box clouds don’t mix
well
! Real 3D cloud imposters can be expensive and are
often not needed
! Limited to 2D planes above the camera clouds can be
rendered with volumetric properties
! Sample a 2D texture (cloud density) along the view dir
! For each sample point sample along the direction to sun
! Adjust number of samples along both directions to fit
into shader limits, save fill-rate, etc.
Distance Clouds
! Use the accumulated density to calc attenuation
and factor in current sun / sky light
Distance Clouds at different times of day *
Volumetric Lightning Using Per-
Pixel Depth
! Similar to Global Volumetric Fog
! Light is emitted from a point falling off radially
! Need to carefully select attenuation
function to be able to integrate it in a
closed form
! Can apply this lighting model just like
global volumetric fog
! Render a full screen pass
Volumetric Lightning Model
(
f ( x, y , z ) =
T
) v
i
1 + a ⋅ l − ( x, y , z )
T 2
v v v
v (t ) = o + td
( ) v
f (v (t ))dt = ∫ f (ox + td x , o y + td y , oz + td z ) d dt
1
v
∫
T
⎡ ⎛ v + 2w ⎞ ⎛ v ⎞⎤
⎢ arctan⎜⎜ ⎟ − arctan ⎜ ⎟⎥
⎟ ⎜ ⎟
2⎢ ⎝ 4uw − v 2 ⎠ ⎝ 4uw − v 2 ⎠⎥
= 2i d x + d y + d z
2 2
⎢ 4uw − v 2 ⎥
⎢ ⎥
⎢⎣ ⎥⎦
v (3)
= F (v (t ))
f – light attenuation function i – source light intensity
l – lightning source pos a – global attenuation control value
v – view ray from camera (o) to target pos (o+d), t=1
F – amount of light gathered along v
Volumetric Lightning Using Per-
Pixel Depth: Results
*
River shading
! Rivers (and water areas in general)
! Special fog volume type: Plane
! Under water fog rendered as described earlier
(using a simpler uniform density fog model
though)
! Shader for water surface enhanced to softly
blend out at riverside (difference between
pixel depth of water surface and previously
stored scene depth)
River shading: Results
River shading –
Screens taken from a hidden section of the E3 2006 demo *
Ocean shading
! Very similar to river shading, however…
! Underwater part uses more complex
model for light attenuation and in-
scattering
! Assume horizontal water plane, uniform
density distribution and light always
falling in top down
! Can be described as follows…
Ocean shading
( )
f ( x, y, z ) , t = e −c ( s − z )e −ct = e −c ( s − z +t )
T
v v v
v (t ) = o + td
v v v
p−o
d= v v
p−o
v v
l = p−o
v
attenuation = f ( p, l ) = e − c ( s − p z +l )
v
l
v
inscatter = ∫ f (v (t ), t )dt = ∫ e − c ( s − (o z + td z )+ t )
d dt
0
l
−c ( s −oz ) − ct (1− d z )
=e ∫e
0
dt
⎡ e (d z −1)ct ⎤
l
−c ( s −oz )
=e ⎢ ⎥
⎣ (d z − 1 ) c ⎦0
−c ( s −oz ) ⎡ e (d z −1)cl − 1⎤
=e ⎢ ⎥
⎣ (d z − 1)c ⎦
finalCol = attenuation ⋅ sceneCol + inscatter ⋅ envlightCol (4)
Ocean shading: Results
Underwater view: from ground up (1st row), from
underneath the surface down (2nd row). Same lighting
settings apply. Higher density on the right column. *
Scene depth based rendering and
MSAA
! Several problems
! Cannot bind multi-sampled RT as texture
! Shading per pixel and not per sample
! Need to resolve depth RT which
produces wrong values at silhouettes Î
potentially causes outlines in later
shading steps
! Two problems we ran into
! Fog
! River / Ocean
Scene depth based rendering and
MSAA: Fog
! Fog color doesn’t changed drastically for
neighboring pixel while density does
! Have fog density computed while laying out
depth (two channel RT)
! During volumetric fog full screen pass only
compute fog color and read density from
resolved RT
! Averaging density during resolve works
reasonably well compared to depth
Scene depth based rendering and
MSAA: River / Ocean
! Shader assumes dest depth > plane depth
(otherwise pixel would have be rejected by z-
test)
! With resolved depth RT this cannot be
guaranteed (depends on pixel coverage of
object silhouettes)
! Need to enforce original assumption by finding
max depth of current pixel and all neighbors
(direct neighbors suffice)
Scene depth based rendering and
MSAA: Results
Fog full screen pass with MSAA disabled (left) / enabled (right)
River / Ocean shading artifact (left) and fix (right)
Conclusion
! Depth Based Rendering offers lot’s of opportunities
! Demonstrated several ways of how it is used in
CryEngine2
! Integration issues (alpha-transparent geometry, MSAA)
Kualoa Ranch on Hawaii –
Real world photo (left), internal replica rendered with CryEngine2 (right)
References
! [Hargreaves04] Shawn Hargreaves, “Deferred
Shading,” Game Developers Conference, D3D
Tutorial Day, March, 2004.
! [Nishita93] Tomoyuki Nishita, et al., “Display of the
Earth Taking into Account Atmospheric Scattering,”
In Proceedings of SIGGRAPH 1993, pages 175-182.
! [Wang03] Niniane Wang, “Realistic and Fast Cloud
Rendering in Computer Games,” In Proceedings of
SIGGRAPH 2003.
! [Wenzel06] Carsten Wenzel, “Real-time Atmospheric
Effects in Games,” SIGGRAPH 2006.
Acknowledgements
! Crytek R&D / Crysis dev team
Questions
???
