GSC Game World‘s
S.T.A.L.K.E.R : Clear
Sky – a showcase for
Direct3D 10.0/1
Speakers:
Igor A. Lobanchikov – Former Lead Gfx Engineer at GSC
Holger Gruen - ISV Engineer AMD GPG
Agenda
Introduction
The X-Ray rendering architecture
Notable special effects
MSAA deferred rendering
10.0/10.1
» G-buffer optimization
» Direct3D 10.1 accelerated effects
» Q&A
»
»
»
»
Introduction
» Jon Peddie mentions Stalker : Clear Sky as one of
his two top games of 08!
»
JON PEDDIE’S TECH WATCH • Volume 9, NUMBER 1
» The first Direct3D 10.0/1 game to be released
with a deferred MSAA renderer
» Contains several Direct3D 10.1 rendering paths
»
»
MSAA alpha test, accelerated sunshaft and shadows
Direct3D 10.1 used for quick prototyping of the
MSSA renderer
» This talk walks you through the Direct3D 10.0/1
and other optimizations done in a joint effort
between GSC and AMD
The X-Ray rendering
architecture
» Rendering stages – list
» G-stage
» Light stage
» Light combine
» Transparent objects
» Bloom/exposition
» Final combine-2
» Post-effects
The X-Ray rendering
architecture: stages
» G-stage
» Output geometry attributes (albedo, specular,
position, normal, ambient occlusion, material).
» Light stage
» Calculate lighting ( diffuse light-RGB, specular
light – intensity only)
» Interleaved rendering with shadowmap
» Draw emissive objects
The X-Ray rendering
architecture: stages
» Light combine
» Deferred lighting is applied here
» Hemisphere lighting is calculated here (both
using OA light-map and SSAO)
» Perform tone-mapping here
» Output Hi and Lo part of tone-mapped image
into 2 RTs
The X-Ray rendering
architecture: stages
» Transparent objects
» Basic forward rendering
» Bloom/exposition
» Use Hi RT as a source for bloom/luminance
estimation
» Final combine-2
» Apply DOF, distortion, bloom here
» Post-effects
» Apply black-outs, film-grain, etc..
Dynamic rain
» Prepare shadowmap as seen along
the direction of rain
» Visible pixels are considered wet
» Apply postprocess to G-Buffer
» Make albedo darker and specular higher
» Fix-up normal
» That's all
Dynamic rain: normal
fix-up
» Horizontal surfaces
» Use tiled volume texture
to animate puddle rings
» Vertical surfaces
» Scroll texture with the
water stream vertically
» All normals are
treated as worldspace ones
Dynamic rain: G-buffer
modification
Dynamic
rain
disabled
Dynamic
rain
enabled
Normal visualization
Combined image
Dynamic rain:
shadowmap
» Use shadowmap to mask pixels
invisible to the rain
» Draw only static geometry
» Snap shadowmap texels to world space
» Use jittering to hide shadowmap aliasing and
simulate wet/dry area border.
Dynamic rain:
shadowmap
4-tap shadowmap
Jittered shadowmap
Dynamic rain:
what’s next?
» Use material ID
» Use more directions for gradient
detection
» Puddle map
» Project additional puddle textures on the
ground for artist-defined behavior
» Use reprojection cache?
» Storing rain shadowmap history from the
previous frame could allow us to use dynamic
objects as rain occluders
Sun Shafts
» Just do ray-marching
» Shadowmap test needs to
be carried out on every
step
» Jitter ray length and
use PCF to hide
banding artifacts
» Use lower single
sample intensity to
hide noise
Sun Shafts performance
considerations
» High sampling shadowmap
coherency due to the high
coherency of positions in G-buffer
(breaks for A-tested geometry)
» Even higher sampling coherency
for dueling frustum case
» Fixed number of steps eliminates
dynamic branching which helps in
low coherency edge cases
Sun Shafts: Cascaded
Shadow Map case
» Just use single cascade for the
whole ray
» Simpler algorithm
» Lower resolution shadowmap reduces banding
for longer rays
» Visible border between cascades
Sun Shafts
MSAA deferred rendering
10.0/10.1
» Deferred MSAA Rendering under DX10
» Main concept
» Stages affected by MSAA rendering
» Easy prototyping with Direct3D 10.1
» DX10 A2C
MSAA deferred rendering
10.0/10.1 main concept
» Render to MSAA G-buffer
» Mask edge pixels
» Process only subsample #0 for plain
pixels. Output to all subsamples
» Process each subsample for edge pixels
independently
MSAA deferred rendering:
MSAA output
» G-stage (subsample geometry data)
» Light stage (subsample lighting)
» Light combine (subsample data
combination)
» Transparent objects
» Bloom/exposition
» Final combine-2
» Post-effects
MSAA deferred rendering:
read from MSAA source
» G-stage
» Light stage (uses G-stage data)
» Light combine (uses G-stage and
light stage data)
» Transparent objects
» Bloom/exposition
» Final combine-2
» Post-effects
MSAA deferred rendering:
MSAA in/out stages
» For each shader
» Plain pixel – run shader at pixel
frequency
» Edge pixel – run at subpixel
frequency
» Early stencil hardware minimizes PS
overhead
MSAA deferred rendering:
MSAA in/out stages
plain pixel pass
Shader
edge pixel pass
Shader
Shader
Shader
Shader
MSAA deferred rendering
» DX10.0 doesn‘t support running
shader at subsample frequency
(DX10.1 does)
» Use DX10.1 for fast prototyping
» For DX10.0 use separate pass
for each subsample: shaders specifies
subsample to read at compile time,
use output mask to allow writing to a
single subsample
MSAA deferred rendering:
DX10.0
plain pixel pass
Shader
edge pixel #0 pass
edge pixel #2 pass
Shader
Shader
edge pixel #1 pass
edge pixel #3 pass
Shader
Shader
MSAA deferred rendering:
DX10.0 A2C
» Alpha-tested geometry can‘t be
MSAA‘d using common technique
» Use A2C to approximate anti-aliasing
» Alpha-channel of all G-Buffers store
geometry attributes: need 2-pass
algorithm:
» Write only depth using A2C
» Write geometry data using Z-equal
G-Buffer Optimization - 1
» Stalker originally used a 3-RT G-buffer
» 3d Pos + materialID => RGBA16F RT0
» Normal + Ambient occl. => RGBA16F RT1
» Color + Gloss => RGBA8 RT2
» At high resolutions/msaa-settings the size of the
G-buffer becomes the bottleneck
» Joint effort optimization effort lead to a 2-RT GBuffer
» Normal+Depth+matID+AO => RGBA16F RT0
» Color + Gloss => RGBA8 RT1
» Trade packing math vs. less G-Buffer texture ops
» Reduces G-Buffer size from 160 to 96 bits per pixel
G-Buffer Optimization - 2
» Reconstruct 3D position from depth
// input SV_POSITION as pos2d
New_pos2d = ( (pos2d.xy) * (2/screenres.xy) )– float2(1,1);
viewSpacePos.x = gbuffer_depth * tan( 90-HORZFOV/2 ) * New_pos2d.x;
viewSpacePos.y = -gbuffer_depth * tan( 90-VERTFOV/2 ) * New_pos2d.y;
viewSpacePos.z = gbuffer_depth;
» Normals get un-/packed from 2D <-> 3D
Packing
Unpacking
float2 pack_normal( float3 norm )
float3 unpack_normal(float2 norm)
{
{
float2 res;
float3 res;
res = 0.5*( norm.xy + float2( 1, 1 ) ) ;
res.xy=( 2.0 * abs(norm) ) – float2(1,1);
res.x *= ( norm.z<0 ? -1.0 : 1.0 );
res.z =(norm.x < 0? -1.0:1.0) *
sqrt(abs(1 – res.x*res.x - res.y*res.y));
return res;
return res;
}
}
G-Buffer Optimization - 2
» Pack AO and matID into the usable bits of the last
16bit fp channel of RT0
»
Pack data into a 32bit uint as a bit pattern that is a valid
16bit fp number
»
Cast the uint to float using asfloat()
»
Cast back for unpacking using asuint()
»
Extract bits
Direct3D 10.1 accelerated
effects - Agenda
» MSAA Alpha test
»
A brief recap
» Shader based A2C
»
Why would you want to do this in a shader?
» Non-hierarchical min-max shadow maps
»
Hybrid plane based/min-max solution
Direct3D 10.1 accelerated
effects – MSAA Alpha Test
» Sample texture once for each
MSAA sub-sample
»
»
Sample locations standardized in
Direct3D 10.1
ddx/ddy used to calculate UV
coordinates for sample locations
» Set SV_COVERAGE for samples passing the AT
» Higher image quality than Direct3D 10.0 A2C!
» One rendering pass only in Stalker
»
A2C needs two passes in Stalker under Direct3D 10.0
»
Good for CPU-limited situations
» More texture-heavy than Direct3D 10.0 A2C
»
Especially at high AA settings (e.g. 8xMSAA)
Direct3D 10.1 accelerated
effects – Shader A2C
» Why would you want to do this?
»
»
MSAA Alpha test slower than A2C at high
(msaa) settings
Control over SV_COVERAGE allows one-pass
shader-based A2C in Stalker
»
»
»
Direct3D 10.0 A2C needs two passes in Stalker
Shader-based A2C only needs to look at one
texture sample
Admittedly lower quality than MSAA AT but
sometimes speed is all you care about
Direct3D 10.1 accelerated
effects – Shader A2C cont.
We tried two methods to implement this
Method 1 at 4xMSAA
15
1 2
5
5
2 3
5
5
3 4
5
5
4 5
5
5
0 bits in SV_COVERAGE set
1 bit in SV_COVERAGE set
2 bits in SV_COVERAGE set
3 bits in SV_COVERAGE set
4 bits in SV_COVERAGE set
SV_POSITION.x+
SV_POSITION.y
used to select bit
pattern
Direct3D 10.1 accelerated
effects – Shader A2C cont.
Method 2 at 4xMSAA
15
SV_POSITION.x+
SV_POSITION.y
used to perturb
alpha
1 2
5
5
2 3
5
5
3 4
5
5
4 5
5
5
SV_COVERAGE = 0
SV_COVERAGE = 1
SV_COVERAGE = 3
SV_COVERAGE = 7
SV_COVERAGE = 15
This method got used in Stalker –
it is simply cheaper!
Direct3D 10.1 accelerated
effects – Shader A2C cont.
No obvious difference in IQ expected –
only a zoom-in shows a difference
Direct3D 10.0 A2C
Direct3D 10.1 shader
based A2C
Direct3D 10.1 accelerated
effects – min/max SM Recap
» Min-Max Shadows Maps introduced at GDC 05 by
K. Dmitriev & Y. Uralsky
» Key idea: Build MIP-chain from a shadow map
»
Store min/max depth of 4 texels in next MIP down
» Allows hierarchical rejection of sub-blocks of a
shadow filter kernel
»
Traverse MIPs and check for overlap of shadow
filter kernel quad with current min-max SM quad
»
If center.z >= max Z => full shadow
»
Else if center.z < min Z => full light
»
Can accelerate large filter kernels
Direct3D 10.1 acc. effects –
non-hierarchical min/max SMs
» Reduce NxN block of SM texels to one minmax texel
MinMax-Z
texel
Full light
minZ
expensive
test
Full
shadow
maxZ
SM texels
Direct3D 10.1 acc. effects –
non-hierarchical min/max SMs
» Can still be used to reject sub-blocks of a
shadow filter kernel
MinMax-Z
texel
Full light
minZ
expensive
test
Full
shadow
maxZ
SM texels
Direct3D 10.1 acc. effects –
non-hierarchical min-max SMs
» Same test logic as hierachical min-max
SMs
MinMax-Z
texel
Full light
minZ
expensive
test
Full
shadow
maxZ
SM texels
Direct3D 10.1 acc. effects –
non-hierarchical min/max SMs
» Higher chance for one-sample quick
rejection test than hierachical min-max SM
MinMax-Z
texel
Full light
minZ
expensive
test
Full
shadow
maxZ
SM texels
Direct3D 10.1 acc. effects –
min/max SM construction
Direct3D 10.1 accelerates min-max SM
construction – e.g. for a 4x4 to 1 reduction
Direct3D 10.0
NxN (4x4) = 16 point
samples if one wants to
find min-max depth
Direct3D 10.1
w
z
w
z
x
y
x
y
w
z
w
z
x
y
x
y
(N/2)x(N/2) = 4 Gather()
samples get all data
Direct3D 10.1 acc. effects –
non-hier. min/max SMs cont.
Things to consider when using the min-max SM ..
Let’s consider a
shadow map
And its min-max SM
Direct3D 10.1 acc. effects –
non-hier. min-max SMs cont.
Things to consider when using the min-max SM ..
A shadow mapping filter
kernel can overlap four
min/max SM texels
It is necessary to sample
all min/max texels that
are touched by the
kernel
Direct3D 10.1 acc. effects –
non-hier. min/max SMs cont.
Things to consider when using the min/max SM ..
Instead one can just
have overlapping minmax construction
kernels
And use only one
sample !
Stalker uses an overlapping filter kernel big enough
to allow quick rejection of sunshaft shadow map
probes and uses Gather() to accelerate min/max SM
lookups for big shadow filter kernels
Direct3D 10.1 acc. effects –
non-hierarchical min/max SMs
Ground planes or
other big planar
features
create problems!
One texel in
the min-max
SM
» All pixels within the projected area are
between minZ/maxZ
»
All go down the expensive path
» Only way around this is a high depth bias
with all its problems
Direct3D 10.1 acc. effects –
non-hier. min/max SMs cont.
z
z0
» Try to fit a plane
through all depth
samples in the block
to convert to min/max
z
y
x
» Store plane equation as RT0:-z0,
RT1:(dz/dx,dz/dy) instead of RT0:maxZ, RT1:
minZ, 0 )
» Shadow shader uses sub-texel offsets to compute
depth D using stored plane equation
» This solves the issues with planar features!
»
Safe to assume that the whole filter kernel is in front of or
behind the plane
Direct3D 10.1 acc. effects –
non-hier. min/max SMs cont.
» Why use min-max SMs in Stalker ?
»
Allows for a higher shadow quality at
high FPS
»
»
»
Rejects most pixels as fully lit or fully shadowed
Expensive 8x8 shadow filter only for pixels that
need them
Min-Max SMs accelerate sunshaft
rendering
»
»
Up to 4 PCF lookups into the shadow map per
step per ray
Uses min-max SM to skip these lookups if
possible
Q&A
[email protected]