Volumetric Global Illumination At Treyarch

Содержание

1. Volumetric Global Illumination At Treyarch
2. Volumetric Global IlluminationGI in volume textureLean texture dataIBL baked from probesConvex blend shapes
3. Presentation OrderWhere we startedEvolution along the wayWhere we ended up
4. Traditional Approach: Lightmaps Could be ok, but…
5. Lightmap Downsides Works poorly on detailed or intersecting geometry
6. Lightmap Downsides Doesn’t work at all on dynamic geometry
7. Lightmap Downsides Software ray-tracing and shading takes forever
8. Lightmap Downsides Results not visible in world editor
9. Process of InventionDeferred RendererReflections already presentSo how do we apply deferred GI?
10. Reflection Probes as Diffuse DataHigher Mips: convolved specular [DROBOT13]Lowest Mip: diffuse irradiance Real time IBL
11. Occlusion Is A Problem
12. Visibility Is A ProblemWhere the probe doesn’t seeLooks like shadows
13. Irradiance Volume[TATARCHUK05]
14. Render a Reflection Probe Per Voxel?138 Volumes
15. Collect Colors From Reflection ProbesRe-project cube mapsCombine to fill holes [BUEHLER01]
16. In Practice4096 rays per voxel15 neighbors consideredMissed rays are in-painted
17. Re-Project From Existing Probes
18. ReprojectionNeighbor candidates sorted based on distanceWhat about spec?
19. ReprojectionAngle and distance to surface defines a solid angle in the cube map
20. ReprojectionSample area validated against depth pyramidIf visible appropriate mip sampled
21. Reprojection Caluclation
22. Reprojection CalculationdistFromUnitCube = √( 1 + u2
23. Biggest BenefitHardware renderingRe-render to get bouncesOnly have to ray-trace and re-project once
24. Texture EncodingFlat Color?
25. Ambient / Highlight / Direction?Texture Encoding
26. Second Order Spherical Harmonic?Texture Encoding
27. Ambient Cube! [MCTAGGART04]BC6H CompressedTexture Encoding
28. Volume Texture Layout
29. Performance BenefitsOnly 3 samplesHardware trilinear filteringEvaluation: color[n]
30. Light Leaking Is A Problem
31. Common ApproachAdjust trilinear Based on normal [SILVENNOINEN15]Our approach needs to be more reliable
32. More Voxel DataPlanesSigned distance fieldBad artifacts
33. Solve With Shaping
34. Click To Size Boxes
35. Click To Add BoxesAuto-parent on creation
36. Voxels Near Walls
37. Consider Backfaces
38. Complex Room Shapes
39. Solution: Convex Shapes
40. Multiface VolumesClick to add and remove faces.
41. Multiface Volume EditingDrag / Cut / Slice / Rotate
42. Subtract ShapesCSG addThen subtract
43. Override VolumesLike priorityOnly two levels.
44. Runtime ImplementationCull against volume AABB’s to build
45. Example GI Volumestruct PlaneGroup { float4
46. Group Size?[6]+[6+6+…?[6]+[4+4+...?[4]+[4+4+…?[8]+[2+2+…?
47. Shader Exampleattenuation = 0;groupAtten = 1;for (
48. Why Not K-DOPs?KDOP – k-sided Discrete Oriented PolytopePairs of planes or slabsInstead of individual planesX
49. Runtime ImplementationSample three ambient cube values depending on normalBlend results between all volumes
50. Challenges
51. Problem: Geo Within Voxels
52. Solution: Smart Centers
53. Invalidate Near Geometry
54. Empty Space Skip
55. Problem: Seams
56. Solution: Volume Smoothing
57. Careful Lighting Artistry
58. Auto Volumes?“Do-Everything Button”
59. Volume Blending And DensityVolume Overdraw Per TileDebug Tools
60. Reflections
61. Reflection Planes[LAGARDE12]
62. Clever Artistry
63. Reflection Plane Parallaxfloat reflectionMip = ( 1
64. Parallax Fade Out
65. Reflection Brightness Correction[LAZAROV13]
66. Reflection Brightness Correction
67. Max specular brightness is just a guess
68. Pros:As good or better quality than light maps
69. Pros:Less than 2ms for reflections and GI
70. Pros:Works on all geometry
71. Pros:Less baking time with incremental baking
72. Pros:Baking is done in editor
73. Pros:Moving and changing GI
74. Pros:Loose connection between light and geo
75. Cons:Takes set up time
76. Cons:Training is hard
77. Cons:Either lower resolution or more memory use in game
78. Cons:Need beefy dev machines (48Gb RAM and 12Gb VRAM)
79. Cons:Development challenges
80. Special ThanksTreyarch:Dimitar Lazarov – Original IdeaKevin Myers
81. References[DROBOT13] DROBOT, M., 2013. Lighting Killzone: Shadow
82. Скачать презентанцию

Volumetric Global IlluminationGI in volume textureLean texture dataIBL baked from probesConvex blend shapes

Главная
Разное
Volumetric Global Illumination At Treyarch

Слайды и текст этой презентации

Слайд 1Volumetric Global Illumination At Treyarch
JT Hooker
Treyarch Senior Graphics Engineer

Слайд 2Volumetric Global Illumination
GI in volume texture
Lean texture data
IBL baked from

probes
Convex blend shapes

Слайд 3Presentation Order
Where we started

Evolution along the way

Where we ended up

Слайд 4Traditional Approach: Lightmaps
Could be ok, but…

Слайд 5Lightmap Downsides
Works poorly on detailed or intersecting geometry

Слайд 6Lightmap Downsides
Doesn’t work at all on dynamic geometry

Слайд 7Lightmap Downsides
Software ray-tracing and shading takes forever

Слайд 8Lightmap Downsides
Results not visible in world editor

Слайд 9Process of Invention
Deferred Renderer
Reflections already present
So how do we apply

deferred GI?

Слайд 10Reflection Probes as Diffuse Data
Higher Mips: convolved specular [DROBOT13]

Lowest Mip: diffuse irradiance
Real time IBL

Слайд 11Occlusion Is A Problem

Слайд 12Visibility Is A Problem
Where the probe doesn’t see
Looks like shadows

Слайд 13Irradiance Volume
[TATARCHUK05]

Слайд 14Render a Reflection Probe Per Voxel?
138 Volumes × 403 Voxels

× 6 Faces
÷ 60 FPS ÷ 60 Seconds
= 14,720 Minutes

(≈ 10 Days)

Render a Reflection Probe Per Voxel?138 Volumes × 403 Voxels × 6 Faces÷ 60 FPS ÷ 60

Слайд 15Collect Colors From Reflection Probes
Re-project cube maps
Combine to fill holes [BUEHLER01]

Слайд 16In Practice
4096 rays per voxel
15 neighbors considered
Missed rays are in-painted

Слайд 17Re-Project From Existing Probes

Слайд 18Reprojection
Neighbor candidates sorted based on distance

What about spec?

Слайд 19Reprojection
Angle and distance to surface defines a solid angle in

the cube map

Слайд 20Reprojection
Sample area validated against depth pyramid

If visible appropriate mip sampled

Слайд 21Reprojection Caluclation

Слайд 22Reprojection Calculation
distFromUnitCube = √( 1 + u2 + v2 );

// Compensation for cube-map shape.
angleOfVoxel = 4 * PI /

numSamples; // Solid angle from voxel.
inSqrt = 1 + distFromVoxel2 * angleOfVoxel * ( angleOfVoxel – 4PI ) / ( 4 * PI2 * distFromProbe2 );
angleOfProbe = 2PI * ( 1 – √inSqrt ); // Solid angle from reflection probe.
cubeRes = 1.0f / √( angleOfProbe * distFromUnitCube3 ); // Resolution needed for sample.
mipLevel = clamp( mipCount – log2( cubeRes ), 0, mipCount ); // Mip level to use.

return mipLevel;

Reprojection CalculationdistFromUnitCube = √( 1 + u2 + v2 ); // Compensation for cube-map shape.angleOfVoxel = 4

Слайд 23Biggest Benefit
Hardware rendering

Re-render to get bounces

Only have to ray-trace and

re-project once

Слайд 24Texture Encoding
Flat Color?

Слайд 25Ambient / Highlight / Direction?
Texture Encoding

Слайд 26Second Order Spherical Harmonic?
Texture Encoding

Слайд 27Ambient Cube! [MCTAGGART04]
BC6H Compressed
Texture Encoding

Слайд 28Volume Texture Layout

Слайд 29Performance Benefits
Only 3 samples

Hardware trilinear filtering

Evaluation: color[n] = normal2 ∙ float3(

Xsample[n], Ysample[n], Zsample[n] )
color = xVolume.SampleLevel( coord ) * normal.x

* normal.x +
yVolume.SampleLevel( coord ) * normal.y * normal.y +
zVolume.SampleLevel( coord ) * normal.z * normal.z;

Performance BenefitsOnly 3 samplesHardware trilinear filteringEvaluation: color[n] = normal2 ∙ float3( Xsample[n], Ysample[n], Zsample[n] )color = xVolume.SampleLevel(

Слайд 30Light Leaking Is A Problem

Слайд 31Common Approach
Adjust trilinear Based on normal [SILVENNOINEN15]

Our approach needs to be more

reliable

Слайд 32More Voxel Data
Planes

Signed distance field

Bad artifacts

Слайд 33Solve With Shaping

Слайд 34Click To Size Boxes

Слайд 35Click To Add Boxes
Auto-parent on creation

Слайд 36Voxels Near Walls

Слайд 37Consider Backfaces

Слайд 38Complex Room Shapes

Слайд 39Solution: Convex Shapes

Слайд 40Multiface Volumes
Click to add and remove faces.

Слайд 41Multiface Volume Editing
Drag / Cut / Slice / Rotate

Слайд 42Subtract Shapes
CSG add
Then subtract

Слайд 43Override Volumes
Like priority
Only two levels.

Слайд 44Runtime Implementation
Cull against volume AABB’s to build a list of

volumes
Per pixel calculate attenuation on visible volumes
Convex hull CSG
Groups of

six planes either extended, combined or subtracted

Runtime ImplementationCull against volume AABB’s to build a list of volumesPer pixel calculate attenuation on visible volumesConvex

Слайд 45Example GI Volume
struct PlaneGroup
{
float4 planes[6]; // Groups

of six planes.
bool subtractive; // Per group, specifies

whether it adds or subtracts.
bool finished; // Per group, whether it should be combined with the previous.
}
struct GIvolume
{
PlaneGroup *groups;
}

// Blends, or “feather”, are pre-multiplied into the plane definition.
planes[i].xyz = planeNormal;
planes[i].w = planeOffset;
planes[i] /= blendWidth; // Blend width is a scalar for how wide the blend is.

Example GI Volumestruct PlaneGroup { float4 planes[6]; // Groups of six planes. bool subtractive; //

Слайд 46Group Size?
[6]+[6+6+…?

[6]+[4+4+...?

[4]+[4+4+…?

[8]+[2+2+…?

Слайд 47Shader Example
attenuation = 0;
groupAtten = 1;
for ( int group =

0; group < numGroups; group++)
{
groupAtten *= saturate(

dot( planes[group][0].xyz, pos ) + planes[group][0].w );
groupAtten *= saturate( dot( planes[group][1].xyz, pos ) + planes[group][1].w );
groupAtten *= saturate( dot( planes[group][2].xyz, pos ) + planes[group][2].w );
groupAtten *= saturate( dot( planes[group][3].xyz, pos ) + planes[group][3].w );
groupAtten *= saturate( dot( planes[group][4].xyz, pos ) + planes[group][4].w );
groupAtten *= saturate( dot( planes[group][5].xyz, pos ) + planes[group][5].w );
if( finished[group] )
{
if( subtractive[group] )
attenuation = max( attenuation, groupAtten );
else
attenuation *= 1.0f - groupAtten;
groupAtten = 1;
}
}
return saturate( attenuation );

Shader Exampleattenuation = 0;groupAtten = 1;for ( int group = 0; group < numGroups; group++) {

Слайд 48Why Not K-DOPs?
KDOP – k-sided Discrete Oriented Polytope

Pairs of planes

or slabs
Instead of individual planes
X

Слайд 49Runtime Implementation
Sample three ambient cube values depending on normal
Blend results

between all volumes

Слайд 50Challenges

Слайд 51Problem: Geo Within Voxels

Слайд 52Solution: Smart Centers

Слайд 53Invalidate Near Geometry

Слайд 54Empty Space Skip

Слайд 55Problem: Seams

Слайд 56Solution: Volume Smoothing

Слайд 57Careful Lighting Artistry

Слайд 58Auto Volumes?
“Do-Everything Button”

Слайд 59Volume Blending And Density
Volume Overdraw Per Tile
Debug Tools

Слайд 60Reflections

Слайд 61Reflection Planes
[LAGARDE12]

Слайд 62Clever Artistry

Слайд 63Reflection Plane Parallax
float reflectionMip = ( 1 – gloss )

* numMips;

// as things get rougher “fade off” parallax correction
//

by pushing out intersection planes
float minDist = saturate( ( reflectionMip – 2.5 ) / ( numMips – 2.5 ) ) * 100;
distanceToPlane = max( abs( distanceToPlane ), minDist );

float intersectionDist = abs( distanceToPlane / -dot( direction, plane.xyz ) );

Reflection Plane Parallaxfloat reflectionMip = ( 1 – gloss ) * numMips;// as things get rougher “fade

Слайд 64Parallax Fade Out

Слайд 65Reflection Brightness Correction
[LAZAROV13]

Слайд 66Reflection Brightness Correction

Слайд 67Max specular brightness is just a guess based on angle

and gloss
The light mostly probably came from a certain direction

and area
This defines a cone of a certain size
Actual numbers determined experimentally

We know the diffuse light is a function of what’s in the specular map
Assume a maximum specular brightness that created this diffuse brightness
Adjust the sampled specular against the theoretical maximum

float maximumSpecValue = max3( 1.26816,
9.13681 * exp2( 6.85741 - 2 * mip ) * nDotV,
9.70809 * exp2( 7.085 - mip - 0.403181 * mip2) * nDotV );

float adjustedMaxSpec = diffuseGILum * maximumSpecValue;
float3 specLum = luminance( cubeMapSample );
float3 reflection = cubeMapSample *
adjustedMaxSpec / ( adjustedMaxSpec + speculum );

Brightness Correction

Max specular brightness is just a guess based on angle and glossThe light mostly probably came from

Слайд 68Pros:
As good or better quality than light maps

Слайд 69Pros:
Less than 2ms for reflections and GI

Слайд 70Pros:
Works on all geometry

Слайд 71Pros:
Less baking time with incremental baking

Слайд 72Pros:
Baking is done in editor

Слайд 73Pros:
Moving and changing GI

Слайд 74Pros:
Loose connection between light and geo

Слайд 75Cons:
Takes set up time

Слайд 76Cons:
Training is hard

Слайд 77Cons:
Either lower resolution or more memory use in game

Слайд 78Cons:
Need beefy dev machines (48Gb RAM and 12Gb VRAM)

Слайд 79Cons:
Development challenges

Слайд 80Special Thanks
Treyarch:
Dimitar Lazarov – Original Idea
Kevin Myers – Baking Code
Everyone

Else at Treyarch
Activision Central Tech:
Peter-Pike Sloan – Lots of Math
Josiah

Manson – Light Bake Features
Angelo Pesce – Reflection Solutions

Special ThanksTreyarch:Dimitar Lazarov – Original IdeaKevin Myers – Baking CodeEveryone Else at TreyarchActivision Central Tech:Peter-Pike Sloan –

Слайд 81References
[DROBOT13] DROBOT, M., 2013. Lighting Killzone: Shadow Fall, Digital Dragons

[TATARCHUCK05]

TATARCHUK, N., 2005. Irradiance Volumes for Games, GDC Europe

[BUEHLER01] BUEHLER,

C., BOSSE, M., MCMILLAN, L., GORTLER, S., COHEN, M., 2001. Unstructured Lumigraph Rendering, SIGGRAPH

[MCTAGGART04] MCTAGGART, G., 2004. Half-Life 2 / Valve Source Shading, Game Developers Conference

[SILVENNOINEN15] SILVENNOINEN, A., TIMONEN, V., 2015. Multi-Scale Global Illumination in Quantum Break, SIGGRAPH

[LAGARDE12] LAGARDE, S., ZANUTTINI, A., 2012. Local Image-based Lighting With Parallax-corrected Cubemaps, SIGGRAPH

[LAZAROV13] LAZAROV, D., 2013. Getting More Physical in Call of Duty: Black Ops II, SIGGRAPH

References[DROBOT13] DROBOT, M., 2013. Lighting Killzone: Shadow Fall, Digital Dragons[TATARCHUCK05] TATARCHUK, N., 2005. Irradiance Volumes for Games,

Скачать презентацию

Разделы презентаций

Volumetric Global Illumination At Treyarch

Содержание

Слайды и текст этой презентации

Слайд 1Volumetric Global Illumination At TreyarchJT HookerTreyarch Senior Graphics Engineer

Слайд 2Volumetric Global IlluminationGI in volume textureLean texture dataIBL baked from

probesConvex blend shapes

Слайд 3Presentation OrderWhere we startedEvolution along the wayWhere we ended up

Слайд 4Traditional Approach: Lightmaps Could be ok, but…

Слайд 5Lightmap Downsides Works poorly on detailed or intersecting geometry

Слайд 6Lightmap Downsides Doesn’t work at all on dynamic geometry

Слайд 7Lightmap Downsides Software ray-tracing and shading takes forever

Слайд 8Lightmap Downsides Results not visible in world editor

Слайд 9Process of InventionDeferred RendererReflections already presentSo how do we apply