Advanced Computer Graphics Bas Zalmstra Marries van de Hoef Tim de Jager
Ask questions
Who the.. are these guys?! Graphics teaching assistants
Who the.. are these guys?! Marries van de Hoef Experimental Game
Seven Dimensions PLAY NAO ->
Who the.. are these guys?! Bas Zalmstra Lead Abbey
Who the.. are these guys?! Tim de Jager Game Abbey
Reus
What are we talking about? Deferred shading variants Culling Variance Shadow mapping Subsurface scattering GPU Particle Physics
DEFERRED SHADING VARIANTS
Dynamic lighting – Calculating lighting real-time – Moving objects – Moving light sources
A lot of variants Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
FORWARD SHADING Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Forward shading Multiple light sources – Graphics practical 3 assignment E3 – Additive blending of light sources
Forward shading Lighting is calculated while rendering geometry – Calculated for the entire surface Two methods: Add multiple light sources in single shader – Variable number of light sources in a shader Render the object multiple times (additively) – Unnecessary vertex shader calculations
DEFERRED SHADING Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Deferred shading Render lighting per light source – Render the bounding volume
Deferred shading What do we need for (Lambertian) shading? – Object: Normals Position Surface color (albedo) – Light source: Position Lighting color Object data is not available when rendering light volumes
Deferred shading Store object properties in buffers (Texture2D’s)! – Everything has to be rendered once – Lighting uses data from buffers Geometry-Buffer (G-buffer): NormalsDepth (Position)Albedo
Deferred shading For each light source – Use data from the G-buffer – Light source properties as shader parameters NormalsDepth (Position)Albedo
DEFERRED LIGHTING Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Deferred lighting Deferred shading has limited object variability. – Defined by the properties in the G-Buffer. What if we want a different shader per object?
Deferred lighting (Also called Light pre-pass rendering) Only calculate lighting – Specular lighting, diffuse lighting – DON’T combine with diffuse/specular color yet
Deferred lighting Render all objects again For each object – Use lighting buffer as lighting information – We can use a different shader for each object!
Deferred lighting Recap: – Render each object to G-Buffer: – Use G-Buffer to calculate lighting: – Render each object again: Use lighting buffer for shading
Deferred lighting We also want different lighting per object Still not possible…
INFERRED LIGHTING Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Inferred lighting What about transparency – A buffer can store 1 value per pixel – Use stippling for the lighting Search for surface when fetching the lighting
Inferred lighting Lighting does not vary much over a surface – Lighting could be calculated at a lower resolution Also search for the correct surface when sampling from the lighting buffer
Inferred lighting At lower resolution – Render each object to G-Buffer: – Use G-Buffer to calculate lighting: Render each object again: – Use lighting buffer for shading
TILE BASED DEFERRED SHADING Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Tile based deferred shading Rendering light volumes is slow – A lot of overdraw Render lighting per tile – List of active light sources per tile Frustum culling (on GPU?)
FORWARD+ Forward Shading Deferred Shading Deferred Lighting Inferred Lighting Tile based deferred shading Forward+ Light indexed deferred shading
Forward+ We want different lighting per object Don’t calculate lighting, but store which light sources are active Store light indices per tile – Similar to tile based deferred rendering
Forward+ Light indexed deferred rendering – Store light indices for each pixel in a texture Limited number of light indices Forward+ – Published in 2012 – Store light indices in DX11 linked list structure
Usage in games Forward shading: Half-life 2, Portal, etc. Deferred shading: Killzone 2, Starcraft 2, etc. Deferred lighting: Most games Inferred lighting: Red faction / Saints Row Tile based def. sh.: Battlefield 3, Uncharted, …? Forward+ : Dirt Showdown (so far)
Questions?
CULLING
What is culling? You No culling
What is culling? You No culling
What is culling? You Frustum culling Cannot see that!
What is culling? You Frustum + Occlusion culling Cannot see that!
CULLING Frustum culling Occlusion culling
Frustum culling: Traditional approaches Culling a few thousand objects is slow – Need optimizations Create spatial hierarchy – Useful for quickly culling a lot of objects – Bounding Volume Hierarchy (BVH) Battlefield 1 and 2 use a Bounding Sphere Hierarchy
Bounding Sphere Hierarchy A B C A BC B+C A+B+C
Problems Hierarchical culling doesn’t scale well – Most computers have 4 or more cores Hard to handle dynamic objects Streaming in new parts of a level
Solution Just do it parallel brute force – Scales over all available cores – No overhead of tree – CPU can do a lot more low level optimizations Statistics (DICE): – 3 times faster – 1/5 of the original code – Seamlessly handles dynamic objects – Easy to optimize further
Grid
CULLING Frustum culling Occlusion culling
Oclusion culling: Traditional approaches Generate a static occlusion culling data structure Used in Quake1-3 Also used in newer engines like Unity3 Umbra (used in Unreal Engine 3)
Occlusion grid
Visible from within the yellow cell Not visible from within the yellow cell
Problems Doesn’t handle dynamic objects as occluders CANT BLOW UP THE EVERYTHING!!
Solution 1 Simply render the object and see if pixels made it on the screen – Low detail version Supported by todays hardware – Hardware Occlusion Queries Downside – The GPU is already busy enough with cool effects – CPU has to wait for the GPU to answer Stalls the CPU
Solution 2 Do the same thing but on the CPU – Moves the burden from the GPU to the CPU – Used in the Frostbite engine and the Cryengine Artists make occlusion geometry for occluders – Very large objects CPU rasterizes these occluders – 256x114 depth buffer – Parallel Test all objects in the frustum in parallel against the depth buffer
Software rasterizer
Recap Frustum culling – Selecting the objects in the viewing volume Occlusion culling – GPU: Hardware Occlusion Queries – CPU: Software Rasterizer References – Battlefield 3: Culling the Battlefield - Daniel Collin (DICE) – Practical Occlusion Culling on the PS3 - Will Vale
Questions?
VARIANCE SHADOW MAPPING
Regular shadow mapping Stage 1:
Regular shadow mapping Stage 2:
Shadow mapping problems Problem is: – Biasing Hard edges Depth
Shadow mapping problems Biasing: – Shadow acne
Improving shadow mapping We can’t filter the images
Improving shadow maps Percentage Close Filtering PCF Result Regular
Filtering We need to think of something So that we can filter...
Variance shadow maps Represent not as a single depth but as a distribution Use first and second moments to recover distribution: – M1 & M2
Algorithm 1. Store both depth and squared depth 2. Recover mean and variance using M1 & M2 3. Use Chebychev’s inequality to compute an upper bound 4. Use the upper bound to compute the lighting percentage
Expected value
First and Second Moment First and second moment can be recovered by: The mean and variance can be recovered by:
Chebychev!
Single occluder and surface
Computing the mean and variance: Where p represent the fraction that is onoccluded
Single occluder and surface Chebychev’s inequality returns the fraction that is unfiltered:
Single occluder and surface
Results
Problems Light bleeding: Solutions offered by: – – s3_ch08.html s3_ch08.html
Questions?
SUBSURFACE SCATTERING
Subsurface Scattering Simulate partial translucency
Subsurface Scattering Affected by volume and internal structure Makes some objects look more lively
Subsurface Scattering Light scatters diffusely through the material Very complex
Subsurface Scattering To use this in games: approximate (cheat) As long as it looks good
Subsurface Scattering By Colin Barré-Brisebois and Marc Bouchard from EA – Used in Battlefield 3
Subsurface Scattering Translucency depends on – The position of the light source – The position of the camera V · -L
Subsurface Scattering Translucency depends on – The position of the light source – The position of the camera – The thickness of the object Camera thickness Light
Subsurface Scattering Precompute local thickness
Subsurface Scattering Local thickness map – Add color to simulate internal color – Add detail to simulate internal structure
Subsurface Scattering
Can be incorrect with concave objects
Subsurface Scattering More information? Check
Questions?
GPU PARTICLE PHYSICS
GPU Particle Physics Particles colliding with the environment CPU processing not an option – Slow CPU-GPU communication – A lot of physics
GPU Particle Physics By Chris Tchou from Bungie – Used in Halo Reach Do everything on the GPU
GPU Particle Physics Particles are stored in buffers (Texture2D’s) Texture2D for: – Position – Velocity – Particle data (lifetime, appearance, lighting,...) One pixel in the Texture2D contains data for one particle
GPU Particle Physics Spawning new particles – CPU draws on the next pixels in the Texture2D – Set position, velocity, etc Old particles Add new particles here
GPU Particle Physics
GPU Particle Physics Collision detection Normal collision detection is not an option – For every particle, world geometry would have to be retrieved and checked... – Very inefficient
GPU Particle Physics Use deferred shading buffers instead! NormalsDepth (Position)
GPU Particle Physics Detect collision – Compare depth of particle with depth in buffer Depth (Position) Camera Particle Depthbuffer Depth to Particle Depth from buffer
GPU Particle Physics Detect collision – Compare depth of particle with depth in buffer Depth (Position) Camera Particle Depthbuffer Depth to Particle Depth from buffer
GPU Particle Physics Handle collision – Retrieve surface normal Camera Particle Normals Normal Velocity
GPU Particle Physics Handle collision – Retrieve surface normal Reflect velocity Apply damping Camera Particle Normals Normal Velocity
GPU Particle Physics
GPU Particle Physics Disadvantages: – Particles can only bounce against objects on screen – Collision errors can occur with high speed particles
Questions?
Thanks! Thanks for listening
Shout outs Wolfgang Hürst DGDARC for the juizz DGG for being awesome Vleeschtorpedo