Bending the Graphics Pipeline

Bending the Graphics Pipeline Johan Andersson DICE

Overview Give a taste of a few rendering techniques we are using & experimenting with how they interact, or would like to interact, with the graphics pipeline Tile-based Deferred Shading Morphological Antialiasing Analytical Ambient Occlusion 08/01/10 Beyond Programmable Shading, SIGGRAPH 2010

Tile-based deferred shading Tile-based culling & lighting Cull lights per screen-space tile Lighting kernel runs per tile Minimizes bandwidth/setup cost DX11: GPU compute shader Covered in the course last year [Andersson09] PS3: SPU jobs GPU renders gbuffer SPU does light culling & full lighting evaluation for each pixel 08/01/10 Beyond Programmable Shading, SIGGRAPH 2010

Standard phong Metallic Skin Translucent Multiple deferred lighting models Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Working with tiles Tile culling optimizations Cull lights & shadows with tile normal cone Detect tile specular=0 Detect tile lighting model Tile lighting kernel permutations Specular on/off Lighting models More in the future Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

SPU-based Deferred Shading Ported DX11 compute shader to SPU job Offloads PS3 GPU SPU processing in parallel with GPU rendering 32x16 pixel tiles Explicit SoA vectorization instead of implicit C/C++ on SPU - HLSL on GPU Not a problem for such a relative small kernel But not ideal data-parallel programming model Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

SPU vs GPU architecture 6 execution contexts vs 1+ million (each pixel) Explicit SIMD vs implicit SIMD C/C++ vs HLSL Explicit async DMA vs implicit latency hiding What can we learn? Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Issues & challenges going forward More lighting models SIMD & branching efficiency Transparent decal surfaces & volumes Fixed function blending doesn’t work well with deferred Higher-quality antialiasing Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Flexible lighting models Want both more & more flexible models: Custom gbuffer layout per material Quality & performance tradeoffs Examples: Hair / anisotropic materials Requires more lighting model parameters in gbuffer Foliage Massive overdraw with alpha-tested simple shaders, few parameters Write to as simple gbuffer as possible to reduce ROP/bandwidth bottleneck Skin Sub-surface scattering approximation Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

The SIMD efficiency problem Lighting models through dynamic branches GPU shader model can be problematic: Increased register pressure = overall slower shader Requires good screen-space SIMD coherency for performance win Potential solutions: Reshuffle pixels to improve coherency ? Within each tile, sort pixels by model, compute lighting & then scatter back GRAMPS -style queing? [Sugerman09] Attractive & powerful high-level programming model Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 Alpha-tested foliage has far from ideal coherency

Decals & deferred shading Decals blend selectively against gbuffer Include: Diffuse albedo (gbuffer1.rgb) Normal (gbuffer0.rgb) Want to include (but can’t in single pass): Specular albedo (gbuffer1.a) Specular smoothness (gbuffer0.a) Exclude: Material id (can’t blend) Object lighting (inherit from below surface) Fixed function blending doesn’t work well Pixel shader can’t write out both alpha & blend factor! Consoles doesn’t have blend mode per MRT Linear blend doesn’t work for all components Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 See Destruction Masking in Frostbite 2 using Volume Distance Fields [Kihl10] for more details about decal use case

Need programmable blending Benefits: Write out gbuffer alpha channels indepenently of blend factor Treat channels & targets however you see fit Non-linear blending & renormalizing blends Can do overlapping dependent blending Read current normal, add bumps relative to it, write out What approach? LRB-style pixel shader framebuffer read/modify/write [Lalonde09] Ideal general solution for developers How to hide synchronization latency? Implicit / explicit? Blend shader Yet another stage in a fixed pipeline No R/M/W, not ideal More? Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

The deferred shading + MSAA problem Huge storage & bandwidth requirements with deferred 1920 x 1080 x 5 x 4 x 4 = 165 MB Doesn’t scale! Adding 1 bit of precision = 2x more memory 4x MSAA is not enough Esp. for thin geometry in a distance Prohibitive performance and bandwidth in general with deferred shading But don’t miss Andrew Lauritzen’s talk later in the course: Deferred Rendering for Current and Future Rendering Pipelines There are alternatives to MSAA... Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

MLAA – Morphological Antialiasing Post-effect antialiasing Introduced in [Reshetov09] Implementations: Intel CPU reference implementation [Reshetov09] Sony PS3 SPU implementation [Perthuis10] GPU compute? [Biri10] Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

MLAA workings Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 From [Reshetov09]

MLAA comparisons (PS3) Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 No AA MLAA

MLAA takeaways Awesome AA for still pictures Moving pictures good, but: No sub-pixel information = edges snap to pixels Doesn’t solve aliasing on fine detail geometry Overall still a very good benefit! Focus/exclude effect based on framebuffer alpha & thresholds Unique requirements per game/app Not good to use on some UI, mark in alpha (or apply before) Variable post-effect , trade perf vs quality! Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

MLAA future (PC) GPU compute shader implementation Combine with MSAA & sub-pixel samples Simple MSAA box filter downsampling is a big waste Sort of similar to A Directionally Adaptive Edge Anti-Aliasing Filter [Yang09] A must to reduce the edge snapping of pure MLAA Not fully clear how it should work (sample distribution) Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Current dynamic AO Horizon-based Ambient Occlusion See [Bavoil09] for complete details Based on screen-space depth-buffer ( SSAO ) Very high quality sampling But only screen-space info is a big limitation Creates false occlusion artifacts Render in half-res for improved performance Bilateral upsampling + gaussian blur Can also do dual-resolution to reduce artifacts Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Horizon-based Ambient Occlusion Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 False occlusion halo from thin geometry

HBAO limitations False halo occlusion artifacts around small geometry Such as: fences & poles Extra visible when moving the camera Very noisy sampling for detailed zbuffers Common with alpha-tested foliage Difficult sampling problem Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Analytical Ambient Occlusion Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

HBAO vs AAO Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Analytical Ambient Occlusion Using Ambient Occlusion Volumes [McGuire10] Experimental implementation in Frostbite 2 With some good help from Morgan McGuire and Louis Bavoil Geometry-based technique Not screen-space! Say what? Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

AOV idea Extrude prism for each triangle (GS) Extrusion distance is where occlusion=0 Rasterize primitives in prism With depth-test enabled, near depth clip disabled Finds visible points inside volume Need to handle case with camera inside volume Accumulate analytical occlusion contribution for visible pixels (PS) Uses pixel normal & depth values from gbuffer Subtractive blend Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 HBAO

Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 HBAO AOV

AOV in practice Render geometry again in separate AO pass Uses depth & normal buffer from deferred rendering Half-res or lower with bilateral upsampling Culling should consider extrusion distance Separate paths for dynamic & rigid objects Can pre-compute rigid extruded AOV & reduce overdraw Doesn’t work with alpha-tested surfaces Simulate with per-surface or per-triangle approx. coverage factor Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Overdarkening (extra occlusion) Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Varying overdraw with AO distance Beyond Programmable Shading, SIGGRAPH 2010 08/01/10 0.1 m 0.2 m 0.5 m

AOV pros & cons Pros: Very high quality - close to raytracing ground truth Noise free (when full res) Perfectly stable with view changes Supports arbitrary dynamic polygon soups Cons: Requires massive fillrate Geometry cost Overdarkening, may require content tweaks Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

AOV future optimizations Reduce the massive overdraw Cull / restrict prisms that only extend out to empty air? Clamp screen-space prism size Not correct, but practical tradeoff. HBAO does this More optimal prism geometry GS is limited to triangle strip output  Precompute using quads for rigid objects Geometry LOD / mix with higher-order geometry representations Also see AO volume texture & analytical capsule techniques [Hill10] Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

AOV takeaways Major improvement in visual quality compared to SSAO Interesting use of geometry & rasterization pipelines Builds on existing HW-, SW- & content pipelines Quite simple brute force drop-in (but not as simple as SSAO ) Siggraph interactive framerates™ today, but lots of potential: Performance highly dependent on occlusion distance Optimizations / less brute force? Use for high-end / reference / precompute / beauty shots initially Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Conclusions New graphics pipeline usages are opened up with improved HW performance Often not efficient to do with pure compute Continue to give us more performance & bandwidth! We need to continue to break down some fixed graphics pipeline barriers 08/01/10 Beyond Programmable Shading, SIGGRAPH 2010

Acknowledgments Morgan McGuire Louis Bavoil David Luebke Andrew Lauritzen Robert Kihl Christina Coffin SCEE Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Questions? email: [email_address] blog: http://repi.se twitter: @repi For more DICE talks: http://publications.dice.se Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

References [Andersson09] Johan Andersson, “Parallel Graphics in Frostbite - Current & Future”, Beyond Programmable Shading Course – Siggraph 2009 http://s09.idav.ucdavis.edu / [Lalonde09] Paul Lalonde “Innovating in a Software Graphics Pipeline” Beyond Programmable Shading Course – Siggraph 2009 http:// s09.idav.ucdavis.edu/ [Reshetov09] Alexander Reshetov, ”Morphological Antialiasing” [Yang09] Jason C. Yang et al, High Performance Graphics 2009, ” A Directionally Adaptive Edge Anti-Aliasing Filter” [McGuire10] Morgan McGuire, High Performance Graphics 2010, ”Ambient Occlusion Volumes” http://graphics.cs.williams.edu/papers/AOVHPG10/ [Biri10] Venceslas Biri et al, Siggraph 2010, “Practical morphological antialiasing on the GPU” [Bavoil08] Louis Bavoil & Miguel Sainz, Siggraph 2008 “ Image-Space Horizon-Based Ambient Occlusion” http://developer.nvidia.com/object/siggraph-2008-HBAO.html [Hill10] Stephen Hill, Game Developers Conference 2010 ”Rendering with Conviction” [Kihl10] Robert Kihl, Advanced in Real-time Rendering in 3D Graphics and Games, Siggraph 2010, ”Destruction Masking in Frostbite 2 using Volume Distance Fields” http://publications.dice.se [Sugerman09] Jeremy Sugerman et al - ACM Transactions on Graphics January, 2009 ”GRAMPS: A Programming Model for Graphics Pipelines” http://graphics.stanford.edu/papers/gramps-tog/ [Perthuis10] Cedric Perthuis, ”MLAA in God of War 3” (PS3 registered developers only) Beyond Programmable Shading, SIGGRAPH 2010 08/01/10

Bending the Graphics Pipeline

More Related Content

What's hot

Viewers also liked

Similar to Bending the Graphics Pipeline

More from Electronic Arts / DICE

Bending the Graphics Pipeline

Editor's Notes