Movatterモバイル変換

“ A bit more Deferred” - CryEngine 3 Martin Mittring – Lead Graphics Programmer Triangle Game Conference 2009

Crytek Main office: Germany Frankfurt More studios: Kiev, Budapest, Sofia, Nottingham, Seoul English as company language 30+ nationalities CryEngine 1: PC only (Far Cry …) CryEngine 2: PC only (Crysis …) CryEngine 3: PC, XBox360, PS3 (announced GDC09)

32/64 Bit, WinXP/Vista, DX9/10, Multi CPU/GPU WYSIWYP ResourceCompiler Source asset -> Platform specific Direct Light: Shadow mapping Indirect Light (AO): SSAO / RAM / ... No precomputed lighting Production time saver, Memory, Consistency, Dynamic content Übershaders [Mittring07]

Goals after CryEngine 2 PS3 / XBox360 GPU, CPU, memory Improve streaming Improve multithreading Improve lighting More predictable performance Tackle the shader combination issue

What is the shader combination issue? Übershader is one shader with many features (e.g. 0..4 lights, light types, CM reflection, fog, detail texture, normalmap, specular texture) Compiling all possible permutations is a memory, production and performance problem Usual solutions: dynamic branching / separating into multiple passes / reducing combinations and accepting less functionality and less performance Asynchronous shader compiling Distributed Job System to compile the shader cache

Why Deferred Rendering? Rendering is a multi-dimensional query: View x Geometry x Material x Light Classic Forward Rendering: for each light render geometry from scene query with shader Classic Deferred Rendering: render geometry from scene query outputting GBuffer render each light from scene query and shade with GBuffer => Decouples geometric complexity from lighting and shading => Helps on shader combination issue and predictable performance

GBuffer in CryEngine 2 Minimal GBuffer (depth) Slower Early Z pass when outputting linear depth Formats: R16 R16G16 R32 Proved to be very useful Z

Deferred in CryEngine 2 Main use: Deferred shadows, Per pixel fog Additionally: Soft Z clipped Particles, Motion Blur, Beach/Ocean, EdgeAA, Sun Rays, SSAO, Fake lights, 2.5D TerrainAO SSAO ShadowMask

Deferred Lighting in CryEngine 3 Passes: 1) Forward GBuffer generation 2) Deferred light accumulation into texture (Phong) 3) Forward shading with light accumulation texture => No deferred shading Deferred Lighting* + Multiple light primitives are possible + even Image Based Lighting (IBL) + easy to extend Compared to Deferred Shading + Less bandwidth and memory problems (10MB EDRAM XBox360) + More flexibility on shading (besides Phong) Z (native) Normal Specular Power * [Geldreich09] aka Light Pre-Pass Renderer [Engel08]

Options for the light accumulation texture 6 channels: Diffuse and Specular two 7e3 7e3 7e3, A16R16G16B16f or A8R8G8B8* 4 channels: Diffuse and Specular strength a single A16R16G16B16f or A8R8G8B8* (specular approximated by diffuse*strength) The following pictures show lighting with two differently coloured lights: 6 channels (correct) 4 channels (fast) * sRGB helps to distribute more details in dark areas

Light accumulation texture in IBL The following pictures show lighting with Diffuse and Specular Cubemaps: Diffuse RGB Specular RGB High Quality (left) Diffuse RGB Specular Strength Fast Rendering (right) Difference often neglectable (depends on environment)

Storing normals in the GBuffer XYZ world space 8 bit: problematic with extreme reflections/specular 10 bit: good, but what about specular power and PS3 Solving Quantization Artefacts Detail Normalmaps, Noise, Dither XY view space (Z reconstruct) 8/10/16 bit, negate Z bit (perspective and normal mapping) [Lee09] [Lob09] => Problematic

Alternative: VS Normal in 2 scalars -1..1 => -1..1 Normal to GBufffer: G= normalize( N .xy)*sqrt( N .z*0.5+0.5) GBuffer to Normal: N .z=length2( G .xy)*2-1 N .xy=normalize( G .xy)*sqrt(1- N .z* N .z) + more precision where it matters (bright part) + framebuffer blending friendly + no z reconstruction issues wasted area more ALU than WS => still, WS normals are faster

Improved SSAO (with normals) Video

Light rasterization in 2D (Rectangle) or 3D (Convex Object) 2D + cheap WS position reconstruction (Interpolator+MAD) + Combining multiple lights Stencil prepass (if not fullscreen) Coarse blocks can be rejected based on z min/max 3D + Z buffer + tighter bounding object (less pixels to process) Depth bounds test (only on some HW)

Deferred Light Types 1/3: Directional light optional with cloud shadows, multiple shadowmaps Point/Projector lights optional with projector texture Procedural Caustics (before this was multi-pass, one drawcall for each object under water including terrain) Interleaved Shadowmap lookups no extra memory less bandwidth needed no limits on shadow mask channel count

Deferred Light Types 2/3: Image Based Lighting (IBL) Light Probes are the high quality solution for distant light Cubemaps allow efficient HDR lighting in real-time Diffuse CM can be computed from specular CM Mip adjusted lookup allows different specular power values Improves shading in ambient lighting condition by adding normal dependent and specular lighting Light Probes can be generated at specified level positions Deferred Lighting allows blending of localized Light Probes Looks even better with SSAO

Ambient without SSAO with hemispherical lighting

Bright ambient SSAO Black ambient Shadow casting light source SSAO Grey ambient (hemispherical) Shadow casting light source SSAO IBL ambient (Specular and Diffuse) Shadow casting light source SSAO

IBL ambient (Specular and Diffuse) SSAO Video * brightened up for better display

Deferred Light Types 3/3: Real-time Dynamic Global Illumination Details will be presented at upcoming Siggraph 2009 by Anton Kaplanyan who developed that at Crytek Implemented and fast on XBox360, PS3 and PC No precomputation Fully dynamic (geometry, materials and lights) Unified for static and dynamic objects

black ambient (to emphasize where GI affects the image) color bleeding bump without light fully dynamic real-time Global Illumination off Global Illumination on

Global Illumination * brightened up for better display

Something missing? Transparency => falling back to well known techniques: Per pixel global fog and fog volumes (deferred) Back to front sorted alpha transparent objects Volume texture clouds, Imposter clouds, Distance clouds Particle systems avoiding per particle sorting Anti-aliasing => Nasty but possible EdgeAA, … ... we work on it

References [Mittring07] “Finding Next Gen CryEngine2“ Siggraph 2007, Martin Mittring http://ati.amd.com/developer/gdc/2007/mittring-finding_nextgen_cryengine2(siggraph07).pdf [Engel08] “The Light Pre-Pass Renderer“ ShaderX7, Wolfgang Engel http://diaryofagraphicsprogrammer.blogspot.com/2008/03/light-pre-pass-renderer.html [Lee09a] “Prelighting“ Mark Lee http://www.insomniacgames.com/tech/articles/0209/files/prelighting.pdf [Lee09b] “Pre-lighting in Resistance 2“ GDC 2009, Mark Lee http://www.gdconf.com/conference/Tutorial%20Handouts/200_insomniac/gdc09_insomniac_prelighting.pdf [Geldreich09] ”Deferred Lighting and Shading” GDC 2009, Rich Geldreich, Matt Pritchard, John Brooks http://archive.gdconf.com/gdc_2004/pritchard_matt.ppt [Shish05] “Deferred Shading in S.T.A.L.K.E.R.“ GPU Gems 2, Oles Shishkovtsov http://http.developer.nvidia.com/GPUGems2/gpugems2_chapter09.html [Lob09] S.T.A.L.K.E.R : Clear Sky – a showcase for Direct3D 10.0/1 GDC 2009, Igor A. Lobanchikov, Holger Gruen http://www.gdconf.com/conference/Tutorial%20Handouts/100_Advanced%20Visual%20Effects%20with%20Direct3D/100_Handout%202.pdf [Valient07] “Deferred Rendering in Killzone 2“ Develop Conference 2007, Michal Valient http:// www.guerrilla-games.com / publications /dr_kz2_rsx_dev07.pdf Slides should be soon at http://www.crytek.com/technology/presentations Special thanks to all the passionate people at Crytek

Movatterモバイル変換

Change Language

A Bit More Deferred Cry Engine3

Embed presentation

Recommended

More Related Content

What's hot

Similar to A Bit More Deferred Cry Engine3

Recently uploaded

A Bit More Deferred Cry Engine3

Editor's Notes