Introduction to Direct3D 10 Course Porting Game Engines to Direct3D 10: Crysis / CryEngine2 Carsten Wenzel.

Introduction to Direct3D 10 Course

Introduction to Direct3D 10 Course

Porting Game Engines to Direct3D 10: Porting Game Engines to Direct3D 10:

Crysis / CryEngine2Crysis / CryEngine2

Carsten WenzelCarsten Wenzel

OverviewOverview

• IntroductionIntroduction

• D3D10 & CryEngine2D3D10 & CryEngine2

• Porting, integration issues and optimizationsPorting, integration issues and optimizations

• Random bitsRandom bits

• ConclusionConclusion

IntroductionIntroduction

• Early adaptationEarly adaptation

• Spec reviews / discussions started in 2005Spec reviews / discussions started in 2005

• Actual work began middle 2006Actual work began middle 2006

• Support from MS and IHVs was crucialSupport from MS and IHVs was crucial

D3D10 in context of a multi-platform engine (1/3)D3D10 in context of a multi-platform engine (1/3)

• API redesign offers a more concise view on API redesign offers a more concise view on how modern graphics hardware workshow modern graphics hardware works

• For the sake of performance it “offloads” a lot For the sake of performance it “offloads” a lot of responsibility to the API userof responsibility to the API user

D3D10 in context of a multi-platform engine (2/3)D3D10 in context of a multi-platform engine (2/3)

• CryEngine2 is designed to be multi-platformCryEngine2 is designed to be multi-platform

• Render implementations share one interface Render implementations share one interface and a lot of common codeand a lot of common code

– State & resource managementState & resource management

– Draw list processingDraw list processing

– Shader subsystem, etc.Shader subsystem, etc.

D3D10 in context of a multi-platform engine (3/3)D3D10 in context of a multi-platform engine (3/3)

• Certain API changes required special considerationCertain API changes required special consideration

– State objectsState objects

– Shader model 4.0Shader model 4.0

– Constant buffersConstant buffers

– Strict API validation (shader linkage, etc)Strict API validation (shader linkage, etc)

– Buffer updatesBuffer updates

• Support for new features Support for new features

– E.g. Geometry Shader, Stream Out, Texture ArraysE.g. Geometry Shader, Stream Out, Texture Arrays

Porting and making it work (1/2)Porting and making it work (1/2)

• Render code separates platform / API Render code separates platform / API dependent parts from independent partsdependent parts from independent parts

• D3D10 renderer specializes base rendererD3D10 renderer specializes base renderer

• Inherits interface and all common codeInherits interface and all common code

Porting and making it work (2/2)Porting and making it work (2/2)

• Platform / API dependent parts wrapped Platform / API dependent parts wrapped and called from common codeand called from common code

• Provided stubs implemented one at a timeProvided stubs implemented one at a time

• #ifdef#ifdef where API characteristics were where API characteristics were awkward to map to common code flowawkward to map to common code flow

Special considerations (1/5):State objects

Special considerations (1/5):State objects

• Need fine grained control over render states!Need fine grained control over render states!

• Wrap state manipulationWrap state manipulation

• Internally caches state objectsInternally caches state objects

• Currently build state objects on demandCurrently build state objects on demand

• Pre-warming cache might be worthwhilePre-warming cache might be worthwhile

Special considerations (2/5):Shader model 4.0

Special considerations (2/5):Shader model 4.0

• Use custom shader / effect systemUse custom shader / effect system

• Updated to support new API features and Updated to support new API features and enhanced HLSL syntaxenhanced HLSL syntax

• Query resource bindings Query resource bindings vvia Shader Reflectionia Shader Reflection

• Compile existing shaders as is (/Gec)Compile existing shaders as is (/Gec)

• Refactored to utilize load balancing Refactored to utilize load balancing

• Still use shader cacheStill use shader cache

Special considerations (3/5):Constant buffers

Special considerations (3/5):Constant buffers

• Group constants by frequency of updateGroup constants by frequency of update

• Each group maps to a unique CB indexEach group maps to a unique CB index

• Constants bound via semantic which encodes group idConstants bound via semantic which encodes group id

• Allows preprocessor to move constants into their Allows preprocessor to move constants into their appropriate CB scopeappropriate CB scope

– Existing shaders automatically benefit from efficient bindingExisting shaders automatically benefit from efficient binding

– Source shader scripts devoid of API specific annotationsSource shader scripts devoid of API specific annotations

– Shader authors don’t have to worry about APIShader authors don’t have to worry about API

Special considerations (4/5):Strict API validation

Special considerations (4/5):Strict API validation

• API rigorously checks how user drives itAPI rigorously checks how user drives it

• Simplifies lower API / driver levelsSimplifies lower API / driver levels

• Example: Shader linkageExample: Shader linkage

– Vertex layout / output signature has to match input Vertex layout / output signature has to match input signatures of following shadersignatures of following shader

– Dimension of each passed element has to match dimension Dimension of each passed element has to match dimension expected in following shaderexpected in following shader

– Took some time to clean up / unify signaturesTook some time to clean up / unify signatures

– Nice side effect: Signature-fixed shaders run faster on DX9Nice side effect: Signature-fixed shaders run faster on DX9

Special considerations (5/5):Buffer updates

Special considerations (5/5):Buffer updates

• Update methods for truly static / dynamic Update methods for truly static / dynamic buffers remain as in D3D9buffers remain as in D3D9

• Occasional updates of static buffers can Occasional updates of static buffers can cause severe pipeline stallscause severe pipeline stalls(e.g. updating terrain)(e.g. updating terrain)

• Staging resources as intermediate storageStaging resources as intermediate storage

• This way mesh update can be pipelinedThis way mesh update can be pipelined

Random DX10 bits used in CryEngine2Random DX10 bits used in CryEngine2

• GS used for particles and shadowsGS used for particles and shadows

• Viewports for texture atlasViewports for texture atlas

• Filtered texture lookups in vertex shader Filtered texture lookups in vertex shader (FFT water)(FFT water)

• Multi-sampled texture lookupMulti-sampled texture lookup(HDR post-processing)(HDR post-processing)

• Sample depth from z bufferSample depth from z buffer

• Render to volume textureRender to volume texture

• New D3D10 features are great New D3D10 features are great

• Perf improvements are definitely therePerf improvements are definitely there

• Expect several code revisions to drive the Expect several code revisions to drive the API in the most efficient wayAPI in the most efficient way

ConclusionConclusion

More detailsMore details

• [1] “Finding Next-Gen”, Martin Mittring, [1] “Finding Next-Gen”, Martin Mittring, SIGGRAPH 2007SIGGRAPH 2007

– Part of course 28, “Advanced Real-time Rendering Part of course 28, “Advanced Real-time Rendering in 3D Graphics in Games”in 3D Graphics in Games”

– Wed Aug 8Wed Aug 8thth, Room 7AB, 8:30 am - 5:30 pm, Room 7AB, 8:30 am - 5:30 pm

• [2] “Real-time Atmospheric Effects in Games [2] “Real-time Atmospheric Effects in Games Revisited”, Carsten Wenzel, GDC 2007Revisited”, Carsten Wenzel, GDC 2007

– Slides and course notes available onlineSlides and course notes available online

AcknowledgementsAcknowledgements

• Crytek R&D Graphics GroupCrytek R&D Graphics Group

• Many thanks to MS, NVIDIA and ATI for the Many thanks to MS, NVIDIA and ATI for the supportsupport

Questions?Questions?