Movatterモバイル変換

Eliminating TextureWaste: Borderless PtexJohn McDonald, NVIDIACorporation

NVIDIA Corporation © 2013Memory ConsumptionModern games consume a lotof memoryThe largest class of memoryusage is texturesBut lots of texture is wasted!Waste costs both memory andincreased load timesBack/FrontGbufferTexturesVB/IBSimulation

NVIDIA Corporation © 2013Wasted?!Two sources of texture waste:Unmapped texture storage (major)Duplicated texels to helpalleviate visible seams (minor)This cannot eliminate seams.http://www.boogotti.com/root/images/face/dffuse_texture.jpgWaste WasteWasteWasteWaste

NVIDIA Corporation © 2013Wasted?!Two sources of texture waste:Unmapped texture storage (major)Duplicated texels to helpalleviate visible seams (minor)This cannot eliminate seams.http://www.boogotti.com/root/images/face/dffuse_texture.jpg

NVIDIA Corporation © 2013Enter Ptex (a quick recap)The soul of Ptex:Model with Quads instead of TrianglesYou’re doing this for your next-gen engine anyways, right?Every Quad gets its own entire texture UV-spaceUV orientation is implicit in surfacedefinitionNo explicit UV parameterizationResolution of each face isindependent of neighbors.

NVIDIA Corporation © 2013Ptex (cont’d)Invented by Brent Burley at Walt Disney Animation StudiosUsed in every animated film at Disney since 20076 features and all shorts, plus everything inproduction now and for the foreseeablefutureUsed on ~100% of surfacesRapid adoption in DCC toolsWidespread usage throughoutthe film industry

NVIDIA Corporation © 2013Ptex benefitsNo UV unwrapsAllow artists to work at any resolution they wantPerform an offline pass on assets to decide what to ship for eachplatform based on capabilitiesShip a texture pack later for tail revenueReduce your load times. And your memory footprint. Improveyour visual fidelity.Reduce the cost of production’s long pole—art.

NVIDIA Corporation © 2013DemoDemo is running on a Titan.Sorry, it’s what we have at the show. I’ve run on 430-680—perf scales linearly with Texture/FB.Could run on any Dx11 capable GPU.Could also run on Dx10 capable GPUs with small adaptations.OpenGL 4—no vendor-specific extensions.

NVIDIA Corporation © 2013Roadmap: Realtime Ptex v1LoadModelRenderPreprocessDraw TimeBucketandSortGenerateMipmapsFillBordersPackTextureArraysReorderIndexBufferPackPatchConstantsRed: Vertex and Index dataGreen: Patch Constant informationBlue: Texel dataOrange: Adjacency information

NVIDIA Corporation © 2013Roadmap: Realtime Ptex v2LoadModelRenderPreprocessDraw TimePackTextureArraysPackPatchConstantsRed: Vertex and Index dataGreen: Patch Constant informationBlue: Texel dataOrange: Adjacency information

NVIDIA Corporation © 2013Realtime Ptex v2Instead of copying texels into a border region, just go look atthem.Use clamp to edge (border color), with a border color of (0,0,0,0)This makes those lookups fast.Also lets you know how close to the edge you areWe’ll need to transform our UVs into their UV spaceAnd accumulate the resultsWaste factor? 0*.

NVIDIA Corporation © 2013Load ModelVertex DataAny geometry arranged as a quad-based meshExample: Wavefront OBJPatch TexturePower-of-two texture imagesAdjacency Information4 Neighbors of each quad patchEasily load texture and adjacency with OSS library available fromhttp://ptex.us/

NVIDIA Corporation © 2013Texture ArraysLike 3D / Volume Textures, except:No filtering between 2D slicesOnly X and Y decrease with mipmap level (Z doesn’t)Z indexed by integer index, not [0,1]E.g. (0.5, 0.5, 4) would be (0.5, 0.5) from the 5th sliceAPI SupportDirect3D 10+: Texture2DArrayOpenGL 3.0+: GL_TEXTURE_2D_ARRAY

NVIDIA Corporation © 2013Arrays of Texture ArraysBoth GLSL and HLSL* support arrays of TextureArrays.This allows for stupidly powerful abuse of texturing.Texture2DArray albedo[32]; // D3Duniform sampler2DArray albedo[32]; // OpenGL* HLSL support requires a little codegen—but it’s entirely a compile-timeexercise, no runtime impact.

NVIDIA Corporation © 2013Pack Patch ConstantsCreate a constant-buffer indexed byPrimitiveID. Each entry contains:Your Array Index and Slice in theTexture2DArraysYour four neighbors across the edgesEach neighbor’s UV orientation(Again, can be prepared at baking time)If rendering too many primitivesto fit into a constant buffer,you can use Structured Buffers / SSBO for storage.struct PTexParameters {ushort usNgbrIndex[4];ushort usNgbrXform[4];ushort usTexIndex;ushort usTexSlice;};uniform ptxDiffuseUBO {PTexParameters ptxDiffuse[PRIMS];};

NVIDIA Corporation © 2013Rendering Time (DS)In the domain shader, we need to generate our UVs.Use SV_DomainLocation.Exact mapping is dependent onDCC tool used to generatethe meshIncorrect surface orientation

NVIDIA Corporation © 2013Rendering Time (PS)Conceptually, a ptex lookup is:Sample our surface (use SV_PrimitiveID to determine our data).For each neighbor:Transform our UV into their UV spacePerform a lookup in that surface with transformed UVsAccumulate the result, correct for base-level differences and return

NVIDIA Corporation © 2013Bad seamingAll your baseBase level differences, wah?When a 512x512 neighbors a 256x256, their base levels aredifferent.This is an issue because samples are constant-sized in texel(integer) space, not UV (float) space

NVIDIA Corporation © 20130% Waste?Okay, not quite 0.Need a global set of textures that match ptex resolutions used.“Standard Candles”But they are one-channel, and can be massively compressed (4 bitsper pixel)<5 megs of overhead, regardless of texture footprintFor actual games, more like 1K of overhead.Could be eliminated, but at the cost of some shader complexity.Not needed for:Textures without alphaTextures used for Normal MapsTextures less than 32 bytes per pixel

NVIDIA Corporation © 2013A brief interlude on the expense of retrievingtexels from textured surfacesTexture lookups by themselves are not expensive.There are fundamentally two types of lookups:Independent readsDependent readsIndependent reads can be pipelined.The first lookup “costs” ~150 clocksThe second costs ~5 clocks.Dependent reads must wait for previous resultsThe first lookup costs ~150 clocksThe second costs ~150 clocks.Try to have no more than 2-3 “levels” of dependent reads in a singleshader

NVIDIA Corporation © 2013Putting it all togetherFUDRLF.(u, v) = ( 0.5, 0.5 )R.(u, v) = ( 0.5, -0.5 )U.(u, v) = ( 0.5, 1.5 )L.(u, v) = ( 1.5, 0.5 )D.(u, v) = ( 0.5, -0.5 )In this situation, texture lookups in R, U, L and D will return theborder color (0, 0, 0, 0)F lookup will return alpha of 1—so the weight will be exactly 1.

NVIDIA Corporation © 2013Putting it all togetherFUDRLF.(u, v) = ( 1.0, 0.5 )R.(u, v) = ( 0.5, 0.0 )U.(u, v) = ( 0.0, 1.5 )L.(u, v) = ( 2.0, 0.5 )D.(u, v) = ( 0.0, -0.5 )In this situation, texture lookups in U, L and D will return the border color(0, 0, 0, 0)If R and F are the same resolution, they will each return an alpha of 0.5.If R and F are not the same resolution, alpha will not be 1.0—renormalizationwill be necessary.

Movatterモバイル変換

Change Language

Borderless Per Face Texture Mapping

Embed presentation

Recommended

More Related Content

What's hot

Similar to Borderless Per Face Texture Mapping

Recently uploaded

Borderless Per Face Texture Mapping

Editor's Notes