Hardware-Compatible Vertex Compression Using Quantization and ...

More documents

Recommendations

Info

System Overview - PreprocessingSystem Overview - RuntimeInputOptimizationOutputInputVertex ProgramOutputModelUserParameterBPVSimplifyPackQuantizeCompressedModelCompressionBlueprintCompressedModelCompressionBlueprintDecompress• Right Shift•Mod• Left shift• CombineUser DefinedTransformandLightingBit AllocationGoal:“Optimize allocation of bits to each vertexattribute component (x, y, z, nx, ny, nz,r, g, b, …) subject to the target bits pervertex and the error metric.”Image-Space Error MetricBased on Lindstrom and Turk 2000’s workon simplifying 3D meshes.•Quantify error of allocating bits to multipleattributes (position vs normals vs colors)automatically.•Account for any desired renderingalgorithm and shader.Image-Space Error Metric (cont.)QuantizationAlgorithm (n = 20):•Render n images from n viewpoints.•Compare n images of altered models torenderings of original model.•RMS of difference images.•Do not consider background pixels.(a)(c)(b)(d)Vertex attributes:• Position• Normals• Texture coord.RocketCar model courtesy of NVIDIA Corporation
Quantization (cont.)Quantization (cont.)Original (896 KB)(xyz, normal, texcoord) =(32 32 32 32 32 32 32 32)Error = 0.0Our result at 64 bpv (224 KB)(xyz, normal, texcoord) =(11 11 12 5 5 5 7 8)Error = 0.018Our result at 64 bpv (224 KB)(xyz, normal, texcoord) = 11 11 12 5 5 5 7 8Error = 0.018DirectX at 96 bpv (336 KB)(xyz, normal, texcoord) = 10 10 10 10 10 10 16 16Error = 0.021Quantization (cont.)Bit Allocation AlgorithmOur result at 64 bpv (224 KB)(xyz, normal, texcoord) = 11 11 12 5 5 5 7 8Error = 0.018OpenGL at 96 bpv (336 KB)(xyz, normal, texcoord) = 8 8 8 8 8 8 16 16Error = 0.040Input: BPV• Initialize bit allocation• Iterate until error no longer decreasesIncreaseFor each attribute component, increase one bit andmeasure current errorPick the one with the lowest error.DecreaseFor each attribute component, decrease one bit andmeasure current errorPick the one with the lowest error.PackingCombining Quantization andSimplificationRocketCar 64 bpv• (xyz, normals, texcoord) = 11 11 12 5 5 5 7 8• Stored into a vectorized data element of 4unsigned shorts.[0][1][2][3]Input: Target model sizeGoal:“Compute quantization and simplification thatmaximize the rendering quality for thespecified size”xyznxnynzuv
Page 1: Hardware-Compatible VertexCompressi
Page 6 and 7: Extracting BitsExtracting Bits2. Fr
Page 8: Future WorkExploiting graphics hard

Hardware-Compatible Vertex Compression Using Quantization and ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?