Adaptive NURBS Tessellation with CUDA

Transcription

1 Adaptive NURBS with CUDA Presenter: Dr. Qiu Jie Multi-plAtform Game Innovation Centre (MAGIC), Nanyang Technological University (NTU), Singapore

2 About MAGIC Established in Nanyang Technological University (NTU), Singapore on 1 Nov 2012 Cross-disciplinary research with 26 professors, 33 researchers, 11 PhD To translate scientific ideas into technological products and services To increase the capability of Future Studios in media & entertainment Launched Future Studios Research Lab (FSR Lab) on 21 Jan

3 Adaptive NURBS Convert NURBS to a set of polygons Adaptive tessellation NURBS surface 3 According to Curvature View (camera distance to object) Non-adaptive tessellation Adaptive tessellation

4 Tessellating NURBS surface 4

5 Adaptive NURBS How to be adaptive? Divide the surface into patches Choose accurate, effective criterion to tessellate each patch How to handle the connection of neighbor parts? Extract neighbour info of each patch to check boundary consistency Stitch gaps if not consistent How to speed up? Implementation on GPU 5

6 Proposed Framework To be adaptive Handle boundaries Bézier Conversion Factors Calculation Stitch Gap 6 GPU implementation

7 Bézier Conversion Bézier Conversion Factors Calculation convert NURBS surface to a set of Bézier patches Stitch Gap 7 More accurate for tessellation factor calculation

8 Factors Calculation Bézier Conversion Factors Calculation Stitch Gap n m 8

9 Factors Calculation One traditional method 2 2 i i ε tess max( ( w P ) + ( r ) w ) i i P i : control points w i : weights of control points Related to the origin of the coordinate system Heuristic method: Center of bounding box as origin r 9 r

10 Factors Calculation Our method Using a new objective function to find origin O n n wp i i 2 ( Pj O) + wi ( O) 2 O w min( ) Model j= 0 i= 0 i Old method (ɛ =0.1) Our method (ɛ =0.1) 10 # of vertices

11 Bézier Conversion Factors Calculation Stitch Gap (s,t) Sst (,) = n m i= 0 j= 0 n m i= 0 j= 0 w P B() sb() t ij, ij, i j w B() sb() t ij, i j 11

12 Stitch Gap Bézier Conversion Factors Calculation Stitch Gap 12 Problem: inconsistent tessellation factors between patches gap

13 Stitch Gap 1 side 4 sides Re-design boundary connections 13

14 Implementation on GPU level of parallelism Thread assignment (16,16) (16,16) (16,16) (16,16) (16,16) (16,16) 14

15 Implementation on GPU data management Bézier Conversion Factors Calculation Stitch Gap Minimize data transfer 15

16 Implementation on GPU NURBS surface to Bézier patches Compute Bézier control points 16

17 Implementation on GPU Copy control points needed by each block to shared memory Avoid duplicate memory read for global memory 17 4 X 259 = 1036 points for a block 19 X 19 = 361 points for a block

18 Implementation on GPU Bézier Conversion Factors Calculation Stitch Gap Store tessellation factors of all patches in order to extract neighbor info 18

19 Implementation on GPU Calculate tessellation factors Calculate tessellation factors (2 directions) Store them in global memory n Patch ID 0 1 Direction 1 Direction 2 h 0 v 0 h 1 v 1 2 h 2 v 2 m 19 k h k v k

20 Implementation on GPU Bézier Conversion Factors Calculation Stitch Gap Pre-compute blending functions 20

21 Implementation on GPU using regular pattern Fast evaluation Each row has the same B i (s) Each column has the same (t) ij, ij, i j B j i= 0 j= 0 Sst (,) = n m n m i= 0 j= 0 w P B() sb() t w B() sb() t ij, i j 21

22 Implementation on GPU Bézier Conversion Factors Calculation Stitch Gap 22 Parallel scan to determine vertex/index offset for each patch

23 Implementation on GPU Vertex offset Determine # of vertices for each patch based on tessellation factors Determine vertex offset for each patch Patch ID # vertices vertex offset 0 (#v) (#v) 1 (#v) 2 Prefix sum (#v) 0 (#v) 0 +(#v) 1 23 k (#v) k SUM((#V) 0, (#V) k-1 )

24 Implementation on GPU Transition region and index offset Determine # of indices on each patch (include transition regions) Determine index offset for each patch Patch ID # indices index offset 0 (#id) (#id) 1 (#id) 2 Prefix sum (#id) 0 (#id) 0 +(#id) 1 24 k (#id) k SUM((#id) 0, (#id) k-1 )

25 Implementation on GPU Bézier Conversion Factors Calculation Stitch Gap Minimize data transfer Store patch info for easy neighbour info extraction Regular tessellation pattern Parallel scan 25

26 Video Demonstration 26

27 Results 27 Head Old man face Stegosaurus Patch No Triangle No

28 Results CPU (intel i GHz) vs GPU (GTX CUDA cores 160GB/s bandwidth) 28 Run time / ms Head Old man face Stegosaurus CPU GPU

29 Results Car model Error tolerance surfaces 3,081,930 triangles 10.4 ms 29

30 Results Submarine model Error tolerance surfaces 2,235,934 triangles 9.16 ms 30

31 Acknowledgement This research is supported by Multi-plAtform Game Innovation Centre (MAGIC), funded by the Singapore National Research Foundation under its IDM Futures Funding Initiative and administered by the Interactive & Digital Media Programme Office, Media Development Authority of Singapore, and the National Science Foundation of China (Grant no , ). 31

32 Thank you!