CS535 Fall Department of Computer Science Purdue University

Transcription

1 Culling CS535 Fall 2010 Daniel G Aliaga Daniel G. Aliaga Department of Computer Science Purdue University

2 Types of culling View frustum culling Visibility culling e.g., backface culling, portal tlculling, textured t depth mesh culling Occlusion culling e.g., in hardware, hierarchical occlusion maps, hierarchical z buffer

3 Portal Culling Increase interactive rendering performance of architectural walkthroughs Cull away non visible areas Replace geometry behind portals (doorways and windows) with images warped to the current eye position Reduce the rendering complexity with a small number of image samples

4 Cells and Portals

5 Portal Images

6 Portal Image Warping McMillan & Bishop Warping Equation: x 2 = δ(x 1 ) P 2-1 (c 1 -c 2 ) + P 2-1 P 1 x 1 Move pixels based on distance to eye ~Texture mapping Per pixel distance values are used to warp pixels to their correct location for the current eye position

7 Portal Image Warping eye Reference COP Projecting the eye onto the reference image partitions it into 1, 2 or 4 sheets A raster scan of each sheet produces a back to front ordering of warped pixels

8 Creating Portal Images Ideal portal image would be one sampled from the current eye position eye portal

9 Creating Portal Images Reference COPs A large number of static reference images (~120) needed to eliminate popping portal

10 Creating Portal Images Reference COPs With warping we use a much smaller number of reference images portal

11 Creating Portal Images Reference COPs Choose closest reference image and warp to current eye position eye portal

12 Problem: Exposure Events We observe tears in regions with unknown visibility data

13 Reducing Exposure Events To address this we: 1. Render multiple reference images 2. Allow warped pixels to persist and fill-in minor tears

14 Parallelization: Sheets eye Projection plane Eye view direction similar to reference COP view direction Since the projection plane is perpendicular p to the view direction, we usually have 4 sheets We warp each sheet on a different processor

15 Sample Path: 1 Image/Portal Re endering Time (sec.) Portal Culling only Portal Image Frame

16 Sample Path: 2 Images/Portal Re endering Time (sec.) Portal Culling only Portal Image Frame

17 Textured Depth Mesh Culling Model Space Partitioning Preference to important areas Far Geometry Replace with textured depth meshes Near Geometry Could use LOD, occlusion and view frustum culling 17

18 Model Space Partitioning View Preference Function (VPF) Similar to progressive rendering, allocate resources (e.g. time, storage, CPU, bandwidth) to most important areas first 18

19 Model Space Partitioning Subdivide important areas into cells Cells Model space 19

20 Model Space Partitioning Associate a cull box with each cell r 6 Cull box 44 20

21 Model Space Partitioning Cull box divides near and far geometry Far geometry Near geometry 21

22 Far Geometry Replace far geometry with ihtextured ddepth Meshes (TDMs) Near geometry TDM 0 22

23 Far Geometry Replace far geometry with ihtextured ddepth Meshes (TDMs) TDM 1 Near geometry TDM 0 23

24 Far Geometry Replace far geometry with ihtextured ddepth Meshes (TDMs) TDM 1 TDM 2 TDM 0 Near geometry 24

25 Far Geometry Replace far geometry with ihtextured ddepth Meshes (TDMs) TDM 1 TDM 2 TDM 0 Near geometry TDM 3 25

26 Textured Depth Mesh ImageI Depth D himage Simplify Preprocessing Texture Mesh Run-time Rendering TDM 26

27 Example TDM Foreground Geometry Background TDM with mesh outlined in white 27

28 TDM Skins Example of a simple TDM A B Eyepoint 28

29 TDM Skins Example of a simple TDM TDM A B Eyepoint 29

30 TDM Skins Disocclusions appear as skins A AA B Skin B Eyepoint 30

31 Near Geometry Cull to view frustum Select LOD based on distance View Frustum Eyepoint Cull Box 31

32 Near Geometry If sufficiently complex, do occlusion culling e.g., Hierarchical Occlusion Maps View Frustum Eyepoint Cull Box 32

33 What size cull box? Given a fixed primitive budget and a fixed set of cells: Large cull boxes require coarse LODs Small cull boxes exaggerate skins in the TDMs (visibility culling does not alter the visual content) 33

34 Balancing Near Far Geometry Error metrics LOD: deviation in pixels from original surface TDM: maximum skin size, in screen pixels Reduce and balance thecombined error Adjust cull box size and LODs to meet budget and balance errors 34

35 Multiple Processor Pipeline Thread 0 Frame interval Occlusion cull (n+1) 1 Cell cull (n+1) View-frustum cull (n+1) Level-of-detail selection (n+1) Occlusion cull (n+1) Occluder TDM render Geometry render 2 render (n+1) (n) (n) 3 Prefetching (asynchronous) 35

36 Power Plant Frame Rates 30 VFC Only Cold Cache Warm Cache 25 Fra ames Per Second Frame Number 36

37 Questions Order of the acceleration techniques affects the results, so which is best? Cull Box VF Culling LOD Occl. Culling Cull Box LOD VF Culling Occl. Culling VF Culling Occl. Culling Cull Box LOD (a) (b) (c) 37

38 Occlusion Culling As opposed to visibility culling, in occlusion culling the algorithm determines what is occluded (instead of what is visible) This leads to the confounding, but true, equation: not visible occluded

39 Occlusion Culling in Hardware Occlusion Query (e.g., nv_occlusion_query) Performed at the polygon level Tells the caller whether any pixels (or maybe how many) of a rendered object/polygon actually became visible Thus, can render a box instead of a complex object if the entire box is occluded, then complex object is occluded Result is potentially faster rendering performance

40 Occlusion Culling in Hardware Early Z rejection Performed at the pixel level If a fragment is not visible (because the depth test failed), rejection occurs immediately Cansave unnecessary texture fetching and execution of GPU level code Result is a saving of memorybandwidth

41 Hierarchical Occlusion Maps Occluded geometry Visible geometry Occluder geometry (all images from Hansong Zhang s dissertation)

42 Hierarchical Occlusion Maps Motivation Occlusion is cumulative Occluders can be fused

43 Occlusion Culling Algorithm Progressive Culling Con? M t t i f t t b k d t t b t Must traverse in front to back order to get best use of culling

44 Occlusion Culling Algorithm Progressive Culling Con? Multiple l passes (could ldb be many )

45 Occlusion Culling Algorithm Multi pass occlusion culling

46 A Definition of Occlusion where R and R o are opaque and not opaque regions

47 Hierarchical Occlusion Map (HOM) The occlusion images can be readily generated by hardware; how? Just render the scene with white=visible and black=background Note that this process is fast but is also done at some selected resolution So lets change the resolution!

48 Hierarchical Occlusion Map (HOM) And we get this:

49 Hierarchical Occlusion Map (HOM) And we get this:

50 Hierarchical Occlusion Map (HOM) Creation algorithm for HOMs:

51 Hierarchical Occlusion Map (HOM) How do you use an HOM? First cull using the coarsest HOM Then the next coarsest... Until the finest level of the HOM If something is visible ibl in all levels, l then it is visible ibl So what is the gain?

52 Hierarchical Occlusion Map (HOM) Observation Even at the finest level, it is not known if the object is truly visible But at the resolution you care about, it is

53 Hierarchical Occlusion Map (HOM) Conclusion The hierarchy is conservative Generate from the coarest HOM to the finest level you care about Only use those and select at run time depending on performance

54 Example Results

55 Some questions? For a 2x2 grid of pixels in the occlusion map, what occlusion value do you choose? Min? Max? Average? Wl Welcome to hierarchical z buffers!