Look Ma, No Jutter! Optimizing Performance Across Oculus Mobile
2 likes2,256 views
The document outlines optimization techniques for mobile applications on Oculus, focusing on performance and file size efficiency. It discusses methods like texture compression, batching, and lighting strategies, emphasizing trade-offs between quality and performance. Additionally, it includes profiling insights to determine if a game is CPU or GPU bound and provides recommendations for shader and geometry simplification to enhance rendering efficacy.
Look Ma, No Jutter! Optimizing Performance Across Oculus Mobile
- 2. Optimizing Performance Across Oculus Mobile Gabor Szauer Developer Relations Engineer, Oculus
- 3. Best balance of quality to file size Maybe not for UI ASTC Texture Compression
- 7. Half your draw calls! OpenGL Extension Renders to frame buffer array CPU side win Single Pass Stereo
- 8. MULTI PASS SetPass Calls: 46 SINGLE PASS SetPass Calls: 22
- 9. Batch Meshes Atlas Textures Reduce Number Of Materials Further Reducing Draw Calls
- 11. LOTS OF TEXTURES Unique Materials More Draw Calls SINGLE ATLAS Shared Material Less Draw Calls
- 12. 4,306 POLYGONS 2 MATERIALS 4 TEXTURES 2,477 POLYGONS 1 MATERIAL 1 TEXTURE
- 13. Pixel Light Count • 1 is ok • 0 is preferred Shadows • hard shadows is ok • disable shadows preferred Bake shadow maps if possible Light & Shadows
- 14. Can't avoid additional resolve Additional resolve: 1ms + Bake environment lights Few dynamic shadow casters Real time shadows
- 15. Avoid shadow projector Re-rendering batched geometry is expensive Simple quad at feet will do in most cases Blob shadows
- 16. Renders model twice First render is the model projected onto plane No self shadowing Same limitations as blob shadows, but looks better Planar Projection Shadows
- 17. Environment should not be lit in real time Bake lighting / shadow maps 4096 is a pretty good size These settings are very project dependent. Bake Lighting Data
- 18. Dynamic characters no longer receive shadows from baked objects Bake Lighting Data
- 19. Sample light map to make moving object kind of match lighting Start with a recast down, then: Renderer.lightmapIndex LightmapData.lightmapColor RaycastHit.lightmapCoord Sample light map for shadow value
- 20. Use light probes! Fast, easy to set up, looks great! Lightprobe shadow volumes
- 21. Use light probes! Fast, easy to set up, looks great! Lightprobe shadow volumes
- 22. Impostor geometry, just a quad with some textures Add particles to make extra fancy Fake lighting effects
- 23. Linear • More Realistic • OpenGL ES 3 Required Gamma • Less Realistic • Works Everywhere Color Space
- 24. GAMMA LINEAR
- 25. Per material color grading Done in forward pass Use Lookup Table Edit lookup table in any image editing application Per Material Color Grading
- 26. Per material color grading Done in forward pass Use Lookup Table Edit lookup table in any image editing application Per Material Color Grading
- 27. Set to 4x Use the "Use Recommended MSAA Level" in OVRManager when possible Anti Aliasing
- 28. 2X MSAA0X MSAA 4X MSAA
- 29. No reason not to use OpenGL ES 3 Vulkan fallback is ok Vulkan is a CPU side win Graphics API
- 30. Multithreaded rendering Static Batching GPU Skinning Turn on
- 31. API Level 23+ Device Filter: ARM Device filter can make about 10 MB difference in APK size Other settings
- 32. Scripting back-end: IL2CPP Prebake collision meshes Keep Loaded Shaders Alive Optimize Mesh Data Consider these
- 33. Anisotropic: off Bone weights: 2 V-Sync: Don't Everything else
- 34. Try to get Render Texture Switches down to 0 if possible * Not possible if using real time shadows Reduce Resolves
- 35. Avoid Using Post Effects Add quad in front of camera Fade quad to fade scene Screen Fade
- 36. VR.WaitForGPU = Wait for Vsynch If you are making frame rate, there should be a few milliseconds of VR.WaitForGPU Frame Times: 30 FPS = 33.3 ms 60 FPS = 16.6 ms 72 FPS = 13.8 ms Reading the profiler
- 38. Internal Unity Profiler Android Unity internal profiler stats: cpu-player> min: 15.8 max: 16.9 avg: 16.4 cpu-ogles-drv> min: 0.0 max: 0.0 avg: 0.0 gpu> min: 0.0 max: 0.0 avg: 0.0 cpu-present> min: 0.3 max: 0.6 avg: 0.3 frametime> min: 16.1 max: 17.3 avg: 16.7 batches> min: 5 max: 5 avg: 5 draw calls> min: 5 max: 5 avg: 5
- 39. Profiling Workflow
- 40. Useful for measuring general performance Not useful for knowing what is slow Not enough info to really optimize Frames per second
- 41. OVR Metrics Tool
- 42. Milliseconds per frame 30 FPS = 33.3 Milliseconds / Frame 60 FPS = 16.6 Milliseconds / Frame 90 FPS = 11.1 Milliseconds / Frame
- 43. Example: Frame Breakdown A single frame takes 21.2 ms and waits 12 ms for the next vsynch The current total frame duration is 33.3 ms To run at 60 FPS, the frame needs to finish in 16.6 ms To hit the 60FPS mark, at least 5.2 ms has to be optimized away from this frame * This example assumes everything runs on one thread to make it easy to read
- 44. CPU or GPU Bound? If the frame rate improved, the game was GPU Bound If the frame rate dod not improve, the game was CPU Bound To determine if a game is CPU or GPU bound: Don't render anything!
- 45. Use Unity Profiler CPU Bound
- 46. If the game is GPU bound, it can be Vertex of Fill bound. To determine which, change render scale to something small, like 0.01 Did performance improve? Yes - Fill / Fragment Bound No - Vertex Bound GPU Bound
- 47. Qualcomm Snapdragon Profiler Mali Graphics Debugger Renderdoc Fill Bound
- 48. Simplify meshes which contain too many vertices Consider combining meshes to reduce draw calls Cull invisible meshes and enable LOD on meshes Vertex Bound
- 49. Rinse and Repeat Disable all rendering. Did performance improve? GPU or CPU Bound No CPU Bound Use instrumented profiler GPU BoundYes Vertex or Fill Bound? Set render scale to 0.01 Did performance improve? No Yes Fill Bound Vertex Bound Simplify Shaders Simplify Geometry Combine Geometry Reduce Draw Calls