Computer Science Thesis Defense
- 1. 1 GPU Ray Tracing with CUDA BY TOM PITKIN Bill Clark, PhD Stu Steiner, MS, PhC
- 2. Objectives Develop a sequential CPU and parallel GPU ray tracer Illustrate the difference in rendering speed and design of a CPU and GPU ray tracer 2
- 3. Outline Introduction to Ray Tracing CUDA Parallelization with CUDA / Results Future Work Questions 3
- 4. What is Ray Tracing? Rendering technique used in computer graphics Simulates the behavior of light Can produce advanced optical effects 4
- 5. Light in the Physical World 5 Light Source Film Object with Red Reflectivity Pinhole
- 6. The Virtual Camera Model Eye Position – camera location in 3D space Reference Point – point in 3D space where the camera is pointing Orientation Vectors (u, v, n) – camera orientation in 3D space Image Plane – projected plane of the camera’s field of view Reference Point v (Up Vector) n u Eye Position 6
- 7. Ray Generation Map the physical screen to the image plane Divide the image plane into a uniform grid of pixel locations 7 Send a ray through the center of each pixel location 𝐼𝑚𝑎𝑔𝑒 𝑃𝑙𝑎𝑛𝑒 𝐻𝑒𝑖𝑔ℎ𝑡 𝑆𝑐𝑟𝑒𝑒𝑛 𝐻𝑒𝑖𝑔ℎ𝑡 Pixel Eye Position 𝐼𝑚𝑎𝑔𝑒 𝑃𝑙𝑎𝑛𝑒 𝑊𝑖𝑑𝑡ℎ 𝑆𝑐𝑟𝑒𝑒𝑛 𝑊𝑖𝑑𝑡ℎ
- 8. Ray Intersection Testing Ray – Sphere Intersection Ray – Triangle Intersection 8
- 10. Specular Reflection Recursive Ray Tracing 10
- 11. Outline Introduction to Ray Tracing CUDA Parallelization with CUDA / Results Future Work Questions 11
- 12. What is CUDA? Compute Unified Device Architecture (CUDA) Parallel computing platform Developed by Nvidia 12
- 13. Kernel Functions Specifies the code to be executed in parallel Single Program, Multiple Data (SPMD) 13
- 15. Memory Model Global Memory Constant Memory Texture Memory Registers Local Memory Shared Memory 15
- 16. Outline Introduction to Ray Tracing CUDA Parallelization with CUDA / Results Future Work Questions 16
- 17. Thread Organization 2D array of blocks 2D array of threads 17 Each thread represents a ray Block (0, 0) Block (1, 0) Block (2, 0) Block (0, 1) Block (1, 1) Block (2, 1) Image Plane
- 18. Testing Environment OS – Ubuntu Gnome Remix 13.04 CPU – Core i7-920 Core Clock – 2.66 GHz GPU – Nvidia GTX 570 Core Clock - 742 MHz CUDA Core - 480 Memory Clock - 3800 MHz Video Memory - GDDR5 1280MB 18
- 19. Test Objects Teapot Surfaces: 1 Triangles: 992 Al Surfaces: 174 Triangles: 7,124 Crocodile Surfaces: 6 Triangles: 34,404 19
- 20. Single Kernel 20 Single Kernel Single Thread 160 (0.16 sec) Teapot (Surfaces: 1) (Triangles: 992) 23,003 (23 sec) 411 (0.41 sec) Al (Surfaces: 174) (Triangles: 7,124) 55,260 (55.26 sec) 5,867 (5.87 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 1,617,160 (26.95 min) 1 10 100 1,000 10,000 Milliseconds 100,000 1,000,000 10,000,000
- 21. Kernel Complexity and Size Driver timeout Register Spilling 21
- 22. Replacing Recursion Iterative Loop Layer based stack Layers store color values returned from rays Final image from convex combination of layers 22
- 23. Multi-Kernel 23 Multi-Kernel Single Kernel (Previous Kernel) 381 (0.38 sec) Teapot (Surfaces: 1) (Triangles: 992) 160 (0.16 sec) 967 (0.97 sec) Al (Surfaces: 174) (Triangles: 7,124) 411 (0.41 sec) 13,217 (13.22 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 5,867 (5.87 sec) 1 10 100 1,000 Milliseconds 10,000 100,000
- 24. Multi-Kernel with Single-Precision Floating Points 24 Multi-Kernel with Single-Precision Floating Points Multi-Kernel (Previous Kernel) 46 (0.05 sec) Teapot (Surfaces: 1) (Triangles: 992) 381 (0.38 sec) 118 (0.12 sec) Al (Surfaces: 174) (Triangles: 7,124) 967 (0.97 sec) 1,556 (1.56 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 13,217 (13.22 sec) 1 10 100 1,000 Milliseconds 10,000 100,000
- 25. Caching Surface Data Object’s surface data stored on shared memory All threads in same block have access to cached surface data Removes duplicate memory requests Data reuse 25
- 26. Multi-Kernel with Surface Caching 26 Multi-Kernel with Surface Caching Multi-Kernel with Single-Precision Floating Points (Previous Kernel) 30 (0.03 sec) Teapot (Surfaces: 1) (Triangles: 992) 46 (0.05 sec) 133 (0.13 sec) Al (Surfaces: 174) (Triangles: 7,124) 118 (0.12 sec) 1,007 (1.01 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 1,556 (1.56 sec) 1 10 100 Milliseconds 1,000 10,000
- 27. Simplifying Mesh Data 27 Triangle data originally stored as three points (vertices) Optimize data by storing triangles as one point (vertex) and two edges Calculate edges on host before kernel call 0.5, 1 0, 0 0.5, 1 1, 0
- 28. Multi-Kernel with Mesh Optimization 28 Multi-Kernel with Mesh Optimization Multi-Kernel with Surface Caching (Previous Kernel) 27 (0.03 sec) Teapot (Surfaces: 1) (Triangles: 992) 30 (0.03 sec) 127 (0.13 sec) Al (Surfaces: 174) (Triangles: 7,124) 133 (0.13 sec) 873 (0.87 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 1,007 (1.01 sec) 1 10 100 Milliseconds 1,000 10,000
- 29. Final Results 29 Multi-Kernel with Intersection Optimization Single Thread 27 (0.03 sec) Teapot (Surfaces: 1) (Triangles: 992) 23,003 (23 sec) 127 (0.13 sec) Al (Surfaces: 174) (Triangles: 7,124) 55,260 (55.26 sec) 873 (0.87 sec) Crocodile (Surfaces: 6) (Triangles: 34,404) 1,617,160 (26.95 min) 1 10 100 1,000 10,000 Milliseconds 100,000 1,000,000 10,000,000
- 30. Outline Introduction to Ray Tracing CUDA Parallelization with CUDA / Results Future Work Questions 30
- 31. Future Work Spatial partitioning Multiple GPUs Optimize code for different GPUs 31
- 32. Questions? 32
Editor's Notes
- #3: Used C++ and CUDA
- #7: Forward Ray TracingBackward Ray Tracing
- #8: Pixel – picture element that represents one point on an image. Consists of a single color
- #9: Don’t forget to mention what happens if a ray misses completely
- #10: Ambient Light – indirect light reflected off of other objects in the sceneDiffuse Light – direct light reflected off the surface in all directionsSpecular light – direct light reflected off the surface in a single direction
- #15: Block and Threads have unique identifier
- #16: Register Memory – 50x faster than Global MemoryL2 Cache – LRU (Least Recently Used)L1 Cache – Spatial Locality (Quickly access memory in nearby location of current memory reference), Caches per-thread stack and other local data structures
- #21: Logarithmic ScaleSingle Pass, 640 x 480
- #30: 1852X speedup!