1. From Turing Machine to Global Illumination

2. Outline My first computer (CASIO fx3600)
Turning machine and von Neumann architecture
GPU pipeline
Local and global illumination
Shadow and reflection through texture
Programmable GPUs

3. Calculator vs. Computer
What is the difference between a calculator and a computer?
Doesn’t a compute-r just “compute”?
The Casio fx3600p calculated can be programmed (38 steps allowed).

4. Turing Machine
Can be adapted to simulates the logic of any computer that could possibly be constructed.
von Neumann architecture implements a universal Turing machine.
Look them up at Wikipedia!

5. Outline My first computer (CASIO fx3600)
Turning machine and von Neumann architecture
GPU pipeline
Local and global illumination
Shadow and reflection through texture
Programmable GPUs

8. Simplified View The Data Flow:
3D Polygons (+Colors, Lights, Normals, Texture Coordinates…etc.)
2D Polygons
2D Pixels (I.e., Output Images)
Transform
(& Lighting)
Rasterization

9. Outline My first computer (CASIO fx3600)
Turning machine and von Neumann architecture
GPU pipeline
Local and global illumination
Shadow and reflection through texture
Programmable GPUs

10. Global Effects
shadow
multiple reflection
translucent surface

11. Local vs. Global

12. How Does GPU Draw This?

13. Quiz
Q1: A straightforward GPU pipeline give us local illumination only. Why?
Q2: What typical effects are missing?
Hint: How is an object drawn? Do they consider the relationship with other objects?
Shadow, reflection, and refraction…

14. Wait but I’ve seen shadow and reflection in games before…
With Shadows
Without Shadows

15. Outline My first computer (CASIO fx3600)
Turning machine and von Neumann architecture
GPU pipeline
Local and global illumination
Shadow and reflection through texture
Programmable GPUs

16. Adding “Memory” to the GPU Computation
Modern GPUs allow:
The usage of multiple textures.
Rendering algorithms that use multiple passes.
Transform
(& Lighting)
Rasterization
Textures

17. Faked Global Illumination
Shadow, Reflection, BRDF…etc.
In theory, real global illumination is not possible in current graphics pipeline:
Conceptually a loop of individual polygons.
No interaction between polygons.
Can this be changed by multi-pass rendering?

18. Case Study: Shadow Map Using two textures: color and depth
Relatively straightforward design using pixel (fragment) shaders on GPUs.

19. Eye’s View
Light’s View
Depth/Shadow Map
Image Source: Cass Everitt et al., “Hardware Shadow Mapping” NVIDIA SDK White Paper

20. Basic Steps of Shadow Maps
Render the scene from the light’s point of view,
Use the light’s depth buffer as a texture (shadow map),
Projectively texture the shadow map onto the scene,
Use “texture color” (comparison result) in fragment shading.

21. Outline My first computer (CASIO fx3600)
Turning machine and von Neumann architecture
GPU pipeline
Local and global illumination
Shadow and reflection through texture
Programmable GPUs

22. PC Graphics Architecture
Two buses on PC: System Bus (CPU-Memory) and Peripheral I/O Bus.
Before AGP: narrow path (I/O Bus) between main memory and graphics memory (for frame buffer, Z buffer, texture, vertex data…etc.)
AGP and PCI-e speed up the link between host PC and graphics processor (GPU)

23. Source: http://www.karbosguide.com/hardware/module2d03a.htm

24. NVIDIA Geforce 6800

25. NVIDIA Geforce 8800

26. NVIDIA Fermi (Geforce 400 and 500 Series)
From NVIDIA Fermi Architecture Whitepaper

27. How to Program a GPU? Writing a 3D graphics application program
Typically in DirectX or OpenGL
Still CPU programming in C/C++
The APIs and drivers do the dirty work for you.
Writing GPU shaders
Typically in GLSL or Cg
Still drawing 3D objects
Working like plug-in’s to the 3D rendering pipeline

28. GPGPU General-purpose GPU computing
No longer restricted to graphics applications.
To utilize the abundant “GFLOPs” in GPU.
Could be implemented in GPU shaders
By clever transformation of problem domains.
Textures to store the data structures
However, shaders could not perform memory writes with calculated addresses (a.k.a. scatter operations)

29. GPU as a Parallel Computing Platform
Treating GPUs as parallel machinery
Not quite the same as shared-memory multi-processor.
A special kind of memory hierarchy.
NVIDIA CUDA
Widely adopted in real-world applications
OpenCL
For non-NV GPUs and multi-core CPUs

30. Branch Divergence on GPU
Warp …
if x1 – x0 > y1 – y0:
xMajorIteration()
else:
yMajorIteration()
GPUs group pixels into “warps” or “wavefronts” of adjacent pixels, each of which runs on a single core of a superscalar processor.
Here we assume the ray is in the first quadrant of the plane. Green threads represent rays where the slope of the ray in screen space is less than one, and blue thread have slope more than 1
At a statement where pixels within a group branch in different directions, the GPU must compute both sides of the branch, issuing no operation to a portion of the vector lanes for each branch and this is the case even if only a single thread in a warp takes a particular branch. So, while branch instructions themselves are often relatively expensive compared to arithmetic instructions, their most important cost is that divergent execution will reduce throughput by one half for each nested branch.
Note that this code only handles the first quadrant; we need 2 more nested branches to handle all four quadrants in naïve implementation
More branches for handling clipping to viewport
Total of 8x reduction in performance for handling all cases, plus overhead for clipping branches against the viewport.
We rederive the algorithm to make it branchless except for the actual hit detection. …
½ performance for each branch!

31. Examples

32. GPU Shading Effects Reflection and refraction Relief on surface
Ambient occlusion and lighting

33. Real-Time Rendering of Splashing Water
Particle system simulation for real-time interaction with terrains and dynamic objects.
Reconstruction of the splash surface with 2D metaballs

34. Ray Tracing on GPU
Using OpenCL or NVIDIA CUDA
Or use NVIDIA OptiX