1 / 40

GPU Programming

GPU Programming. Lecture 2 Comp 175: Computer Graphics April 22, 2014. Interfacing with the GPU. Modern OpenGL Mobile/Web Programming Direct3D Compute Languages Stream, OpenCL , CUDA Future. Modern OpenGL. OpenGL 2.0 and beyond. Client/Server Architecture. Think of like a network

amory
Télécharger la présentation

GPU Programming

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. GPU Programming Lecture 2 Comp 175: Computer Graphics April 22, 2014 Michael Shah

  2. Interfacing with the GPU • Modern OpenGL • Mobile/Web Programming • Direct3D • Compute Languages • Stream, OpenCL, CUDA • Future Michael Shah

  3. Modern OpenGL OpenGL 2.0 and beyond Michael Shah

  4. Client/Server Architecture • Think of like a network • Client = the CPU • Server = the GPU • The less they talk, the better • Less waiting for memory to transfer • Bandwith is almost always the bottleneck • Send data over once, and have it ready to go Michael Shah

  5. Optimizations Available • Vertex Array • Display Lists • Vertex Buffer Objects • Pixel Buffers • Frame Buffer Objects Michael Shah

  6. Immediate Mode • Every ‘gl’ command is being sent to the GPU every iteration. • Very inefficient • Think of it like having an interpreter versus a compiler for C++ Michael Shah

  7. Vertex Array • No more glBegin/glEnd • Store Vertices, indices, normals, colors, texture coordinates in an array • Minimizes OpenGL calls on Client (CPU) • Data still being passed from Client to Server every frame • A little better because it’s all at once • Stylistically, much cleaner code as well Michael Shah

  8. Display List • Compiled geometry • Send it to the Server(GPU) once, and it is stored in memory on Server • Geometry cannot be updated, without deleting display list. someGeometryDisplayList= glGenLists(1); glNewList(someGeometryDisplayList,GL_COMPILE); drawShape(); glEndList(); Michael Shah

  9. Vertex Buffer Object (VBO) • Improves upon Vertex Array • We can store data on the Server (GPU) • Just like display lists! • We get a pointer into the Server (GPU) so that we can update it! • As good as display lists! • Because we have a pointer to the data, multiple clients can access data. • Multiple instances of a program can access data Michael Shah

  10. Buffer ObjectPattern Step 1: • glGenBuffer // Allocate some memory • glBindBuffer // Populate memory • glBufferData // copy data over to GPU Step 2: • glEnable // Enable drawing of Buffer Object • glDrawElements // draw the data Michael Shah

  11. Pixel Buffer Object (PBO) • Same as Vertex Buffer Object • But--We are storing pixel data • Pixels can be stored to make up textures • Static • Dynamic • Streaming • Not directlytied to texture Michael Shah

  12. PBO Use Case Michael Shah

  13. FramebufferObject (FBO) • Two Use Cases • Render Buffer – Controlled by Windows System • This is what we see (Deferred Render Example) • Texture Images • Used for rendering textures to geometry • Can be static, dynamic, or streaming data just like with VBO. • Used to give us cool animations, or simply play videos on shapes. • Video Example Michael Shah

  14. Mobile/Web Graphics OpenGL ES and WebGL Michael Shah

  15. OpenGL ES[1] • OpenGL ES 1.X • Fixed Function pipeline (OpenGL 1.5) on mobile platforms • OpenGL ES 2.X • Programmable Pipeline only • OpenGL ES 3.X • Many more features, texture compression the big one! Michael Shah [1] http://www.khronos.org/opengles/

  16. OpenGL ES versus OpenGL • ES – Stands for embedded systems • Simpler API (Meaning, less functions implemented) • No geometry shader support for OpenGL ES 2.0 (or 3.0) • 2.0 version supports the GLSL 1.2 version we are using • Which again means no fixed-function pipeline Michael Shah

  17. WebGL • Javascript API – currently in version 2.0 • Identical to OpenGL ES • Runs in your browser • Programmable Pipeline only • No fixed-function pipeline available Michael Shah

  18. What Does it Look Like • WebGL Samples • http://webglplayground.net/ Michael Shah

  19. Direct3D (D3D) Microsoft’s Graphics API Michael Shah

  20. Microsoft’s Equivalent Graphics API • Very similar graphics API, but for Microsoft Platforms only • X-Box, PC, Zune, Surface, Windows Phone • Has High Level Shading Language(HLSL) similar to GLSL Michael Shah

  21. Which is Better? • Learn both • (Or the one your company is interviewing you for) • Use D3D for Microsoft platforms Michael Shah

  22. Managed Languages • For managed languages (Visual Basic, C#, other .net languages) • XNA Software Development Kit for Microsoft platforms) – Use for X-Box 360 programming Michael Shah

  23. GPGPU (General Purpose GPU) Languages CUDA OpenCL Compute Michael Shah

  24. CUDA Created by nVidia • Terminology • Device – GPU • Host – CPU • Kernel – Function that runs on device(GPU) • Gives us access to 100s or 1000s of processors on GPU • We write code in a kernel (think of like a shader) • One kernel executers at a time • Then 100s or 1000s of threads are spawned to compute result. • We can pre-compile the kernel! (But we must do so for each architecture we want to support) Michael Shah

  25. What GPU Looks like • Each Kernel gets a grid which has blocks • Each block has threads that share memory Michael Shah

  26. Example CUDA (Host) Michael Shah

  27. Example CUDA Kernel (Executed on Device) Michael Shah

  28. CUDA + OpenGL • Shared memory for all the buffer objects • So it’s still fast! It all lives on the Device (GPU) • Special CUDA Calls to register buffers • cuGLRegisterBufferObject(GLuintbufferobj ); • Simulations/programs where graphics are tied to data should use CUDA • e.g. Volume Rendering Michael Shah

  29. A Graphics Example [1] Michael Shah

  30. OpenCL • Performs same job as CUDA • Speed-wise • OpenCL and CUDA give you same benefits • OpenCL offers better portability • CUDA generally more tied to nVidia Michael Shah

  31. Compute • Microsoft’s version of CUDA or OpenCL • Programs written called “compute shaders” Michael Shah

  32. When to Use each? • Compute with DirectX suite • OpenCL with OpenGL and browser applications (webCL exists now!) • CUDA if only designing for one platform • CUDA is the most mature platform Michael Shah

  33. Future What to do in the next 1.5 weeks “Great Scott!” – Dr. Emmett Brown Michael Shah

  34. Where to Go • Finish your final project • Think about graphics applied to non-graphics applications • Check out the state of the art and build lots of little demoes Michael Shah

  35. Stuff to Check Out • Unity3D Michael Shah

  36. Stuff to Check Out • Unreal Engine (State of the Art) Michael Shah

  37. Renderman • Pixar’s tools • If you liked Ray tracer, you will want to take a look Michael Shah

  38. Next 15 minutes and Next Class • Project Updates in 1 minute • Next 15 minutes • Course evaluation (link) • Next Class • Mini in-class lab • Most likely VBO’s Michael Shah

  39. Supporting Slides In class questions come up Michael Shah

  40. Ambient Occlusion Map Michael Shah

More Related