The University of North Carolina - Chapel Hill COMP Spring GPGP - Background Lecture Graphics Pipelines, OpenGL, Fragment Programs, First GPGP Programs
The University of North Carolina - Chapel Hill COMP Spring Purpose Give sufficient background to non- graphics students to program a simple GPGP program –Attempted to minimize any assumptions prior knowledge of Graphics
The University of North Carolina - Chapel Hill COMP Spring My Background in GPGP 06/ Worked with Bill Baxter on paint work using GPU to transform volume representation of paint to RGB 01/ Worked with Brandon Lloyd on shadow work –Didn’t touch GPU work, but understood what it did 09/ Taught COMP136 and we talked about GPU 10/ Wrote a simple program to remove radial distortion from images in real time 01/ Prepared for this lecture
The University of North Carolina - Chapel Hill COMP Spring Overview Raster Graphics –Where did this start? Graphics Pipeline –What’s the hardware like? OpenGL –How to talk to the hardware? GPU Programming –How to make it do something non-standard? A First GPGP Program –Hello World!
The University of North Carolina - Chapel Hill COMP Spring Raster Graphics -- “Utah” Rendering 1960s-1970s The University of Utah led raster graphics research –Ivan Sutherland How do we convert mathematical representations of objects into images on the screen? –Conversion of continuous to discrete –Solution of light-surface-eye interactions
The University of North Carolina - Chapel Hill COMP Spring Object Representations MANY continuous representations exist for objects –Planar polygons, quadrics, splines, M- Reps, general equations … Discrete representations are sparser –Generally some interpretation of an array of numbers
The University of North Carolina - Chapel Hill COMP Spring The Winner! In modern raster graphics, the triangle is king –It is the simplest continuous representation that can approximate all other surface-types –All other continuous representations must be triangulated before being rasterized Unless GPGP is used!
The University of North Carolina - Chapel Hill COMP Spring Rasterization Terms Tessellate - convert some other continuous representation to planar polygons Triangulate - converting some other continuous representation to triangles Vertex - a point in some nD space Rasterize - convert a triangle* to fragments Fragment - a potential pixel**
The University of North Carolina - Chapel Hill COMP Spring “Utah” Graphics Pipeline Obtain triangulation of model Affine Transforms Projective Transforms Clip to viewable region Rasterize
The University of North Carolina - Chapel Hill COMP Spring Raster Graphics –Where did this start? Graphics Pipeline –What’s the hardware like? OpenGL –How to talk to the hardware? GPU Programming –How to make it do something non-standard? A First GPGP Program –Hello World! Overview
The University of North Carolina - Chapel Hill COMP Spring Why Specialty Hardware? CPU can do all Turing complete operations
The University of North Carolina - Chapel Hill COMP Spring Earliest Commodity Hardware Vertex Transformation Clipping Rasterization Fragment Operations Visibility State of pipeline initialized Vertices come down the pipe Framebuffer and depth buffer set by the end
The University of North Carolina - Chapel Hill COMP Spring Cutting-edge Commodity Hardware Okay, not much changed Orange denotes programmability Power of standard set of settings increased Output can go to any/many buffers Vertex Transformation Clipping Rasterization Fragment Operations Visibility
The University of North Carolina - Chapel Hill COMP Spring Overview Raster Graphics –Where did this start? Graphics Pipeline –What’s the hardware like? OpenGL –How to talk to the hardware? GPU Programming –How to make it do something non-standard? A First GPGP Program –Hello World!
The University of North Carolina - Chapel Hill COMP Spring Talking to the GPU Two parts: –STATE: The majority of OpenGL calls modify the state machine –INPUT: Vertices Three vertices make a triangle Once a triangle is complete, the GPU runs with it
The University of North Carolina - Chapel Hill COMP Spring OpenGL Overview Most of the features in OpenGL will probably never be used in this class For the majority of GPGP work, you render a quad (two triangles) that fills the screen on a one-input- texture-to-one-output-pixel basis
The University of North Carolina - Chapel Hill COMP Spring Note on OpenGL Although OpenGL calls are supported by nVIDIA or ATi drivers, some windowing system must be used –Native to OS - a pain –GLUT - quick, easy, small, has some issues with the “nicities of coding” –Almost all windowing toolkits support OpenGL FLTK, Qt, WxWindows, etc.
The University of North Carolina - Chapel Hill COMP Spring Let’s go to the code All code is available at asses/GPGPLecture asses/GPGPLecture
The University of North Carolina - Chapel Hill COMP Spring OpenGL Gnitty-Gritties Three more important OpenGL features –Multi-pass rendering –Read-backs –Extensions
The University of North Carolina - Chapel Hill COMP Spring Overview Raster Graphics –Where did this start? Graphics Pipeline –What’s the hardware like? OpenGL –How to talk to the hardware? GPU Programming –How to make it do something non-standard? A First GPGP Program –Hello World!
The University of North Carolina - Chapel Hill COMP Spring History of Commodity GPU Programming Pre Basic rasterizers –Some texture combination ability –Vertex transformation occurs on CPU Slightly configurable –Cube maps, signed math ops –Vertex transforms added to GPU Vertex programmability 2002-present - Vertex/fragment programmability
The University of North Carolina - Chapel Hill COMP Spring History of GPU non- Commodity Programmability mid-1990’s - UNC PixelFlow later-1990’s - Stanford RTSL
The University of North Carolina - Chapel Hill COMP Spring GPU Programming Languages Assembly language Cg and HLSL GLSL
The University of North Carolina - Chapel Hill COMP Spring Types of GPU Programs Vertex Programs –Required Outputs: Vertex position and Vertex color –Optional Outputs: Hardware/language dependant maximum number of output values Fragment Programs –Required Outputs: RGBA color –Optional Outputs: Writing to multiple sources
The University of North Carolina - Chapel Hill COMP Spring Communicating with GPU Programs There are two ways of sending information to GPU Programs: –Explicitly setting parameters using specified function calls –Sending down standard values by setting OpenGL state
The University of North Carolina - Chapel Hill COMP Spring For More Info Tutorials, sample code, etc. –Go to Cg Tutorial –Amazon: Tutorial-Definitive-Programmable-Real- Time/dp/ http:// Tutorial-Definitive-Programmable-Real- Time/dp/ GLSL (Orange Book) –Amazon: Shading-Language- 2nd/dp/ /sr=1- 1/qid= /ref=pd_bbs_1/ ?ie=UTF8&s=bookshttp:// Shading-Language- 2nd/dp/ /sr=1- 1/qid= /ref=pd_bbs_1/ ?ie=UTF8&s=books
The University of North Carolina - Chapel Hill COMP Spring Overview Raster Graphics –Where did this start? Graphics Pipeline –What’s the hardware like? OpenGL –How to talk to the hardware? GPU Programming –How to make it do something non-standard? A First GPGP Program –Hello World!
The University of North Carolina - Chapel Hill COMP Spring Let’s go to the code Borrowed heavily from gpgpu.org/developer All code is available at asses/GPGPLecture asses/GPGPLecture
The University of North Carolina - Chapel Hill COMP Spring Notes No Vertex program! … no use for one. The framebuffer-to-texture transfers we were doing are slow Use the framebuffer object class available from GPGPU.org/developer GLEW is downloadable from glew.sourceforge.net/ We only passed one parameter down in this example
The University of North Carolina - Chapel Hill COMP Spring Reading the Data Back to the CPU See function SnapShot at the bottom of the last file
The University of North Carolina - Chapel Hill COMP Spring Debugging IMDebug by Bill Baxter: – mdebug/index.html
The University of North Carolina - Chapel Hill COMP Spring Questions?