A Crash Course on Programmable Graphics Hardware LiYi

A Crash Course on Programmable Graphics Hardware Li-Yi Wei 2005 at Tsinghua University, Beijing

Why do we need graphics hardware?

The evolution of graphics hardware SGI Origin 3400 NVIDIA Geforce 7800

7 years of graphics accelenation. com/? doc=123&page=1

Ray tracing General & flexible Intuitive Global illumination Hard to accelerate

Polygonal graphics pipeline Local computation Easy to accelerate Not general Unintuitive

Graphics hierarchy Layered approach Like network layers Encapsulation Easy programming Driver optimization Driver workaround Driver simulation Protection Hardware error check

Overview Graphics pipeline Only high level overview (so you can program), not necessarily real hardware GPU programming

Graphics pipeline

Application Mostly on CPU High level work User interface Control Simulation Physics Artificial intelligence

Host Gatekeeper of GPU Command processing Error checking State management Context switch

Geometry Vertex processor Primitive assembly Clip & cull Viewport transform

Vertex Processor Process one vertex at one time No information on other vertices Programmable Transformation Lighting

Transformation Global to eye coordinate system

Lighting Diffuse Specular

Transform & Light on Vertex Processor A sequence of assembly instructions (more on this later)

Primitive Assembly Assemble individual vertices into triangle (or line or point) Performance implication A triangle is ready only when all 3 vertices are Vertex coherence & caching

Clipping & Culling Backface culling Remove triangles facing away from view Eliminate ½ of the triangles in theory Clipping against view frustum Triangles may become quadrilaterals

Viewport transform From floating point range [-1, 1] x [-1, 1] to integer range [0, height-1] x [0, width-1]

Rasterization Convert primitives (triangles, lines) into pixels Barycentric coordinate Attribute interpolation

Triangles into pixels

Attribute interpolation Interpolation Barycentric

Perspective correct interpolation incorrect

Fragment processor Fragment: corresponds to a single pixel and includes color, depth, and sometimes texture-coordinate values. Compute color and depth for each pixel Most interesting part of GPU

Texture Optional (though hard to avoid) Cache data Hide latency from FB Sampling/filtering I told you this last time

ROP (Raster Operation) Write to framebuffer Comparison Z, stencil, alpha, window

Framebuffer Storing buffers and textures Connect to display Characteristics Size Bandwidth Latency

Conceptual programming model Inputs (read-only) Attributes Constants Textures Registers (read-write) Used by shader Outputs (write-only) Attributes

Simple example HPOS: position COL 0: diffuse color MOV o[HPOS], v[HPOS]; MOV o[COL 0], v[COL 0];

More complex example o[COL 0] = v[COL 0] + constant*v[HPOS]; MOV o[HPOS], v[HPOS]; MOV R 0, v[COL 0]; MAD R 0, v[HPOS], c[0], R 0; MOV o[COL 0], R 0;

Sample instruction set

A real example

High-level shading language Writing assembly is Painful Not portable Not optimize-able High level shading language solves these Cg, HLSL

Cg example

Applications Too many of them for me to describe here The only way to learn is try to program Useless for you even if I try to describe Look at developer website NVIDIA, ATI, GPGPU

Homework Try to program GPU! Even without NVIDIA GPU, you can download the emulator Stanford course on graphics hardware http: //www. graphics. stanford. edu/courses/cs 448 a-01 -fall/ History of graphics hardware 7 years of graphics accelenation. com/? doc=123&page=1