Introduction - Advanced Graphics Deferred Rendering Introduction 6. Advanced Graphics 2012-2013 Fast...

Advanced Graphics 2012-2013 Advanced Graphics 2012-2013

Advanced Graphics

Introduction

Introduction

Advanced Graphics 2012-2013

Teacher

Marries van de Hoef

Under supervision of Arjan Egges and Frank van der Stappen

Introduction 2


Teacher

Graduating Master student

– Graphics research since Bachelor

– Three graphics research projects in Master

Introduction 3


Dynamic Lighting Architectures

Introduction 4


Reprojection 3D & Head Tracking

Introduction 5


Hybrid Deferred Rendering

Introduction 6


Fast Dynamic Radiosity

Introduction 7


Teacher

Graphics research internship at Nixxes – Worked on a major AAA-title

Assisted the Graphics course twice – Advised on practical assignment setup

– Gave multiple lectures

Much more I won’t bore you with

Introduction 8


COURSE OVERVIEW Introduction 9


Focus

Computer Graphics research area is very wide – Computer animation

– Geometry

– Simulation

– Rendering

– Much more…

This course focuses purely on Rendering

Introduction 10


Focus

The rendering research area can be divided into two extremes:

– Optimization for real-time performance

• Real-time global illumination

– Simulation of visual effects

• Realistic hair rendering

Introduction 11


Focus

This course leans towards the real-time side, but also covers the principles of simulation – Relevance to games (and game-tools)

Context of GPGPU programming – Will become increasingly more important

– Future of the GPU (and CPU?)

– Also useful for other areas

Introduction 12

GPGPU = General-Purpose computing on the GPU


Course Format

Meetings: – 8 lectures

– Guest lecture

– 8 paper presentation sessions

Presence is mandatory

Two practical assignments – Forum for questions

Introduction 13


Grading

Grading rules:

40% - Paper presentation

10%+40% - Practical assignments

10% - Quality of attendance

Need at least 5.0 for presentation and practical assignments

Introduction 14


COURSE DETAILS Introduction


Paper presentation

Form teams of two

– Post configuration on the forum

– Deadline: Saturday February 9th

– Required to continue with the course

– Can’t find a partner? Notify me

Introduction 16


Paper presentation

Paper and presentation slot are assigned

– Fully randomized

Paper subject extends on a lecture subject

Requirements and details on the website

– Read it before you start preparing

Introduction 17


Website

Download papers from the course website

Secured with password

– Username: student

– Password: the course code of this course

– (both are lower case)

Introduction


Practical Assignments

Have to be done individually Focus on GPGPU-style programming Additional features

– Large part of your grade – Free to choose – Creativity is encouraged – Specialize in your own interest

Introduction 19



Requirements

– C++

– Direct3D 11

– Visual Studio 2012

Works on DirectX 10 hardware and Windows 7

Work in BBL-175 or GameHall

Introduction 20



Forum

– Ask questions

– Answer questions

– Saves a lot of time

– And headaches

FORUM!!!!

Introduction 21


GPU Mandelbrot

Introduction 22


GPU Mandelbrot

Just a warming up Create it from scratch

– Learn what’s really possible

A lot of work

– The deadline is soon

Read the assignment document

Introduction 23


GPU Path Tracer

Introduction 24


GPU Path Tracer

Practice advanced shader programming & learn the practical side of path tracing

Relevant for games – GPGPU-style programming

– Rays are commonly used in other techniques

– Several smaller parts directly usable in games

Theoretically more complex than Mandelbrot

Introduction 25


Forum

http://www.uu.nl/blackboard

Click on the Advanced Graphics course – Post your paper presentation team

• In the Paper Presentation forum

– Ask/answer questions about the assignments…

– …or lectures

– Course announcements

Introduction 26




Points

Gamification

Several ways to earn points during the course

Small competition amongst you and your friends

– Also converted to a small part of your grade

Introduction 27


Points

Ways to earn points

– Questions about the previous lecture

• Each meeting will start with 10 random students each receiving 1 question. You have to study!

– Find errors in the slides

• Real mistakes, not disputable things or typos

– Help other students on the forum

• Top two helpers get points

Introduction 28


Points

Conversion to points

– Quality of Attendance term (10%)

– Maximum QoA grade is 15 (out of 10)

• 5 bonus points compensate for any bad luck

– Maximum final grade is 10.5

Introduction 29


Points

Calculation of QoA grade Lecture questions always give 1 opportunity

– And only a point when you are correct (and present)

Other points earned increment both variables equally

Introduction 30

uint point, opportunity; float qoaGrade = point / (float)opportunity * 15;


Points

10 questions at the start of each meeting – Be on time! – No peeking and no telling – Student selection is fully random

• You could be selected twice in one meeting

Procedure – A student is selected (you) – Stand up – The slide with the question is shown – Think for a few seconds and answer – The correct answer is revealed

Introduction 31


Points

Practice 3 questions

– Not for real

– Only easy questions

– Real questions will focus on facts (and insight)

– Hard questions (roughly equal difficulty)

• You must study!

– Thursday will be for real

• About the hardware part of this lecture

Introduction 32


Practice Questions

Introduction 33


Question 1

What is the first name of the teacher?

Introduction 34


Question 2

Which one is not a graphics (related) API?

a) Direct3D

b) OpenGL

c) Havok

d) CUDA

e) Direct2D

Introduction 35


Question 3

Given two perpendicular 3D vectors, how to make an orthogonal basis?

Introduction 36


Questions

Introduction 37


Advanced Graphics

Hardware

Hardware


Parallel Data Processing

We need to process a lot of data for graphics – Over 200 GB per second

– Using over 3TFLOPS

Requires a different programming architecture

Get a feeling for how this works – Important to write fast graphics code

Hardware 39



Requirements for massive data processing

– Massive processing power

– Massive bandwidth to/from memory

• Deal with memory latency

We can do with simple processing units

– Limited instruction set

– Share instruction stream

Hardware 40



Use architecture based on SIMD – Single Instruction, Multiple Data

Good when working with Vector3 (add, multiply, etc.) – Same instruction executed for all 3 components anyway

Can we go “wider”? – What if we can execute an instruction for

8 components simultaneously?

Hardware 41


GPU’s don’t parallelize on Vector3 level – But instead on pixel/vertex level

Execute the same shader instructions

for all threads (pixels, vertices, etc.) – Also called SIMT

Keeps functional units slim

– And parallelization manageable


Hardware 42

Instruction feeding

F F F F

F F F F

F F F F

F F F F

Local memory

1 Core


Parallel Hardware

Multiple functional units form a core – up to 64

Multiple cores on a GPU – up to 32

Conceptual number of threads executed simultaneously (warp)

– 16 (Intel) – 32 (Nvidia – warp) – 64 (AMD – wavefront)

(Disclaimer: many exceptions)

Not necessarily 1 thread per functional unit Some GPU’s execute multiple instructions simultaneously

Hardware 43


Parallel Hardware

Consequences of thread-level SIMD

– Vector4 operations cost more than Vector3

– What about branching code?

Branches conflict with thread-level SIMD

– No good solution

– Execute both (diverging) branches for all threads

• Ignoring the results of the wrong branch

Hardware 44


Threading SIMD Example

How can we program using this architecture?

Calculate the sum of 16 numbers in a list

– Assume a warp size of 8

Hardware 45

1 2 3 4 5 6 7 8


Memory Latency

Waiting for memory takes very long

– It takes 400-800 processor cycles!

– Optimized for throughput at the cost of latency

Don’t wait, do something else

– Work on another warp

– …repeat if that warp needs memory access as well

Extended form of CPU hyper-threading

Hardware 46


Memory Latency

Hardware 47

Warp 1

Warp 1

Stall Warp 2

Stall Warp 3

Stall Warp 4

Stall


Memory Latency

Warp contexts are stored in local memory – Contains state for all functional units

– Context size depends on the shader

If the shader requires many temporary registers – Fewer warps, thus less latency hiding

Hardware 48

Instruction feeding

F F F F

F F F F

F F F F

F F F F

Local memory

1 Core

Ctx Ctx Ctx

Ctx Ctx Ctx


Bandwidth Bottleneck

Latency hiding is great! No more stalling, right?

– Still limited by the available bandwidth

Very common bottleneck

– Bandwidth is more precious than computing power!

– If you can avoid reading a texture by doing more calculations: do it!

Hardware 49


CPU SIMD

CPU also has SIMD capabilities

– SSE instructions

– Not threaded SIMD like the GPU

Can operate on 4 floats simultaneously

– Vector4 + Vector4 is 1 operation

– Always on 4 floats, even if you use 3 floats

Hardware 50


Simple SIMD Example

Calculate the sum of 16 numbers in a list

– Store the numbers in 4 SSE vectors

– Add the vectors

– Use special instructions to do internal addition

Hardware 51

+ + + =


CPU SIMD

SSE caveats

– Expensive to move data in/out of SSE registers

– Storage requires memory alignment

New AVX instructions double SIMD width

– Operations on 8 floats simultaneously!

– How can we use this for graphics?

Hardware 52


INTERACTING WITH THE GPU Hardware 53


Graphics API

Use Graphics API to access GPU – Well defined programming interface

– Agreement between software developer and hardware vendor

Main Graphics API’s – Direct3D

– OpenGL

Hardware


Direct3D vs. OpenGL

Religious war

OpenGL used to be “behind” Direct3D, however…

– Rise of mobile devices and Mac

– Direct3D hasn’t had new real features since 2009

Hardware 55


Direct3D

We will use Direct3D

– Logic choice after using XNA

– Easier for hardware requirements

– Object oriented structure

Hardware 56


Direct3D

Important Direct3D11 objects – DXGI

– ID3D11Device

– ID3D11DeviceContext

Matrices – Can use both Left/Right handed coordinate systems

– HLSL (Shader) default is column-major matrices

– DirectXMath/XNA is always row-major matrices

Hardware 57


Below Direct3D

Hardware 58

Game

D3D11 User

Mode Driver

DXGI

KernelMode Driver

GPU PCI express


GPU Architecture

Graphics API requires order preservation

– But the GPU is massively parallel!

Hardware 59

workload

Buffer reorder

Buffer reorder …etc

(Vertex Shader) (Rasterizer) (Pixel shader…)


GPU Architecture

Hardware 60


GPU Architecture

Hardware 61


GPU Architecture

Hardware 62


GRAPHICS PIPELINE Hardware 63


– Input Assembler

– Vertex Shader

– Primitive Assembler

Pipeline overview – XNA/D3D9

Hardware

VS IA PA RS PS OM

Memory

– Rasterizer

– Pixel Shader

– Output Merger


– Input Assembler

– Vertex Shader


Pipeline overview – D3D10

Hardware

VS IA PA RS PS OM

Memory

– Geometry Shader

– Stream Output

– Rasterizer

– Pixel Shader

– Output Merger

GS

SO


– Input Assembler

– Vertex Shader


– Hull Shader

– Tesselator

– Domain Shader

Pipeline overview – D3D11

Hardware

VS IA PA HS TS DS GS

SO

RS PS OM CS

Memory

– Geometry Shader

– Stream Output

– Rasterizer

– Pixel Shader

– Output Merger

– Compute Shader

PA


Rasterizer

Hierarchical and parallel rasterization – Test for each rectangle/pixel if inside triangle

Interwoven with hierarchical early-z testing – Test depth buffer before executing pixel shader – Be careful when changing depth/depth settings

Hardware 67


Rasterizer

Projection is not a linear transformation – Divide by w component – After interpolation over triangle – Perspective division is always done!

Perspective correct interpolation

Hardware 68


Rasterizer

Vertex shader outputs “homogeneous clip space”

After persp. division: Normalized Device Coordinates

Viewport transform brings to pixel coordinates – Applies 0.5 pixel offset (DX10+)

Hardware 69


Pixel Shader

Always executed in groups of 2x2

– Required for easy uv derivatives (for mipmapping)

– Some pixels might not be visible

– Small/thin triangles are a waste!

The shader compiler is an aggressive optimizer

– But don’t count on it

Hardware 70


Compute Shader

Additional features

– Write at random positions in UAV (Unordered)

– Synchronization primitives

– Some atomic operations

– Thread Group Shared Memory

Threading SIMD program

Hardware 71


Forum

Reminder:

– Check out the forum

– Work on first practical assignment

– Make paper presentation teams

• Post it on the forum

FORUM FORUM FORUM FORUM FORUM

Hardware 72

Introduction - Advanced Graphics Deferred Rendering Introduction 6. Advanced Graphics 2012-2013 Fast...

Documents

Transcript of Introduction - Advanced Graphics Deferred Rendering Introduction 6. Advanced Graphics 2012-2013 Fast...