seminar ii: rendering architectures yan cui love joy mendoza oscar kozlowski john tang

23
Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Upload: barnaby-owens

Post on 16-Jan-2016

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Seminar II:Rendering Architectures

Yan CuiLove Joy MendozaOscar KozlowskiJohn Tang

Page 2: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Contents

• Graphics Pipeline

• Current Consumer Renderer Architectures

• Parallelisation of Rendering Pipeline

• Synchronisation issues in parallelisation

• Parallel and Cluster Architectures examples

Page 3: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Graphics Pipeline

• Geometric TransformationMapping of triangles from a 3D coordinate system

(object space) to a 2D coordinate system (image space).

• Rasterisation

Converts transformed triangles into pixel values to be shown on the computer screen.

Page 4: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Graphics Pipeline

Model andViewing

TransformLighting

ProjectionTransform

ClippingPerspective

DivisionViewportMapping

Modelling transformation positions primitives with respect to each other.Viewing transformation orients the resulting primitives to the user viewpoint.

Evaluates the colour of the vertices.Projects objects onto the screenHides primitives not included in the viewing frustrum.Converts the vertex to Cartesian coordinates. Performs final scaling and translation.

Geometric transformation

Page 5: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Graphics Pipeline

AlphaBlending

Rasterisation

TextureMapping

DepthTest

Scan Conversion

Decomposes triangle into a set of pixels & calculates the attributes of each pixel.

Wraps a 2D texture image on the surface of a 3D object.Compares the value of the current pixel against the depth value of the pixel at the corresponding XY coordinate of the frame buffer.

Colour of old & new pixels is composited according to their alpha value.

Page 6: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Consumer PC Architecture

GeForce 7800 GTX

• Clock Freq. 430Mhz• 256Mb DDR3 VRAM• 24 Pixel Shaders• 8 Vertex Shaders• Over 300 billion floating

point operations per sec.• 44.4 Gb/s Texture

bandwidth

Page 7: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Parallelisation

• Parallel architectures– Multiple processors/renderers on

a close-proximity, low-latency bus– Becoming common in consumer

market currently:• HyperThreading• SLI

• Cluster Architecture– Multiple computers networked– Use of low cost, consumer PC’s

“SGI Graphics Cluster™: The Cluster Architecture Challenges, the SGITM Solution”, SGI, 2001

Page 8: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Parallelisation Strategies- Sort middle

• Natural way to parallelise

• No overhead on geometry

computation

• Access to renderer memory required

– Specialised renderers required

• Network requirements depend on:

– Number of primitives

– Amount of overlap between tiles

Page 9: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Parallelisation Strategies- Sort first

• Extra bounding box calculation

• Overhead on geometry computation and rasterisation due to primitives overlapping multiple screen tiles

– However, temporal coherence

• Consumer renderers useable

• Lower network usage

• Load balancing difficult

Page 10: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Parallelisation Strategies- Sort last

• No overhead in geometry and

rasterisation

• Load balancing easily possible

• High-bandwidth network required

– Sparse: send only rendered pixels

– Full: send full image

• Compositor design is difficult

• Transparency is almost impossible

• Anti-aliasing is very expensive

Page 11: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Synchronisation issues

• All rendering pipelines must have synchronised input data– Implied for shared memory parallel system– Cluster architectures require same solution

as distributed databases

• All rendering pipelines must synchronise render output depending on display type– Not needed for polarised display

• Failure to synchronise results in incorrect rendered image

• SWAPBUFFERS

Page 12: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Dynamic data synchronization

• Two types of dynamic data:– control information,– changing/dynamic data set information

• Dynamic Data generated from raw real time Stimulus Data i.e. input devices

• SYNC: Ensure images on each node are computed on coherent data sets.

• Block until everyone accepts stimulus data

Page 13: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Video Synchronisation

• Output to a Display driven by Signal from each graphics card

• Signal provides:– Display Image– Synchronisation

• SYNC: Ensure video signals are synchronised

• Signal and it’s synchronisation are not something that can be controlled via software on commodity graphics cards

Page 14: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Video Synchronisation

• Genlock (high-end apps)– Most precise way of ensuring synchronisation– Here the graphics system ensures pixel-level

synchronisation by using a PLL to lock onto the line rate to derive the pixel rate (or pixel clock)

– Lock is fine enough to allow phase adjustments for each pixel

• Frame Lock (low-end apps)– Less-precise method– Synchronises once per frame at the end of the blanking

period

Page 15: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Cluster Architecture

• Pixel Planes– Heterogeneous multi-computer system– Parallel processors

Page 16: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Cluster Architecture

Each PixelPlanes renderer unit consists of

Page 17: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

PixelPlanes Render

Rendered in 1 second

Page 18: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Parallel Architecture

• Silicon Graphics InfiniteReality architecture– third-generation graphics system

– designed to deliver 60Hz steady frame rate high-quality rendering of complex scenes

– support for OpenGL

– boardset consists of 3 distinct board types:• Geometry• Raster memory• Display generator

Page 19: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

InfiniteReality Architecture

Geometry pipeline

Geometry distributor (sort-middle)

Rasteriser

Page 20: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

InfiniteReality Render

Page 21: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Clustering System Examples

• Stanford's Chromium Toolkit• Fraunhofer Institute for Industrial Engineering

IAO's HiPI-6 (mature installation)• ARS Electronica ARSBox (commercial)• Solutions based on VRJuggler (opensource)

– ClusterJuggler– NetJuggler

• Unreal Tournament CAVE

Page 22: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Summary

• Graphics Pipeline• PC, Parallel and Cluster architectures• Clusters of consumer Graphics cards starting to replace

specialised parallel architectures due to cost and availability

• Parallelisation strategies– Sort-middle– Sort-first– Sort-last

• Issues of synchronisation and how to resolve– Input data synchronisation– Output render synchronisation

Page 23: Seminar II: Rendering Architectures Yan Cui Love Joy Mendoza Oscar Kozlowski John Tang

Further Reading

• “Three-dimensional computer graphics Architecture”, Mitra et Al., 2000

• “An Overview of Cluster Solutions for Immersive Displays”, Steed et. al, (http://www.cs.ucl.ac.uk/staff/a.steed/cluster-solutions-with-figures.htm)

• SGI Graphics Cluster™: The Cluster Architecture Challenges, the SGI™ Solution (http://www.sgi.com/products/legacy/pdf/architecture.pdf)