A Sorting Classification of Parallel Rendering

Molnar et al., 1994

What’s this about? Describes a classification scheme for

comparing parallel rendering systems Based on where the sort from object

coordinates to screen coordinates occurs

Supports analysis of computational and communication costs and encompasses current and proposed highly parallel renderers (both hardware and software)

Parallel rendering as a sorting problem The classification scheme is for “Fully Parallel” systems, i.e.

rendering rate is high enough that geometry processing and rasterization must be done in parallel

geometry processing (transformation, clipping, lighting) rasterization (scan conversion, shading, visibility determination)

Usually geometry is parallelized by assigning each processor a subset of primitives (objects) in a scene

And rasterization is parallelized by assigning each processor a portion of the pixel calculations

Mainly by rendering we are trying to calculate the effect each primitive (object) has on a pixel

Since a primitive may be on or off the screen, rendering can be viewed as a problem of sorting primitives to the screen [Sutherland]

For “fully parallel” renders, sorting involves a redistribution of data between processors because the responsibility for primitives and pixels is distributed

Parallel rendering (cont) The location of the sort determines the

structure of the resulting parallel rendering system and in general the sort can take place anywhere in the rendering pipeline. sort-first : sort during geometry processing

redistributing “raw” primitives before their screen space parameters are known

sort-middle : sort between geometry processing and rasterization

redistributing screen-space primitives sort-last : sort during rasterization

redistributing pixels, samples or pixel fragments

Sort-first No known implementations? Goal is to distribute primitives early in the rendering pipeline to

processors that can do the remaining rendering calculations. Done by dividing the screen into disjoints regions and making

processors (renderers) responsible for all rendering calculations that affect their screen region.

Initially, primitives are assigned to renderers in an arbitrary fashion.

At the start of rendering, each renderer does enough transformation to determine which region(s) each primitive falls. This is pre-transformation.

If a renderer contains a primitive which does not belong in it’s screen region, the primitive is redistributed over an interconnect network to the appropriate renderer(s) which perform the remaining geometry processing and rasterization for the primitive.

Redistribution (at the beginning of rendering) is the distinguishing feature of sort-first.

Analysis Advantages:

Communication requirements are low when the tessellation ratio and the degree of oversampling are high, or when frame-to-frame coherence can be exploited

Processing nodes implement the entire rendering pipeline for a portion of the screen

Disadvantages: Susceptibility to load imbalance. Primitives may clump

into regions, concentrating the work on a few renders. Necessity of retained mode and complex data handling

code to take advantage of frame-to-frame coherence.

Sort-middle Most common approach. Primitives are redistributed in the middle of the rendering

pipeline between geometry processing and rasterization. At the point of redistribution, primitives have been

transformed into screen coordinates and are ready for rasterization.

Many systems use separate processors for geometry processing and rasterization so this is a natural point to divide the pipeline.

In sort-middle, geometry processors are assigned arbitrary subsets of the primitives to be displayed and rasterizer(s) are assigned a portion of the display screen

During each frame, geometry processors transform, light, etc their portion of the primitives and classify them with respect to screen region boundaries.

Then they transmit all these screen-space primitives to the appropriate rasterizer(s).

Analysis Advantages:

General and straightforward Redistribution occurs at natural place in the

pipeline

Disadvantages: High communication costs if the tessellation

ratio is high Susceptibility to load imbalance between

rasterizers when primitives are distributed unevenly over the screen

Sort-last Sorting deferred until end of rendering pipeline,

after primitives have been rasterized Each renderer is assigned an arbitrary subset of

the primitives Each renderer computes pixel values for its

primitives Pixel values are then transmitted over interconnect

network to compositing processors which resolve the visibility of the pixels from each renderer

Two approaches: SL-sparse: minimizes communication by distributing only

those pixels actuall produced by rasterization SL-full: stores and transfers a full image from each

renderer

Analysis Advantages:

Processing nodes implement the entire rendering pipeline for a portion of the primitives

Less prone to load imbalance SL-full merging can be embedded in a linear

network, making it linearly scalable

Disadvantages: Pixel traffic can be extremely high

Talisman: Commodity Realtime 3D Graphics for the

PC

Jay Torborg and Jim Kajiya, Microsoft Corporation

1996

What is it?

New architecture for 3D graphics Cost $200-$300 Requirements? Smooth motion,

synchronized with sound and video and low-latency interaction (want real-time at 72-85Hz)

Limitations of Traditional Architectures

High Memory Bandwidth Requirement

System Latency Cost/Memory Cost

Composited Image Layers Independent image layer for each non-

interpenetrating object in the scene Each object can be updated independently (optimize

updates based on priority) Layers can be arbitrary size and shape

Image layers are composited at video rates Support image layer transformations at video

rates (scaling, rotation) Typically, the same rendered image can be

used for 4 frames

Image Compression

Used for textures and image layers Lossless and lossy compression

supported Significantly reduces bandwidth

and capacity requirements 16:1 texture compression 5:1 image layer compression

Chunking Each image layer is rendered in 32x32

chunks All polygons for a 32x32 chunk are

rendered before proceeding to next chunk Allows 32x32 depth buffer to be on-chip

Anti-aliasing supported with depth buffering and translucency using on-chip fragment buffer

High Quality Rendering

Anisotropic filtering of textures Multipass rendering

Shadows, spot lights, fog Antialiasing

Reference Hardware Targets high-end consumer PC market Uses PCI expansion bus Replaces:

Windows accelerator board 3D accelerator board MPEG playback board Video conferencing board Sound board modem

Polygon Object Processor Polygons are processed in 32x32 chunks Initial Evaluation

Computes intersection of a chunk with a triangle and computes the values for color, transparency, depth and texture coordinates for the starting point of the triangle within the chunk

Pixel Engine Performs pixel level calculations (compositing, depth

buffering, fragment generation) for pixels which are partially covered

Fragment Resolve Performs final anti-aliasing by resolving depth sorted

pixel fragments with partial coverage or transparency

Image Layer Compositor Responsible for generating the graphics

output from a collection of depth sorted image layers

Locked to the video refresh Data structure maintains z-sorted list of

image layers visible in each 32 scanline region

Performs affine transforms (scaling, rotation)

Passes pixel and alpha data to compositing buffer at four pixels per clock cycle

Compositing Buffer

Simple specialty memory Contains two 32-scanline buffers

for double buffering of scanline regions one sl buffer for compositing the

other for display

Software

DirectDraw, Direct3D, DirectSound Media DSP Real-time kernel on DSP for

scheduling and load balancing

Performance

High Resolution Display 1344x1024 @ 75Hz

24 bit color at all resolutions 20-30k polygon scene complexity 40 MPix/sec w/ anisotropic

texturing and anti-aliasing