2005.6.2 Jiwon Choo

24
The Setup for Triangle The Setup for Triangle Rasterization Rasterization By Andres Kugler Eurographics ’96 By Andres Kugler Eurographics ’96 Computer Graphics Lab in Tubingen Univ Computer Graphics Lab in Tubingen Univ 2005.6.2 Jiwon Choo

description

The Setup for Triangle Rasterization By Andres Kugler Eurographics ’ 96 Computer Graphics Lab in Tubingen Univ. 2005.6.2 Jiwon Choo. Contents. Abstract & Introduction The Graphic Rendering Pipeline Scan Conversion pipeline Initialization Data Latency Intensive Computation - PowerPoint PPT Presentation

Transcript of 2005.6.2 Jiwon Choo

Page 1: 2005.6.2 Jiwon Choo

The Setup for Triangle RasterizaThe Setup for Triangle Rasterizationtion

By Andres Kugler Eurographics ’96By Andres Kugler Eurographics ’96Computer Graphics Lab in Tubingen UnivComputer Graphics Lab in Tubingen Univ

2005.6.2Jiwon Choo

Page 2: 2005.6.2 Jiwon Choo

Contents

Abstract & Introduction The Graphic Rendering Pipeline Scan Conversion pipeline Initialization Data Latency Intensive Computation Setup Architecture and Bandwidth Conclusion

Page 3: 2005.6.2 Jiwon Choo

Abstract

Integrating Slope and setup calculation to rasterizer increase 3D system performance host processing is greatly reduced Data flow from host to sub system is also

greatly reduced.

Page 4: 2005.6.2 Jiwon Choo

Intoroduction

Rasterizer Performance Bottleneck

world coordinate transformation to screen coordinate and computing vertex colors

calculating the triangle edge slope and increment necessary for the scan-conversion

the rate at which the fixed point iterators generate pixel values

the achieved bandwidth into the frame buffer and texture memory

Page 5: 2005.6.2 Jiwon Choo

The Graphic Rendering Pipeline

Page 6: 2005.6.2 Jiwon Choo

The Graphic Rendering Pipeline

Page 7: 2005.6.2 Jiwon Choo

The Graphic Rendering Pipeline

Page 8: 2005.6.2 Jiwon Choo

Scan Conversion pipeline

Page 9: 2005.6.2 Jiwon Choo
Page 10: 2005.6.2 Jiwon Choo
Page 11: 2005.6.2 Jiwon Choo

Initialization Data

Page 12: 2005.6.2 Jiwon Choo

Initialization Data

Page 13: 2005.6.2 Jiwon Choo

Latency Intensive Computation

Page 14: 2005.6.2 Jiwon Choo

Latency Intensive Computation

Multiplication easily parallelized by breaking the

multiplication of two values in two, or four individual multiplication and summing the partial results.

Needs 4 pipeline stage

Page 15: 2005.6.2 Jiwon Choo

Latency Intensive Computation

Division Common way: subtractive and multiplicative method.

pj+1 = rpj – qj+1.D Needs N pipeline stage

Faster way: series of expansion of the reciprocal. Newton-Raphson implementation xi+1 = xi .(2-b.xi) Needs 10 pipeline stage Initial approximation for reciprocal is taken from a

ROM Look-up-table - LUT Additionally guard bit reduce latency & size of rasterizer

Page 16: 2005.6.2 Jiwon Choo

Latency Intensive Computation

Page 17: 2005.6.2 Jiwon Choo

Latency Intensive Computation

Page 18: 2005.6.2 Jiwon Choo

Setup Architecture and Bandwidth

Page 19: 2005.6.2 Jiwon Choo

Setup Architecture and Bandwidth

Page 20: 2005.6.2 Jiwon Choo

Setup Architecture and Bandwidth

Page 21: 2005.6.2 Jiwon Choo

Setup Architecture and Bandwidth

Input stack Register File with Setup Values

Reciprocal evaluation for color, depth, texture coordinate increment

Gradient computation Slope calculation

Page 22: 2005.6.2 Jiwon Choo

Setup Architecture and Bandwidth

Modeling the throughput of setup Pipeline Setup & raster pipeline clock 80MHz. Producing 1 rendered pixel per cycle. Standard VRAM&DRAM are used for frame buffer and z-buffer. Page mode with a memory controller operating at 80MHz Z buffer has 10way inter leaving

Page 23: 2005.6.2 Jiwon Choo

Conclusion

Page 24: 2005.6.2 Jiwon Choo

Conclusion

Moving the setup from the geometry processor to the rasterizer greatly improves the available input

bandwidth to the rasterizer virtually permits to send triangles to the

setup at the same rate as they can be written into the frame buffer by rasterizer.