+ All Categories
Home > Documents > Hardware Implementation of Micropolygon Rasterization with...

Hardware Implementation of Micropolygon Rasterization with...

Date post: 30-Mar-2021
Category:
Upload: others
View: 5 times
Download: 0 times
Share this document with a friend
1
Hardware Implementation of Micropolygon Rasterization with Motion Blur and Depth Of Field ABSTRACT Current GPUs rasterize micropolygons (polygons approximately one pixel in size) inefficiently. Additionally, modern systems do not natively support rasterization with jittered sampling, defocus, and motion blur. This work is a follow on to the work presented in [FLB09], which indicates that hardware micropolygon rasterization would be required for real-time micropolygon rendering. In this paper we present three designs for NoBlur rasterization, Interleave rasterization, and a hybrid unit which supports both render modes. We found that for all three designs it was efficient to perform sample tests at low precision and to operate on quads (triangle pairs who share an edge). In the case of an efficient NoBlur design, we found that a 2x2 sample raster stamp should be utilized which contradicts the prior work [FLB09]. In the case of an Interleave rasterization unit we confirmed the prior works results that a sample parallelism of 2-3 across micropolygons is efficient and interleave rasterization would cost 7-10x more than NoBlur rasterization. Finally, we present a hybrid design which balances the performance requirements of both NoBlur and Interleave by serializing the blur specific hardware and utilizing a 2x2 sample raster stamp. BACKGROUND In order to render scenes containing complex geometry, triangles about the size of a pixel or micropolygon are required. This is consistent with all off-line rendering systems utilized for cinematic rendering. John S. Brunhaver II, Kayvon Fatahalion, Pat Hanrahan OBJECTIVE The goal is to implement a system with orders of magnitude improvement over a software implementation. Hardware Design Parameterize micro-architecture Use architectural simulation to predict IPC Use Synthesis to predict clock cycle time Explore micro-architecture and physical design space Verify globally optimal strategies RESULTS CONCLUSIONS Our results suggest that micropolygon rasterization in hardware is more efficient than a software implementation on the order of 100x. We have also shown the an optimal configuration of this design. A remaining issue is understanding the role that large polygons play in a micropolygon rendering pipeline and how efficiently to design a micropolygon rasterization back end with support for large polygons. Additionally, it is not uncommon in off-line render systems for the artist to slightly increase the size of blurred micropolygons for increased performance; an implementation which automatically detects and optimizes for high blur polygons in this manner would be a natural follow on to this work. RESULTS For additional information please contact: John S. Brunhaver II Electrical Engineering Department Stanford University [email protected] Globally Optimal Micro-Architecture Parameters Cost to rasterize at a 300 Million Triangle per second rate Cost to rasterize at a 5 Billion Triangle per second rate
Transcript
Page 1: Hardware Implementation of Micropolygon Rasterization with ...forum.stanford.edu/events/posterslides/Micropolygon...modern systems do not natively support rasterization with jittered

Hardware Implementation of Micropolygon Rasterization

with Motion Blur and Depth Of Field

ABSTRACT

Current GPUs rasterize micropolygons (polygons

approximately one pixel in size) inefficiently. Additionally,

modern systems do not natively support rasterization with

jittered sampling, defocus, and motion blur. This work is a

follow on to the work presented in [FLB09], which indicates

that hardware micropolygon rasterization would be required

for real-time micropolygon rendering. In this paper we present

three designs for NoBlur rasterization, Interleave

rasterization, and a hybrid unit which supports both render

modes. We found that for all three designs it was efficient to

perform sample tests at low precision and to operate on

quads (triangle pairs who share an edge). In the case of an

efficient NoBlur design, we found that a 2x2 sample raster

stamp should be utilized which contradicts the prior work

[FLB09]. In the case of an Interleave rasterization unit we

confirmed the prior works results that a sample parallelism of

2-3 across micropolygons is efficient and interleave

rasterization would cost 7-10x more than NoBlur rasterization.

Finally, we present a hybrid design which balances the

performance requirements of both NoBlur and Interleave by

serializing the blur specific hardware and utilizing a 2x2

sample raster stamp.

BACKGROUND

In order to render scenes containing complex geometry,triangles about the size of a pixel or micropolygon arerequired. This is consistent with all off-line renderingsystems utilized for cinematic rendering.

John S. Brunhaver II, Kayvon Fatahalion, Pat Hanrahan

OBJECTIVE

The goal is to implement a system with orders ofmagnitude improvement over a software implementation.

Hardware Design

Parameterize micro-architecture

Use architectural simulation to predict IPC

Use Synthesis to predict clock cycle time

Explore micro-architecture and physical design space

Verify globally optimal strategies

RESULTS

CONCLUSIONS

Our results suggest that micropolygon rasterization inhardware is more efficient than a software implementationon the order of 100x. We have also shown the an optimalconfiguration of this design.

A remaining issue is understanding the role that largepolygons play in a micropolygon rendering pipeline and howefficiently to design a micropolygon rasterization back endwith support for large polygons. Additionally, it is notuncommon in off-line render systems for the artist to slightlyincrease the size of blurred micropolygons for increasedperformance; an implementation which automaticallydetects and optimizes for high blur polygons in this mannerwould be a natural follow on to this work.

RESULTS

For additional information please contact:

John S. Brunhaver II

Electrical Engineering Department

Stanford University

[email protected]

Globally Optimal Micro-Architecture Parameters

Cost to rasterize at a 300 Million Triangle per second rate

Cost to rasterize at a 5 Billion Triangle per second rate

Recommended