Power Analysis of Mobile 3D Graphics - DAC

1

2


-
Based Workload
Estimation for Mobile 3D Graphics

Bren Mochocki*
†
, Kanishka Lahiri*,

Srihari Cadambi*, Xiaobo Sharon Hu
†

*NEC Laboratories America,
†
University of Notre Dame

DAC 2006

3

Mobile Graphics Technology

2000

2001

2002

2003

2004

2005

2006

2007

Basic 3D

Graphics


Technology

Video clips

Advanced 3D

1997

2D color

Time

Increasing resource load


•

Performance (Speed)

•

Lifetime (Energy)

4

Meeting Performance/Lifetime Requirements

System
-

Level
Optimizations

Graphics

Algorithms

Hardware

Solutions

Tack, 04

•

LoD control for
mobile terminals

Kameyama, 03

•

low
-
power 3D ASIC

Woo, 04

•

low
-
power 3D ASIC

Akenine
-
Moller, 03

•

Texture compression


for mobile terminals

Mochocki, Lahiri, Cadambi, 06

•

DVFS for mobile 3D graphics

Accurate workload prediction is critical

Gu, Chakraborty, Ooi, 06

•

Games are up for DVFS

5

Mobile 3D Workload Estimation


Why?



Adapt architectural parameters



Adapt application parameters



Better on
-
line resource management


Desirable properties



Speed
–

must be performed on
-
line



Accuracy



Compact

6

Workload
-
Estimation Spectrum


General purpose history
-
based predictors provide poor
prediction accuracy for rapidly changing workloads


Highly accurate analytical schemes are too complex for
use at run time

General Purpose

Simplicity

Application specific

Accuracy

History
-
Based
Predictors

Analytical
Predictors

7

Workload
-
Estimation Spectrum


Uses combination of history and application
-
specific
parameters (the signature) to predict future workload


Strikes a balance between simplicity and accuracy


Preserves both cause AND effect


Preserves substantial history

General Purpose

Simplicity

Application specific

Accuracy

Signature
-
Based Predictor

8

Outline


Introduction and Motivation


Background



3D
-
pipeline Basics



Challenges in workload Estimation


Signature
-
Based Workload Prediction


Experimental Results


Conclusions


9

3D Pipeline Basics


3D representation


2D image

World View

Camera View

Raster View

Frame Buffer

Geometry

Setup

Rendering

•

Transformations

•

Lighting


•

Clipping

•

Scan
-
line conversion

•

Pixel rendering

•

Texturing


Texturing

10

Workload Across Applications


Workload varies significantly between applications


Prediction scheme must be flexible

RoomRev

TexCube

0

2

4

6

8

10

12

Execution Cycles (ARM, x10
7
)

Benchmark

11

Workload Within an Application


Workload can change rapidly between frames

0

1

2

3

4

5

6

1

16

31

46

61

76

91

106

121

136

151

166

181

196

Execution Cycles (ARM, x10
7
)

Frame

geometry

render

setup

Race

12

Outline


Introduction and Motivation


Background


Signature
-
Based Workload Prediction



Signature Generation



Method Overview



Pipeline Modifications


Experimental Results


Conclusions


13

Example

Signature

Table

Application

Frame

Buffer

Workload
Prediction



Signature

Workload

<6, 2.5>

1.0e4

extract

signature

measure

workload

Default

end

frame

extract

Signature:

<
vertex count
,

avg. area
>

3D Pipeline

14

Example

Signature

Table

Application

Frame

Buffer

Workload
Prediction



Signature

Workload

<6, 2.5>

<6, 2.5>

1.0e4

extract

signature

measure

workload

1.0e4

1.0e4

end

frame

extract

3D Pipeline

Signature:

<
vertex count
,

avg. area
>

15

Example

Signature

Table

Application

Frame

Buffer

Workload
Prediction



Signature

Workload

<6, 2.5>

<6, 2.5>

1.2e4

extract

signature

measure

workload

1.0e4

1.0e4

end

frame

extract

No overlap


(render all pixels)

3D Pipeline

Signature:

<
vertex count
,

avg. area
>

16

Transform

Clipping

Lighting

Scan
-
line

conversion

Per
-
pixel

Operations

Lighting

Scan
-
line

conversion

Per
-
pixel

Operations

Transform

Clipping

Application

Display

Partitioning the 3D pipeline

GEOMETRY

SETUP

RENDER

Application

Display

•

Generally small workload

•

Provides necessary signature elements

Bulk of 3D workload

Transform

+

Clipping

Scan
-
line

conversion

Per
-
pixel

Operations

Lighting

Buffer

ORIGINAL

PARTITIONED

Pre
-
Buffer

Post Buffer

17

Pipeline Workload


Pre
-
buffer workload is
less than 10% of the
total workload


Pre
-
buffer variation is
small


Post
-
buffer workload is
large with significant
variation

post
-
buffer

pre
-
buffer

18

Signature Composition


Can vary by application


May include:

1.
Average Triangle Area

2.
Average Triangle Height

3.
Total vertex count

4.
Lit vertex count

5.
Number of lights

6.
Any measurable parameter


Larger signatures


more accurate


Smaller signatures


less time & space


19

Outline


Introduction & Background


Experimental Framework


Signature
-
Based Workload Prediction


Experimental Results



Evaluation Framework



Signature length vs. accuracy



Frame Rate



Energy


Conclusions


20

Architectural View

Programmable

3D Graphics

Engine

Frame

Buffer

Performance

counter

Memory

Applications

Processor

System
-
level Communication Architecture

Prog. Voltage

Regulator

Prog. PLL

V, F

•

buffer

•

signature table

•

pre
-
buffer

•

signature extraction

post
-
buffer

output

measure workload

21

Evaluation Framework

OpenGL/ES library

Instrumented with

pipeline stage triggers

Hans
-
Martin Will

Fast, cycle
-
accurate

Simulation

W. Qin

Trace simulator of
mobile 3D pipeline

OpenGL/ES 1.0

3D
–

application

3D pipeline

Performance/power

Simit
-
ARM

Cross Compiler

ARM
—

g++

Trace Simulator

Triangle,

Instruction, &

Trigger traces

Workload prediction

scheme

3D application

Vincent

Processor

Energy Model

Architecture Model

Simulation output

22

Workload Accuracy

Average Error (normalized)

<
a
>

2 bytes

<
a
,
b
>

6 bytes

<
a
,
b
,
c
>

10 bytes

<
a
,
b
,
c
,
d
>

14 bytes

Signature Complexity


> 2 fps error at peaks

Peaks < 1 fps

<
a
> triangle count, <
b
> avg. area, <
c
> avg. height, <
d
> vertex count

23

Frame Rate

High peaks result in wasted energy

Low valleys result in poor visual quality

Target

24

Workload prediction for DVFS

Before DVFS

DVFS using signature
-
based
workload Prediction

32% energy
reduction

25

Outline


Introduction & Background


Experimental Framework


Signature
-
Based Workload Prediction


Experimental Results


Conclusions

26

Conclusions


Accurate 3D workload prediction critical for
mobile platforms.


Proposed signature
-
based method



Outperforms conventional history methods



Trade accuracy for time & space


Can be used to meet real time constraints
and conserve energy.


27

Future Work


Automatic selection of signature elements


More sophisticated data structures for
signature storage


Faster comparison and replacement
algorithms

28


-
Based Workload
Estimation for Mobile 3D Graphics

Bren Mochocki*
†
, Kanishka Lahiri*,

Srihari Cadambi*, Xiaobo Sharon Hu
†

*
NEC Laboratories America,
†
University of Notre Dame

DAC 2006

Questions?