a gpu-based architecture for parallel image-aware version ...leomurta/courses/cm/csmr2012.pdf ·...

A GPU-based Architecture for Parallel Image-aware Version Control

Jose Ricardo da Silva Junior Toni PachecoEsteban Clua Leonardo Murta

Instituto de Computação

Universidade Federal Fluminense


IntroductionGPU processingVCS operationsResultsConclusionFuture works

Schedule

2


Version control, nowadays, is considered a vitalcomponent for supporting professionalsoftware development.

Mainly based on files and directories.

Textual artifacts has a well established process.

Unfortunately, VCS for binary data are not yet well established.

Introduction

3


Movie industry

Game Industry

Advertising Industry

Introduction

Normally, has more binary artifacts (sound, 3d models, images) than textual artifacts

So many projects are highly binary data intensive!


Normally, to deal with binary data, two paths are used:

Store the binary data as a whole between each modification.• Loose semantic information!• Requires more storage space.• No processing time.

Implement algorithms to deal with these binary artifacts.• Allows more semantic for the end user.• Requires more processing and time!

Introduction

5


What has been changed between these two revisions?

Storing binary data as a whole

6

Revision 1 Revision 2


Normally, state-based VCS, such as Git, saves binarydata without any delta information.

High network traffic in projects that uses a lot of binary artifacts!Slow down operations of check –in and –out!More disk space required.Loose of productivity!

Storing binary data as a whole

7

Check-out


Store delta information between binary artifacts.Less disk space required.Allows more semantic to be presented to user.On the other hand, normally requires a lot moreprocessing during check-in and –out operations!

Semantic based algorithms

8

Check-out



9

In order to process two single images for diff operation, as example:

Processing of 2.097.152 elements!

Revision 1 Revision 2

Hei

ght:

102

4

Hei

ght:

102

4

Width: 1024Width: 1024

1.048.576 pixels 1.048.576 pixels



10

Normally, images are composed pixels of three (RGB) or four (RGBA) channels, for Red, Green, Blue and Alpha, requiring processing each channel, individually during VCS operation.


Due to these observations, we are aimed to:Give semantic information for image type artifacts;Process as fast as possible diff, patch and merge operations during check-in and –out;Use less space to store delta between revisions.Using GPU due to problem characteristics (data independency).

Motivation

11

DIF

F

MER

GE

PATC

H


Aimed to locate differences between imagesand save its delta.

Uses the XOR operation on each channel to find its delta.

Diff on IMUFF

12

C.At(x,y).R = A.At(x,y).R XOR B.At(x,y).R;C.At(x,y).G = A.At(x,y).G XOR B.At(x,y).G;C.At(x,y).B = A.At(x,y).B XOR B.At(x,y).B;

A B C

Diff =

Parallel


As can be observed, most of our delta image is composed of black colors (zeros).

After compression, this delta leads to small size, requiring less storageand network bandwidth.

Usually, small deltas are expected between two consecutives versions.

Gives the user a high semantic information of the modification.

Diff on IMUFF

13


Aimed to reconstruct others revisions.Like diff, uses the XOR operation on each channel to reconstruct a revision.

Patch on IMUFF

14

B.At(x,y).R = A.At(x,y).R XOR C.At(x,y).R;B.At(x,y).G = A.At(x,y).G XOR C.At(x,y).G;B.At(x,y).B = A.At(x,y).B XOR C.At(x,y).B;

A C B

Patch =

Parallel


Some properties of XOR operation in IMUFF:patch(A,C) = A if C is an empty delta

patch (A,C) = patch(C,A) = Bpatch (A,C) = B and patch(B,C) = A

Patch on IMUFF

15

B.At(x,y).R = A.At(x,y).R XOR C.At(x,y).R;B.At(x,y).G = A.At(x,y).G XOR C.At(x,y).G;B.At(x,y).B = A.At(x,y).B XOR C.At(x,y).B;

B C A

Patch =

Parallel


Performed to conciliate two revisions created in parallel.Uses the previously diff and patch operations.

Merge on IMUFF

16

A"

B" D"

Changes

A"

B" D"Changes

Diff

C"

A"

B" D"Changes

Patch

C"

E"

1 2


Performed to conciliates two revisions created in parallel.

Merge on IMUFF

17

In case the same image area are changed, a conflict in generated, like a common line based VCS.

A

B

C=Diff(A,B)

D

E=Patch(D,C)


In order to perform these operations on IMUFF, a GUI is freely available at http://josericardojunior.com/imuff/.

Results

18

http://josericardojunior.com/imuff/


Processing time:

Results

19

1"

10"

100"

1000"

10000"

512x512" 1024x1024" 2048x2048" 4096x4096"

Time%(m

s)%

Image%Size%

Image%Processing%Time%

CPU"Diff"

GPU"Diff"

CPU"Patch"

GPU"Patch"

CPU"Merge"

GPU"Merge"

*Using log10 scale.


Comparison between storage for the whole binary data and our delta.

Results

20

0,44802&

3&5,841752&

11,777958&

0,006655&

0,024877&0,085436&

0,151375&

0,000001&

0,00001&

0,0001&

0,001&

0,01&

0,1&

1&

10&

100&

512x512& 1024x1024& 2048x2048& 4096x4096&

Size%(M

Bytes)%

Image%Size%

Storage%Space%

Original&

Delta&

*Using the “Where is Waldo” sample.*Using log10 scale.


Aligned images.Reasonable for VCS as its track evolutions.

All evolutions must maintain the same image’s resolution.

Only work for PNG images.

Restrictions

21


IMUFF is not a VCS as it.Instead, provides infrastructure to allow any VCS to better work with images artifact.

We are planning to develop a plugin for Git and Subversion to deal with image artifacts.

Study how to work with movie artifacts.

Future Works

22


Using GPU for VCS processing can speedup up to 55x CPU processing.

Working with delta for image artifacts can reduce greatly the space required to store it.

Using IMUFF gives the user sematic over its image artifacts.

Allow faster check-in and –out of image artifacts, also reducing network bandwidth for distributed VCS systems.

Conclusion

23


Jose Ricardo da Silva Junior Toni PachecoEsteban Clua Leonardo Murta

Instituto de Computação

Universidade Federal Fluminense


GPU (Graphics Processor Unit) is a massivelymulti-threaded processor capable of performalmost thousands of operations/second.

Using GPU for VCS operations

25

Presented in almost every personal computer!


Allows for heterogeneous environment.Both GPU and CPU doing different tasks at the same time.

Using GPU for VCS operations

26

instruction

instruction

GPUCPU

IdleBusy

a gpu-based architecture for parallel image-aware version ...leomurta/courses/cm/csmr2012.pdf ·...

Documents