Implementation of K-means on the Cell Broadband Engine Filip imek - - PowerPoint PPT Presentation

IBM - CVUT Student Research Projects

Implementation of K-means on the Cell Broadband Engine

Filip Šimek (simekf1@fel.cvut.cz)

IBM - CVUT Student Research Projects 2

About K-means

• a data clustering algorithm
• places k centroids in an

n-dimensional space so that intra-cluster variance (sum

• f distances between points

and their nearest centroid) is minimal

• large number of points gives

space for parallelization

⇩

IBM - CVUT Student Research Projects 3

K-means application example

– reducing number of colors of an image

Truecolor original 16 colors found using K-means

IBM - CVUT Student Research Projects 4

Implementation

1

st step - scalar implementation

• the program runs on the PPU of Cell

Broadband Engine with no parallelism

• easy to implement or port from another

scalar implementation

• does not take full advantage of CBE

performance

IBM - CVUT Student Research Projects 5

Implementation

2

nd step - vector implementation

• the PPU distributes work among one or

more SPUs, which use SIMD instructions

• a) multiple serial calls to one SPU thread,
• ccupies only one SPE, the others can

perform another operations in a pipeline

• b) multiple SPUs running in parallel on

multiple data

IBM - CVUT Student Research Projects 6

What's done

• scalar implementation of the algorithm

running on the PPU

• operation on n-dimensional data
• organisation of data into vectors
• PPU loop creating a single SPE thread as

part of the implementation of step 2a)

IBM - CVUT Student Research Projects 7

Future plans

• code running on the SPU
• implementation of step 2b)
• possibility to separate learning and

classification

• call K-means multiple times to reduce error
• use DMA effectively
• implement a unified interface to integrate

this code into a larger toolkit