Advanced Optimizations for VQ Compression on Parallel SystemsAkiyoshi Wakatani(wakatani@konan-u.ac.jp)

Faculty of Science and Engineering, Konan University, Japan

Recently, parallel processing technologies have been applied to a variety of applications to boostits performance. We implemented a parallel algorithm with “double strip-mining” method of opti-mal codeword search for VQ compression on a shared-memory parallel environment and evaluatedthe effectiveness of the parallel algorithm. Namely, by applying strip-mining method to a loop withregard to vectors as well as that to codewords (double strip-mining), the buffer area, which keepsthe intermediate results, can be dramatically reduced and the results of our experiments find thatthis method can keep the same performance as the original strip-mining method in terms of theelapsed time. As shown in Figure 1-(b), the p-dist algorithm with the double strip-mining methodoutperforms the c-dist algorithm by over 10%, when the size of strip-mining is enough large.

Moreover, two optimization methods for sequential computers are presented: the shortcut calcu-lation and the modified expression. The shortcut calculation is that the distance calculation shouldbe terminated if the intermediate result exceeds the tentative minimum in order to reduce the com-putation cost, and the modified expression is that the distance formula should be altered to achieveless computation. As shown in Figure 1-(a), our experimental results show that the elapsed time isimproved by 20% to 40% by using the both sequential optimization methods and the effectivenessdepends on the size of a vector and the size of the codebook, so the effectiveness of the above meth-ods on sequential computers is empirically confirmed and can be also applied to parallel systemseasily.

7

8

9

10

11

12

13

14

15

4 8 16 32 64 128 256 512 10242048

Ela

psed

tim

e (s

ec)

LOC or strip-mining size

PSIZE=4PSIZE=256PSIZE=max

c-dist

(a) Double strip-mining (SMP with 4CPUs, V SIZE � 8192)

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

4 8 16 32 64 128

spee

dup

vector size

s.cut (K=256)mod. (K=256)

s.cut (K=1024)mod. (K=1024)

(b) Shortcut calculation and modifiedexpression (1024�1024)

Figure 1. Results of our experiments

�������������

This work was supported by MEXT ORC (2004-2008), Japan.

2007 Data Compression Conference (DCC'07)0-7695-2791-4/07 $20.00 © 2007