A Floating Point Data Compression Using Inter-Extrapolative Predictor

In recent years, along with advances in processor technology, large-scale numerical simulation dealing with huge floating point data is an active research field. In the case of supercomputing using many processors via the network, the bottleneck is communication speed rather than the calculation one...

Full description

Saved in:
Bibliographic Details
Published in:2018 IEEE 61st International Midwest Symposium on Circuits and Systems (MWSCAS) pp. 546 - 549
Main Authors: Imai, Shinya, Fukuma, Shinji, Mori, Shin-ichiro
Format: Conference Proceeding
Language:English
Japanese
Published: IEEE 01.08.2018
Subjects:
Data compression
Data models
Decoding
Encoding
Extrapolation
Indexes
Interpolation
ISSN:1558-3899
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent years, along with advances in processor technology, large-scale numerical simulation dealing with huge floating point data is an active research field. In the case of supercomputing using many processors via the network, the bottleneck is communication speed rather than the calculation one. Thus reduction of the communication cost is strongly required in the supercomputing and its straightforward technique is data compression. Most conventional floating-point data compression schemes use predictive coding based on an extrapolative prediction [1-2]. In general, it is known that interpolative prediction has higher accuracy than extrapolative one. This paper proposes a floating point data compression using combined predictor with interpolation and extrapolation, namely Inter-Extrapolative predictor. Proposed predictor is defined one-dimensional but it can be is easily expanded into 2-dimensional by cascading in the horizontal-vertical direction. Note that the interpolation violates causality because it uses future data. In order to avoid this problem, transmission order of data is arranged; first, odd index sequence; second, remaining even index one. The first even index sequence is extrapolated, then the odd index sequence. Similar to Lindstrom's method [1], each prediction residual is encoded by the range coder with '0' run-length of bitwise. Thus proposed method can gain the higher compression rate than [1] under the same number of arithmetic operation.
ISSN:1558-3899
DOI:10.1109/MWSCAS.2018.8623986