Efficient Methods to Generate Constant SNs with Considering Trade-Off between Error and Overhead and Its Evaluation

Yudai SAKAMOTO  Shigeru YAMASHITA  

Publication
IEICE TRANSACTIONS on Information and Systems   Vol.E103-D   No.2   pp.321-328
Publication Date: 2020/02/01
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2018EDP7435
Type of Manuscript: PAPER
Category: Computer System
Keyword: 
Stochastic Computing,  Stochastic Number,  correlation,  

Full Text: PDF(844.1KB)>>
Buy this Article




Summary: 
In Stochastic Computing (SC), we need to generate many stochastic numbers (SNs). If we generate one SN conventionally, we need a Stochastic Number Generator (SNG) which consists of a linear-feedback shift register (LFSR) and a comparator. When we calculate an arithmetic function by SC, we need to generate many SNs whose values are equal to constant values used in the arithmetic function. As a consequence, the hardware overhead becomes huge. Accordingly, there has been proposed a method called GMCS (Generating Many Constant SNs from Few SNs) to generate many constant SNs with low hardware overhead. However, if we use GMCS simply, generated constant SNs are correlated highly with each other. This would be a serious problem because the high correlation of SNs make a large error in computation. Therefore, in this paper, we propose efficient methods to generate constant SNs with reasonably low hardware overhead without increasing errors. To reduce the correlations of constant SNs which are generated by GMCS, we use Register based Re-arrangement circuit using a Random bit stream duplicator (RRRD). RRRDs have few influences on the hardware overhead because an RRRD consists of three multiplexers (MUXs) and two 1-bit FFs. We also use a technique to share random number generators with several SNGs to reduce the hardware overhead. We provide some experimental results by which we can confirm that our proposed methods are in general very useful to reduce the hardware overhead for generating constant SNs without increasing errors.