1-5hit |
In this paper we present an Overlapped Block Gauss-Seidel (OBGS) algorithm for the solution of large scale LSEs (Linear System of Equations) based on array architecture which we have already proposed. Better partitioning for processor array usually requires (1) balanced block size, and (2) minimum coupling between blocks for better convergence. These conditions can well be satisfied by overlapping some variables in computation algorithm. The mathematical implication of overlapped partitioning is discussed at first, and some examples show the effectiveness of OBGS algorithm. Conclusion points out that the convergence properties can well be improved by proper choice of overlapped variables. An efficient algorithm is given for choosing block and variables in order to realize above conditions.
A reconfigurable array for solving linear system of equations (LSE) by LU-decomposition method is considered. It has nearly a half of the number of processing elements (PE) and I/O terminals with reduced computation time compared with known conventional array structures.
Ben CHEN Mahoki ONODA Mineo KANEKO
With the development of LSI technologies, conventional circuit simulation using only single type of method has become unsatisfactory, i.e. circuit-level analysis based on device model spends much simulation time and relaxation methods have the problems on their accuracy. Therefore, it is necessary to develop the better methods to realize both high-speed and good accuracy. In this paper, a mixed-mode circuit-timing simulation method has been studied. It uses a new kind of automatic circuit partition approach--dynamic circuit partition process based on checking coupling factors between circuit nodes at every time point for better convergence. This method is based on examining the characteristic of circuit equations rather than circuit topology or function blocks. A mixed-mode simulation program--MMAPC for transient analysis of CMOS large-scale circuit has been developed and some simulation examples have been performed. The results show that MMAPC can be more than two orders of magnitude faster than a standard" circuit-level simulator (SPICE-like) while providing comparable waveform accuracy, and has better convergence property than general timing-level simulators.
Junibakti SANUBARI Keiichi TOKUDA Mahoki ONODA
In this paper, a new time series analysis method is proposed. The proposed method uses the exponential (EXP) model. The residual signal is assumed to be identically and independently distributed (IID). To achieve accurate and efficient estimates, the parameter of the system model is calculated by maximizing the logarithm of the likelihood of the residual signal which is assumed to be IID t-distribution. The EXP model theoretically assures the stability of the system. This model is appropriate for analyzing signals which have not only poles, but also poles and zeroes. The asymptotic efficiency of the EXP model is addressed. The optimal solution is calculated by the Newton-Raphson iteration method. Simulation results show that only a small number of iterations are necessary to reach stationary points which are always local minimum points. When the method is used to estimate the spectrum of synthetic signals, by using small α we can achieve a more accurate and efficient estimate than that with large α. To reduce the calculation burden an alternative algorithm is also proposed. In this algorithm, the estimated parameter is updated in every sampling instant using an imperfect, short-term, gradient method which is similar to the LMS algorithm.
Junibakti SANUBARI Keiichi TOKUDA Mahoki ONODA
In this paper, a new M-estimation technique for the linear prediction analysis of speech is proposed. Since in the conventional linear prediction (CLP) method the obtained estimates are very much affected by the large amplitude residual parts, in the proposed method we use a loss function which assigns large weighting factor for small amplitude residuals and small weighting factor for large amplitude residuals which is for instance caused by the pitch excitations. The loss function is based on the assumption that the residual signal has an independent and identical t-distribution t(α) with α degrees of freedom. The efficiency of this new estimator depends on α. When α=, we get the CLP method. When the proposed method with small α is applied to the problems of estimating the formant frequencies and bandwidths of the synthetic speech by finding the roots of the prediction polynomial, we can achieve a more accurate and a smaller standard deviation (SD) estimate than that with large α. When the signal is very spiky, the proposed method can ahieve more efficient and accurate estimates than that with robust linear prediction (RBLP) method. The loss function is modified in the similar manner as the autocorrelation method. The solution is calculated by the Newton-Raphson iteration technique. The simulation results show that only few iterations are needed to reach a stationary point, the stationary point is always a local minimum and the obtained prediction filter is always minimum phase. Preliminary experiments on the human speech data indicate that the obtained results are insensitive to the placement of the analysis window and a higher spectral resolution than the CLP and RBLP method can be achieved.