Reduce Memory Requirements Research Articles

Numerical wave simulation for interactive audio-visual applications

Sound is an integral part of many interactive applications, such as 3D computer games. The generality and complexity of the simulated domains makes wave simulation a natural fit. The opposing goals of real-time execution and compute intensiveness of wave simulation provide a rich ground for designing novel computational methods. Two instances of such audio-visual interactive techniques will be discussed. The first technique employs precomputed wave simulation for providing acoustic effects in computer games. Time-domain 3D simulation is performed from many prospective player positions, resulting in a 7D field containing billions of acoustic responses. Perceptual encoding is then performed to drastically reduce memory requirements. Interactive motion of player and sources in the game can then be accommodated via quickly decoding and auralizing the pre-calculated wave effects. The second technique explores the limits of real-time parallel numerical acoustic computation on today's desktop graphics cards. The user is presented with a 2D “sandbox” environment for designing virtual 2D musical wind instruments that are instantly simulated in real-time for the entire human auditory bandwidth. Instrument shape editing and performance via manipulating tone-holes/valves can be done on-the-fly. Video demonstrations of both techniques will be shown along with a discussion of the core technical ideas.

Applying vectorization of diagonal sparse matrix to accelerate wind field calculation

Wind field calculation is a critical issue in reaching accurate forest fire propagation predictions. However, when the involved terrain map is large, the amount of memory and the execution time can prevent them from being useful in an operational environment. Wind field calculation involves sparse matrices that are usually stored in CSR storage format. This storage format can cause sparse matrix-vector multiplications to create a bottleneck due to the number of cache misses involved. Moreover, the matrices involved are extremely sparse and follow a very well-defined pattern. Therefore, a new storage system has been designed to reduce memory requirements and cache misses in this particular sparse matrix-vector multiplication. Sparse matrix-vector multiplication has been implemented using this new storage format and taking advantage of the inherent parallelism of the operation. The new method has been implemented in OpenMP, MPI and CUDA and has been tested on different hardware configurations. The results are very promising and the execution time and memory requirements are significantly reduced.

Adding approximate counters

We describe a general framework for adding the values of two approximate counters to produce a new approximate counter value whose expected estimated value is equal to the sum of the expected estimated values of the given approximate counters. (To the best of our knowledge, this is the first published description of any algorithm for adding two approximate counters.) We then work out implementation details for five different kinds of approximate counter and provide optimized pseudocode. For three of them, we present proofs that the variance of a counter value produced by adding two counter values in this way is bounded, and in fact is no worse, or not much worse, than the variance of the value of a single counter to which the same total number of increment operations have been applied. Addition of approximate counters is useful in massively parallel divide-and-conquer algorithms that use a distributed representation for large arrays of counters. We describe two machine-learning algorithms for topic modeling that use millions of integer counters, and confirm that replacing the integer counters with approximate counters is effective, speeding up a GPU-based implementation by over 65% and a CPU-based by nearly 50%, as well as reducing memory requirements, without degrading their statistical effectiveness.

Efficient on-the-fly interpolation technique for Bethe–Salpeter calculations of optical spectra

The Bethe–Salpeter formalism represents the most accurate method available nowadays for computing neutral excitation energies and optical spectra of crystalline systems from first principles. Bethe–Salpeter calculations yield very good agreement with experiment but are notoriously difficult to converge with respect to the sampling of the electronic wavevectors. Well-converged spectra therefore require significant computational and memory resources, even by today’s standards. These bottlenecks hinder the investigation of systems of great technological interest. They are also barriers to the study of derived quantities like piezoreflectance, thermoreflectance or resonant Raman intensities.We present a new methodology that decreases the workload needed to reach a given accuracy. It is based on a double-grid on-the-fly interpolation within the Brillouin zone, combined with the Lanczos algorithm. It achieves significant speed-up and reduction of memory requirements. The technique is benchmarked in terms of accuracy on silicon, gallium arsenide and lithium fluoride. The scaling of the performance of the method as a function of the Brillouin Zone point density is much better than a conventional implementation. We also compare our method with other similar techniques proposed in the literature.

A Fast Method for the Segmentation of Synaptic Junctions and Mitochondria in Serial Electron Microscopic Images of the Brain.

Recent electron microscopy (EM) imaging techniques permit the automatic acquisition of a large number of serial sections from brain samples. Manual segmentation of these images is tedious, time-consuming and requires a high degree of user expertise. Therefore, there is considerable interest in developing automatic segmentation methods. However, currently available methods are computationally demanding in terms of computer time and memory usage, and to work properly many of them require image stacks to be isotropic, that is, voxels must have the same size in the X, Y and Z axes. We present a method that works with anisotropic voxels and that is computationally efficient allowing the segmentation of large image stacks. Our approach involves anisotropy-aware regularization via conditional random field inference and surface smoothing techniques to improve the segmentation and visualization. We have focused on the segmentation of mitochondria and synaptic junctions in EM stacks from the cerebral cortex, and have compared the results to those obtained by other methods. Our method is faster than other methods with similar segmentation results. Our image regularization procedure introduces high-level knowledge about the structure of labels. We have also reduced memory requirements with the introduction of energy optimization in overlapping partitions, which permits the regularization of very large image stacks. Finally, the surface smoothing step improves the appearance of three-dimensional renderings of the segmented volumes. Electronic supplementary material The online version of this article (doi:10.1007/s12021-015-9288-z) contains supplementary material, which is available to authorized users.

Automatic portal generation for 3D audio – from triangle soup to a portal system

The purpose of this paper is to investigate an algorithm for generating an automatic portal system. This has been accomplished based on a given set of triangles. The proposed solution was designed to enhance the performance of a sound beam-tracing engine. This solution can also be used for other areas where portal systems are applicable. The provided technical solution emphasizes the beam tracing engine's requirements. Our approach is based on the work of Haumont et al. (with additional improvements), resulting in improved scene segmentation and lower computational complexity. We examined voxelization techniques and their properties, and have adjusted these to fit the requirements of a beam-tracing engine. As a result of our investigation, a new method for finding portal placement has been developed by adjusting the orientation of the found portals to fit the neighboring scene walls. In addition, we replaced Haumont et al.'s prevoxelization step, which is used for erasing geometrical details (for example, thin walls). This was done by smoothing the distance field that, in effect, eliminated incorrectly positioned portals. The results of our work remove the requirement for walls that separate rooms to have a particular thickness. We also describe a method for building a structure that accelerates real-time queries for determining the area where a given point is located. All of the presented techniques allow for the use of larger sized voxels, which increases performance and reduces memory requirements (not only during the preprocessing phase but also during real-time usage). The proposed solutions were tested using scenarios with scenes of varying complexity.

Lifting based 2d-Discrete Wavelet Transform for Image Compression based on Residue Number System

The high-performance of image compression is needed greater as digital image finds its way into many areas of everyday life. In order to transmit and store digital images, the images must be compressed; otherwise, each image would require a large amount of memory. JPEG 2000 which is one of the bestimage compression standardis used. Unlike the discrete cosine transform, DWT has higher compression ratios, no blocking artifacts and good localization in time and frequency domain, inherent scaling and higher flexibility. Owing to such inherent advantages of DWT, two-dimensional (2D) bi-orthogonal reversible/irreversible DWT is adopted in JPEG2000 still image compression standard. In the proposed work the 2-D Discrete Wavelet Transform (DWT), using 9/7 filter with lifting scheme is designed. Lower computational complexity and reduced memory requirement are the advantages of lifting scheme. Predictor and updater are the two coefficients used in the DWT decomposition process with lifting scheme. In the 2D DWT lifting based structure multipliers are the more power consumed computational blocks. So in order to reduce the power and to increase the speed of the system, the array multipliers are replaced by Residue Number System(RNS) multiplier and adders by RNS adder. Thus the entire lifting structure is designed to operate with RNS. So, the complexity in computation and power consumption are reduced. Using Synopsystool reports of design vision are obtained to verify power and area of both array multiplier lifting structure and RNS lifting structure.

A Multilevel Approach for Computing the Limited-Memory Hessian and its Inverse in Variational Data Assimilation

Use of data assimilation techniques is becoming increasingly common across many application areas. The inverse Hessian (and its square root) plays an important role in several different aspects of these processes. In geophysical and engineering applications, the Hessian-vector product is typically defined by sequential solution of a tangent linear and adjoint problem; for the inverse Hessian, however, no such definition is possible. Frequently, the requirement to work in a matrix-free environment means that compact representation schemes are employed. In this paper, we propose an enhanced approach based on a new algorithm for constructing a multilevel eigenvalue decomposition of a given operator, which results in a much more efficient compact representation of the inverse Hessian (and its square root). After introducing these multilevel approximations, we investigate their accuracy and demonstrate their efficiency (in terms of reducing memory requirements and/or computational time) using the example of preconditioning a Gauss-Newton minimization procedure.

Multiobjective Stochastic Economic Dispatch With Variable Wind Generation Using Scenario-Based Decomposition and Asynchronous Block Iteration

We investigated a multiobjective stochastic economic dispatch (MOSED) problem considering variable wind power integration. We transformed this problem into an equivalent large-scale multiobjective deterministic optimization model based on the scenario method. We simultaneously minimized power purchase costs and polluting gas emissions. We introduced the normal boundary intersection (NBI) method to convert the multiobjective optimization (MOO) model into a series of single-objective optimization (SOO) problems, which we solved using the interior-point method (IPM). In the process used to solve each SOO problem, we rearranged the coefficient matrix of the correction equation in the block bordered diagonal form (BBDF) according to the sequence of the forecast scenario and sampling scenarios. Thus, we were able to decompose this correction equation further into a number of low-dimensional equations corresponding to the forecast scenario and sampling scenarios, respectively, and solve them using the asynchronous block iteration method. Furthermore, we implemented the proposed algorithm on an IEEE 39-bus system and a real-provincial power system, and built a parallel computational framework on high-performance clusters to demonstrate the enhancements in computational speed and the reduced memory requirements obtained by parallelization. Through this framework, one can obtain scheduling of the outputs of generators on a day-ahead basis.

A hybrid broadband analysis approach for surface-wire junctions structures by applying AIM and asymptotic waveform evaluation technique

The asymptotic waveform evaluation (AWE) technique is utilized to deal with broadband electromagnetic problems of surface–wire junction geometries in this paper. To ensure the current continuity, triangles-to-segments junction model is used and the Rao–Wilton–Glisson (RWG) and triangular basis functions are used to solve the distributions of current on perfect electric conductor and wire, respectively. Then, the current on surface conductor, wire, and junctions are expanded in a rational function via Padé approximation. The AWE technique needs to store the high-order derivatives of the self-impedance and mutual-impedance matrices of surface–wire, which increase the memory requirements. Thus, the adaptive integral method is employed to reduce memory requirements and to accelerate the matrix-vector multiplications in an iterative solver. Finally, some numerical examples are shown to demonstrate the efficiency and accuracy of the proposed method

Computing eigenvalue sensitivity coefficients to nuclear data based on the CLUTCH method with RMC code

Recently, there is a need to develop advanced methods of computing eigenvalue sensitivity coefficients to nuclear data in the continuous-energy Monte Carlo codes. One of these methods is the iterated fission probability (IFP) method, which is adopted by most of Monte Carlo codes of having the capabilities of computing sensitivity coefficients, including the Reactor Monte Carlo code RMC. Though it is accurate theoretically, the IFP method faces the challenge of huge memory consumption. Therefore, it may sometimes produce poor sensitivity coefficients since the number of particles in each active cycle is not sufficient enough due to the limitation of computer memory capacity. In this work, two algorithms of the Contribution-Linked eigenvalue sensitivity/Uncertainty estimation via Tracklength importance CHaracterization (CLUTCH) method, namely, the collision-event-based algorithm (C-CLUTCH) which is also implemented in SCALE and the fission-event-based algorithm (F-CLUTCH) which is put forward in this work, are investigated and implemented in RMC to reduce memory requirements for computing eigenvalue sensitivity coefficients. While the C-CLUTCH algorithm requires to store concerning reaction rates of every collision, the F-CLUTCH algorithm only stores concerning reaction rates of every fission point. In addition, the fission matrix method is put forward to generate the adjoint fission source distribution for the CLUTCH method to compute sensitivity coefficients. These newly proposed approaches implemented in RMC code are verified by a SF96 lattice model and the MIT BEAVRS benchmark problem. The numerical results indicate the accuracy of the F-CLUTCH algorithm is the same as the C-CLUTCH algorithm while the F-CLUTCH algorithm requires less memory and has higher computational efficiency.

Preconditioning for Radial Basis Function Partition of Unity Methods

Meshfree radial basis function (RBF) methods are of interest for solving partial differential equations due to attractive convergence properties, flexibility with respect to geometry, and ease of implementation. For global RBF methods, the computational cost grows rapidly with dimension and problem size, so localised approaches, such as partition of unity or stencil based RBF methods, are currently being developed. An RBF partition of unity method (RBF---PUM) approximates functions through a combination of local RBF approximations. The linear systems that arise are locally unstructured, but with a global structure due to the partitioning of the domain. Due to the sparsity of the matrices, for large scale problems, iterative solution methods are needed both for computational reasons and to reduce memory requirements. In this paper we implement and test different algebraic preconditioning strategies based on the structure of the matrix in combination with incomplete factorisations. We compare their performance for different orderings and problem settings and find that a no-fill incomplete factorisation of the central band of the original discretisation matrix provides a robust and efficient preconditioner.

FrWF-Based LMBTC: Memory-Efficient Image Coding for Visual Sensors

After the successful development of JPEG2000, many state-of-the-art wavelet-based image coding algorithms have been developed. However, the traditional discrete wavelet transform (DWT) is implemented with memory intensive and time-consuming algorithms and, therefore, has very high system resource requirements. In particular, the very large requirement of memory poses a serious limitation for multimedia applications on memory-constrained portable devices, such as digital cameras and sensor nodes. In this paper, we propose a novel wavelet-based image coder with low memory requirements and low complexity that preserves the compression efficiency. Our encoder employs the fractional wavelet filter (FrWF) to calculate the DWT coefficients, which are quantized and encoded with a novel low memory block tree coding (LMBTC) algorithm. The LMBTC is a listless form of the wavelet block tree coding algorithm. Simulation results demonstrate that the proposed coder significantly reduces memory requirements and computational complexity and has competitive coding efficiency in comparison with other state-of-the-art coders. The FrWF combined with the LMBTC is, thus, a viable option for image communication over wireless sensor networks.

Efficient Analysis of Surface‐Wire Configurations Using AIM and Best Uniform Rational Approximation

The Adaptive integral method (AIM) in conjunction with the best uniform approximation technique is applied to analyze the electromagnetic problems of surface-wire junction geometries over a wide frequency band. To improve the computation efficiency, the best uniform approximation method is utilized. The AIM can reduce memory requirements and accelerate the matrix-vector multiplications in iterative process. Some numerical examples are shown to demonstrate the efficiency and accuracy of the proposed method. Compared with the direct solution method and Asymptotic waveform evaluation (AWE) technique, the proposed technique is found to be efficient in a broadband with the lower Central processing unit (CPU) time required and without loss of accuracy.

Fast computation of seabed spherical-wave reflection coefficients in geoacoustic inversion.

This paper develops a fast numerical approach to computing spherical-wave reflection coefficients (SWRCs) for layered seabeds, which provides substantial savings in computation time when used as the forward model for geoacoustic inversion of broadband seabed reflectivity data. The approach exploits the Sommerfeld-integral representation of SWRCs as the Hankel transform of a function proportional to the plane-wave reflection coefficient (PWRC), and applies Levin integration to the rapidly oscillating integrand cast as the product of a (pre-computed) media-independent matrix and a vector involving PWRCs at a sparse sampling of integration angles. Compared to conventional Simpson's rule integration for computation of the SWRC, the Levin integration yields speed-up factors of an order of magnitude or more. Further, it results in reduced memory requirements for storage of pre-computed quantities, a desirable property when a graphics processing unit (GPU) is used for parallel computation of SWRCs. The paper applies trans-dimensional Bayesian inversion to investigate the impact of forward modeling in terms of PWRCs and SWRCs on the estimation of geoacoustic parameters and uncertainties. Model comparisons are quantified in simulated- and measured-data inversions by comparing the estimated geoacoustic parameters to the true parameters or core measurements, respectively, and by calculating the deviance information criterion for model selection.

Computational study of radiation torque on arbitrary shaped particles with MLFMA.

The surface integral equation (SIE) method is used for the computational study of radiation torque on arbitrarily shaped homogeneous particles. The Multilevel Fast Multipole Algorithm (MLFMA) is employed to reduce memory requirements and improve the capability of SIE. The resultant matrix equations are solved iteratively to obtain equivalent electric and magnetic currents. Then, radiation torque is computed using the vector flux of the pseudotensor over a spherical surface tightly enclosing the particle. We use, therefore, the analytical electromagnetic field expression for incident waves in the near region, instead of the far-field approximation. This avoids the error which may be caused when describing the incident beam. The numerical results of three kinds of non-spherical particles are presented to illustrate the validity and capability of the developed method. It is shown that our method can be applied to predict, in the rigorous sense, the torque from a beam of any shape on a particle of complex configuration with a size parameter as large as 650. The radiation torques on large ellipsoids are exemplified to show the performance of the method and to study the influence that different aspect ratios have on the results. Then, the code is used for the calculation of radiation torque on objects of complex shape including a biconcave cell-like particle and a motor with a non-smooth surface.

Low‐cost and high‐speed hardware implementation of contrast‐preserving image dynamic range compression for full‐HD video enhancement

This study presents a cost-efficient and high-performance field programmable gate array (FPGA)-based hardware implementation of a contrast-preserving image dynamic range compression algorithm, which is an important function used in modern digital video cameras and displays to improve visual quality of standard dynamic range colour images (8 bits/channel). To achieve this purpose, a hardware-friendly approximation to an existing fast dynamic range compression with local contrast preservation (FDRCLCP) algorithm is proposed. The computation of the proposed approximated FDRCLCP algorithm requires only fixed-point unsigned binary addition, multiplication, and bit-shifting. Moreover, the proposed hardware implementation uses a line buffer instead of a frame buffer to process whole image data. These advantages significantly improve throughput performance and reduce memory requirement of the system. The FPGA implementation of the proposed algorithm requires only about 98 K bits on-chip memory and achieves about 170.24 MHz operating frequency by using an Altera Cyclone II device. This is a large improvement compared with the existing results as it is quick enough to process full high-definition videos (1920 × 1080 pixels) at least 80 frames per second using a low-cost FPGA device.

Design of Modified Canny Edge Detector Based on FPGA for Portable Device

Edge detection is one of the key stages in image processing and object reorganization. The Canny Edge Detector is one of the most widely used edge detection algorithm due to its superior performance. In this paper, we propose a mechanism to implement the Canny algorithm at the block level without any loss in edge detection performance compared with the original frame-based Canny algorithm. Directly applying the original Canny Edge detection algorithm at the blocklevel leads to excessive edges in smooth regions and to loss of significant edges in high-detailed regions since the original Canny computes the high and low thresholds based on the frame-level statistics. To solve this problem, we present a modified Canny edge detection algorithm that adaptively computes the edge detection thresholds based on the block type and the local distribution of the gradients in the image block. Here we propose the design of modified Canny Edge detection algorithm that results in significantly reduced memory requirement, decrease in latency, increase throughput, with no loss in edgedetection performance as compare to original Canny Detector Algorithm. Here we are using matlab to convert image into text/pixel value.

An improved filter of progressive TIN densification for LiDAR point cloud data

Based on the classic filter of progressive triangulated irregular network (TIN) densification, an improved filter is proposed in this paper. In this method, we divide ground points into grids with certain size and select the lowest points in the grids to reconstruct TIN in the process of iteration. Compared with the classic filter of progressive TIN densification (PTD), the improved method can filter out attached objects, avoid the interference of low objects and obtain relatively smooth bare-earth. In addition, this proposed filter can reduce memory requirements and be more efficient in processing huge data volume. The experimental results show that the filtering accuracy and efficiency of this method is higher than that of the PTD method.

Fast randomized approximate string matching with succinct hash data structures

BackgroundThe high throughput of modern NGS sequencers coupled with the huge sizes of genomes currently analysed, poses always higher algorithmic challenges to align short reads quickly and accurately against a reference sequence. A crucial, additional, requirement is that the data structures used should be light. The available modern solutions usually are a compromise between the mentioned constraints: in particular, indexes based on the Burrows-Wheeler transform offer reduced memory requirements at the price of lower sensitivity, while hash-based text indexes guarantee high sensitivity at the price of significant memory consumption.MethodsIn this work we describe a technique that permits to attain the advantages granted by both classes of indexes. This is achieved using Hamming-aware hash functions--hash functions designed to search the entire Hamming sphere in reduced time--which are also homomorphisms on de Bruijn graphs. We show that, using this particular class of hash functions, the corresponding hash index can be represented in linear space introducing only a logarithmic slowdown (in the query length) for the lookup operation. We point out that our data structure reaches its goals without compressing its input: another positive feature, as in biological applications data is often very close to be un-compressible.ResultsThe new data structure introduced in this work is called dB-hash and we show how its implementation--BW-ERNE--maintains the high sensitivity and speed of its (hash-based) predecessor ERNE, while drastically reducing space consumption. Extensive comparison experiments conducted with several popular alignment tools on both simulated and real NGS data, show, finally, that BW-ERNE is able to attain both the positive features of succinct data structures (that is, small space) and hash indexes (that is, sensitivity).ConclusionsIn applications where space and speed are both a concern, standard methods often sacrifice accuracy to obtain competitive throughputs and memory footprints. In this work we show that, combining hashing and succinct indexing techniques, we can attain good performances and accuracy with a memory footprint comparable to that of the most popular compressed indexes.