Antonio M. Vidal

Real-time massive convolution for audio applications on GPU

Massive convolution is the basic operation in multichannel acoustic signal processing. This field has experienced a major development in recent years. One reason for this has been the increase in the number of sound sources used in playback applications available to users. Another reason is the growing need to incorporate new effects and to improve the hearing experience. Massive convolution requires high computing capacity. GPUs offer the possibility of parallelizing these operations. This allows us to obtain the processing result in much shorter time and to free up CPU resources. One important aspect lies in the possibility of overlapping the transfer of data from CPU to GPU and vice versa with the computation, in order to carry out real-time applications. Thus, a synthesis of 3D sound scenes could be achieved with only a peer-to-peer music streaming environment using a simple GPU in your computer, while the CPU in the computer is being used for other tasks. Nowadays, these effects are obtained in theaters or funfairs at a very high cost, requiring a large quantity of resources. Thus, our work focuses on two mains points: to describe an efficient massive convolution implementation and to incorporate this task to real-time multichannel-sound applications.

Fully Parallel GPU Implementation of a Fixed-Complexity Soft-Output MIMO Detector

Multicore and graphic processing units (GPUs) can be combined to efficiently implement signal-processing algorithms for communication systems, due to their parallel processing capabilities. This paper proposes a fully parallel fixed-complexity soft-output detector, which is suitable for GPU implementation and allows a considerable decrease in the computational time required for the data detection stage in multiple-input-multiple-output (MIMO) systems. A novel channel matrix preprocessing stage, based on column-norm ordering, is developed to efficiently match the multicore architecture. The throughput of the implementation is shown to outperform other recent implementations and to support some of the configurations in the long-term evolution (LTE) standard.

Combined K-Best sphere decoder based on the channel matrix condition number

It is known that sphere decoding (SD) methods can provide maximum-likelihood (ML) detection over Gaussian MIMO channels with lower complexity than the exhaustive search. Channel matrix condition number represents an important influence on the performance of usual detectors. Throughout this paper, two particular cases of a SD method called K-Best carry out a combined detection in order to reduce the computational complexity with predictable performance degradation. Algorithm selection is based on channel matrix condition number thresholding. K-Best is a suboptimal SD algorithm for finding the ML solution of a detection problem. It is based on a fixed complexity tree search, set by a parameter called k. The proposed receiver makes use of a low value of k while working with well-conditioned channels and switches to a higher value of k whether the channel gets worse. It is also presented practical algorithms for finding the 1-norm condition number of a given channel matrix and the condition number threshold selection. Finally an algorithm variant that switches between an ML SD and a linear detector is also evaluated.

On the performance of multi-GPU-based expert systems for acoustic localization involving massive microphone arrays

Expert system for passive sound source localization that makes use of multiple GPUs.Fine spatial grids and a high number of microphones provide excellent localization.GPU resources for managing a large expert system are described.A complete set of simulations evaluates the performance of the expert system.Excellent localization accuracy is achieved even in adverse environments. Sound source localization is an important topic in expert systems involving microphone arrays, such as automatic camera steering systems, human-machine interaction, video gaming or audio surveillance. The Steered Response Power with Phase Transform (SRP-PHAT) algorithm is a well-known approach for sound source localization due to its robust performance in noisy and reverberant environments. This algorithm analyzes the sound power captured by an acoustic beamformer on a defined spatial grid, estimating the source location as the point that maximizes the output power. Since localization accuracy can be improved by using high-resolution spatial grids and a high number of microphones, accurate acoustic localization systems require high computational power. Graphics Processing Units (GPUs) are highly parallel programmable co-processors that provide massive computation when the needed operations are properly parallelized. Emerging GPUs offer multiple parallelism levels; however, properly managing their computational resources becomes a very challenging task. In fact, management issues become even more difficult when multiple GPUs are involved, adding one more level of parallelism. In this paper, the performance of an acoustic source localization system using distributed microphones is analyzed over a massive multichannel processing framework in a multi-GPU system. The paper evaluates and points out the influence that the number of microphones and the available computational resources have in the overall system performance. Several acoustic environments are considered to show the impact that noise and reverberation have in the localization accuracy and how the use of massive microphone systems combined with parallelized GPU algorithms can help to mitigate substantially adverse acoustic effects. In this context, the proposed implementation is able to work in real time with high-resolution spatial grids and using up to 48 microphones. These results confirm the advantages of suitable GPU architectures in the development of real-time massive acoustic signal processing systems.

Adaptive step ODE algorithms for the 3D simulation of electric heart activity with graphics processing units

In this paper we studied the implementation and performance of adaptive step methods for large systems of ordinary differential equations systems in graphics processing units, focusing on the simulation of three-dimensional electric cardiac activity. The Rush-Larsen method was applied in all the implemented solvers to improve efficiency. We compared the adaptive methods with the fixed step methods, and we found that the fixed step methods can be faster while the adaptive step methods are better in terms of accuracy and robustness.

Designing polylibraries to speed up linear algebra computations

In this paper, we analyse the design of polylibraries, where the programs call for routines from different libraries according to the characteristics of the problem and of the system used to solve it. An architecture for this type of library is proposed. Our aim is to develop a methodology which can be used in the design of parallel libraries. To evaluate the viability of the proposed method, the typical linear algebra libraries hierarchy has been considered. Experiments have been performed in different systems and with linear algebra routines from different levels of the hierarchy. The results confirm the design of polylibraries as a good technique for speeding up computations.

An Efficient Parallel Algorithm to Solve Block-Toeplitz Systems

In this paper, we present an efficient parallel algorithm to solve Toeplitz–block and block–Toeplitz systems in distributed memory multicomputers. This algorithm parallelizes the Generalized Schur Algorithm to obtain the semi-normal equations. Our parallel implementation reduces the communication cost and optimizes the memory access. The experimental analysis on a cluster of personal computers shows the scalability of the implementation. The algorithm is portable because it is based on standard tools and libraries, such as ScaLAPACK and MPI.

Empirical Modelling of Parallel Linear Algebra Routines

This paper shows some ways of combining empirical studies with the theoretical modellization of parallel linear algebra routines. With this combination the accuracy of the model is improved, and the model can be used to take some decisions which facilitate the reduction of the execution time. Experiments with the QR and Cholesky factorizations are shown.

Parallel Algorithms for the Solution of Toeplitz Systems of Linear Equations

In this paper we present two parallel algorithms to solve non-symmetric Toeplitz systems of linear equations. The first algorithm performs a modified QR factorization of the matrix by using the generalized Schur algorithm. The second one is based on the transformation of the Toeplitz matrix into a Cauchy-like matrix in order to reduce the communication cost. Both sequential methods have small computational cost. This fact makes it difficult to implement efficient parallel algorithms. We have tested the efficiency and stability of the algorithms on a cluster of personal computers. The results show the speed-up reaches the number of processors in many cases and both algorithms offer an accurate solution of the linear system. Besides, we have used public domain computation and communication libraries in order to get portable codes.

A New Parallel Approach to the Toeplitz Inverse Eigenproblem Using Newton-like Methods

In this work we describe several portable sequential and parallel algorithms for solving the inverse eigenproblem for Real Symmetric Toeplitz matrices. The algorithms are based on Newtons method (and some variations), for solving nonlinear systems. We exploit the structure and some properties of Toeplitz matrices to reduce the cost, and use Finite Difference techniques to approximate the Jacobian matrix. With this approach, the storage cost is considerably reduced, compared with parallel algorithms proposed by other authors. Furthermore, all the algorithms are efficient in computational cost terms. We have implemented the parallel algorithms using the parallel numerical linear algebra library SCALAPACK based on the MPI environment. Experimental results have been obtained using two different architectures: a shared memory multiprocessor, the SGI PowerChallenge, and a cluster of Pentium II PCs connected through a Myrinet network. The algorithms obtained show a good scalability in most cases.