The ubiquitous Kronecker product

Fast changing, increasingly complex, and diverse computing platforms pose central problems in scientific computing: How to achieve, with reasonable effort, portable optimal performance? We present SPIRAL, which considers this problem for the performance-critical domain of linear digital signal processing (DSP) transforms. For a specified transform, SPIRAL automatically generates high-performance code that is tuned to the given platform. SPIRAL formulates the tuning as an optimization problem and exploits the domain-specific mathematical structure of transform algorithms to implement a feedback-driven optimizer. Similar to a human expert, for a specified transform, SPIRAL "intelligently" generates and explores algorithmic and implementation choices to find the best match to the computer's microarchitecture. The "intelligence" is provided by search and learning techniques that exploit the structure of the algorithm and implementation space to guide the exploration and optimization. SPIRAL generates high-performance code for a broad set of DSP transforms, including the discrete Fourier transform, other trigonometric transforms, filter transforms, and discrete wavelet transforms. Experimental results show that the code generated by SPIRAL competes with, and sometimes outperforms, the best available human tuned transform library code.

/pdf/spiral-code-generation-for-dsp-transforms-4219bd16gh.pdf

SPIRAL: Code Generation for DSP Transforms

Under consideration is the large body of signal recovery problems that can be formulated as the problem of minimizing the sum of two (not necessarily smooth) lower semicontinuous convex functions in a real Hilbert space. This generic problem is analyzed and a decomposition method is proposed to solve it. The convergence of the method, which is based on the Douglas-Rachford algorithm for monotone operator-splitting, is obtained under general conditions. Applications to non-Gaussian image denoising in a tight frame are also demonstrated.

A Douglas–Rachford Splitting Approach to Nonsmooth Convex Variational Signal Recovery

The FFTW library for computing the discrete Fourier transform (DFT) has gained a wide acceptance in both academia and industry, because it provides excellent performance on a variety of machines (even competitive with or faster than equivalent libraries supplied by vendors). In FFTW, most of the performance-critical code was generated automatically by a special-purpose compiler, called genfft, that outputs C code. Written in Objective Caml, genfft can produce DFT programs for any input length, and it can specialize the DFT program for the common case where the input data are real instead of complex. Unexpectedly, genfft "discovered" algorithms that were previously unknown, and it was able to reduce the arithmetic complexity of some other existing algorithms. This paper describes the internals of this special-purpose compiler in some detail, and it argues that a specialized compiler is a valuable tool.

/pdf/a-fast-fourier-transform-compiler-4w0zmnryds.pdf

A fast Fourier transform compiler

Orthogonal frequency-division multiplexing (OFDM) systems are highly sensitive to synchronization errors. We introduce an algorithm for the blind estimation of symbol timing and carrier frequency offset in wireless OFDM systems. The proposed estimator is an extension of the Gini-Giannakis (see IEEE Trans. Commun., vol.46, p.400-411, 1998) estimator for single-carrier systems. It exploits the cyclostationarity of OFDM signals and relies on second-order statistics only. Our method can be applied to pulse shaping OFDM systems with arbitrary time-frequency guard regions, OFDM based on offset quadrature amplitude modulation, and biorthogonal frequency-division multiplexing systems. We furthermore propose the use of different subcarrier transmit powers (subcarrier weighting) and periodic transmitter precoding to achieve a carrier frequency acquisition range of the entire bandwidth of the OFDM signal, and a symbol timing acquisition range of arbitrary length. Finally, we provide simulation results demonstrating the performance of the new estimator.

Blind estimation of symbol timing and carrier frequency offset in wireless OFDM systems

Contents: Introduction to Abstract Algebra.- Tensor Product and Stride Permutation.- Cooley-Tukey FFF Algorithms.- Variants of FFT Algorithms and Their Implementations.- Good-Thomas PFA.- Linear and Cyclic Convolutions.- Agarwal-Cooley Convolution Algorithm.- Introduction to Multiplicative Fourier Transform Algorithms (MFTA).- MFTA: The Prime Case.- MFTA: Product of Two Distinct Primes.- MFTA: Transform Size N = Mr. M-Composite Integer and r-Prime.- MFTA: Transform Size N = p2.- Periodization and Decimation.- Multiplicative Character and the FFT.- Rationality.- Index.

Algorithms for Discrete Fourier Transform and Convolution

1 Review of algebra.- 1.1 Introduction.- 1.2 Definitions and examples of groups.- 1.3 Subgroups, cosets, and quotients.- 1.4 Ideals.- 1.5 Mappings.- 1.6 Finitely generated abelian groups.- 1.6.1 Cyclic groups.- 1.6.2 Free abelian groups.- References.- 2 Linear algebra and abelian groups.- 2.1 Introduction.- 2.2 Vector space L(A).- 3 Fourier transform over A.- 3.1 Introduction.- 3.2 Character groups.- 3.3 Character formulas.- 3.4 Duality theory.- 3.5 Character group basis.- 3.6 Fourier transform.- 3.7 Shift and multiplication operators.- References.- Problems.- 4 Poisson summation formula.- 4.1 Introduction.- 4.2 Statement and proof.- 4.3 Fourier transform of periodic functions.- 4.4 Periodization-decimation operators.- References.- Problems.- 5 Zak transform.- 5.1 Introduction.- 5.2 Fourier analysis on A x A*.- 5.3 Zak transform.- 5.4 Functional equation.- 5.5 Fourier and Zak transform.- 5.6 Isometry.- 5.7 Algorithm for computing Zak transform.- References.- Problems.- 6 Weyl-Heisenberg systems.- 6.1 Introduction.- 6.2 Translates.- 6.3 W-H systems.- 6.4 Sampling rates.- 6.5 Divide-and-conquer.- References.- Problems.- 7 Zak transform and W-H systems.- 7.1 Introduction.- 7.2 Basic results.- 7.3 Fundamental formulas.- 7.4 Zak space characterization of W-H systems.- 7.4.1 Critical sampling subgroup.- 7.4.2 Integer over-sampling subgroup.- 7.4.3 General sampling subgroup.- 7.5 Zero set characterization.- 7.5.1 Critical sampling subgroup.- 7.5.2 Integer over-sampling subgroup.- Problems.- 8 Algorithms for W-H systems.- 8.1 Introduction.- 8.2 Critical sampling algorithm.- 8.3 Integer over-sampling algorithm.- 8.3.1 Reducing the problem.- 8.4 General over-sampling algorithm.- 8.4.1 Reducing the problem.- References.- Problems.- 9 Orthogonal projection theorem.- 9.1 Introduction.- 9.2 Orthogonal projection algorithm.- 9.3 Iterative W-H coefficient set algorithm.- 10 Cross-ambiguity function.- 10.1 Introduction.- 10.2 Basic results.- 10.3 Direct algorithm.- 10.4 Critical sampling algorithm.- 10.5 Integer over-sampling algorithm.- 10.6 General divide-and-conquer algorithm.- References.- Problems.- 11 Ambiguity surfaces.- 11.1 Introduction.- 11.2 Fourier transform of ambiguity surfaces.- 11.3 Formulas D1 and D2.- References.- 12 Orthonormal W-H systems.- 12.1 Introduction.- 12.2 Orthonormal W-H systems.- 12.2.1 Critical sampling subgroup.- 12.2.2 Integer over-sampling subgroup.- 12.2.3 Over-sampling subgroup ?.- References.- Problems.- 13 Duality.- 13.1 Introduction.- 13.2 Biorthogonal.- 13.3 Algorithms for computing biorthogonals.- 13.3.1 ?-periodization.- 13.3.2 Critical sampling subgroup.- 13.3.3 Integer over-sampling subgroup.- 13.3.4 General over-sampling subgroup.- References.- Problems.- 14 Frames.- 14.1 Introduction.- 14.2 Frame Operator.- 14.2.1 Critical sampling subgroup.- 14.2.2 Integer over-sampling subgroup.- 14.2.3 General over-sampling subgroup.- 14.3 Frames.- 14.4 Frame biorthogonals.- 14.5 Operator interpretation.- 14.6 Tight frames.- References.- 15 Implementation.- 15.1 Introduction.- 15.2 Tensor product.- 15.3 Multidimensional arrays.- 15.3.1 Two-dimensional arrays.- 15.3.2 Multidimensional arrays.- 15.4 Computing the Zak transform.- 15.4.1 1-dimensional Zak transform.- 15.4.2 Two-dimensional Zak transform.- References.- Problems.- 16 Algebra of multirate structures.- 16.1 Introduction.- 16.2 Algebra.- 16.3 Exact sequences.- 16.4 Main theorem.- 16.5 Representatives mapping theorem.- 16.6 Integer matrices.- References.- Problems.- 17 Multirate structures.- 17.1 Introduction.- 17.2 Decimation operator.- 17.2.1 Elementary formulas.- 17.2.2 Shift-invariant operators.- 17.3 Polyphase representation.- 17.3.1 Analysis bands.- 17.3.2 Synthesis bands.- 17.4 Integer rate bands.- 17.5 Reduction theorems.- 17.5.1 Subgroup of a band.- 17.5.2 Nonuniform integer sampling rate filter banks.- 17.5.3 Rational sampling rate filter banks.- 17.6 Decimators and expanders.- References.- 18 A Time-frequency search for stock market anomalies.- 18.1 Introduction.- 18.2 Time-frequency trees.- 18.2.1 Adaptive segmentation via time-frequency trees.- 18.2.2 Recombination to a nondyadic split.- 18.3 Analysis of the log-differenced DJIA and S&P 500 data.- References.

Time-frequency representations

In the following two chapters, we will concentrate on algorithms for computing FFT of size a composite number N. The main idea is to use the additive structure of the indexing set Z/N to define mappings of the input and output data vectors into 2-dimensional arrays. Algorithms are then designed, transforming 2-dimensional arrays which, when combined with these mappings, compute the N-point FFT. The stride permutations of chapter 2 play a major role.

Cooley-Tukey FFT Algorithms

In this and the next chapters, we present several mathematical results needed to design the algorithms of the text. We assume that the reader has some knowledge of groups, rings and vector spaces but no extensive knowledge is required. Instead, we focus on those mathematical objects which will be used repeatedly in this text.

Introduction to Abstract Algebra

Abelian group DSP can be completely described in terms of a special class of signals, the characters, defined by their relationship to the translations defined by abelian group multiplication. The first problem to be faced in extending classical DSP theory is to decide on what is meant by a translation. We have selected certain classes of nonabelian groups and defined translations in terms of left nonabelian group multiplications.

/pdf/group-filters-and-image-processing-1txr0jn96s.pdf

Myoung An

Papers

Algorithms for Discrete Fourier Transform and Convolution

Time-frequency representations

Cooley-Tukey FFT Algorithms

Introduction to Abstract Algebra

Group Filters and Image Processing