Abstract. The paper begins with an axiomatic definition of rounded arithmetic. The concepts of rounding and of rounded arithmetic operations are defined in an axiomatic manner fully independent of special data formats and encodings. Basic properties of floating-point and interval arithmetic can directly be derived from this abstract model. Interval operations are defined as set operations for elements of the set IR of closed and connected sets of real numbers. As such they form an algebraically closed subset of the powerset of the real numbers. This property leads to explicit formulas for the arithmetic operations of floating-point intervals of IF, which are executable on the computer. Arithmetic for intervals of IF forms an exception free calculus, i.e., arithmetic operations for intervals of IF always lead to intervals of IF again. Later sections are concerned with programming support and hardware for interval arithmetic. Section 9 illustrates that interval arithmetic as developed in this paper has already a long tradition. Products based on these ideas have been available since 1980. Implementing what the paper advocates would have a profound effect on mathematical software. Modern processor architecture comes quite close to what is requested in this paper.

Interval arithmetic has emerged to solve problems with uncertain parameters which are represented by upper and lower bounds. In rectangular coordinate systems, the basic interval operations and improved interval algorithms have been developed and adopted in the numerical analysis. On the other hand, in polar coordinate systems, interval arithmetic still suffers from significant issues of complex computation and overestimation. This paper defines a polar affine quantity and develops a polar affine arithmetic (PAA) that extends affine arithmetic to the polar coordinate systems, which performs much better in many aspects than the corresponding polar interval arithmetic (PIA). Basic arithmetic operations are developed based on the complex affine arithmetic. The Chebyshev approximation theory and the min-range approximation theory are used to identify the best affine approximation of quantities. PAA can accurately keep track the interdependency among multiple variables throughout the calculation procedure, which prominently reduces the solution conservativeness. Numerical case studies in MATLAB programs show that, compared with benchmark results from the Monte Carlo (MC) method, the proposed PAA ensures the completeness of the exact solution, while presenting a much more compact solution region than PIA. PAA has a great potential in research fields including numerical analysis, computer graphics, and engineering optimization.

A randomized algorithm for computing a so called UTV factorization efficiently is presented. Given a matrix $A$, the algorithm "randUTV" computes a factorization $A = UTV^{*}$, where $U$ and $V$ have orthonormal columns, and $T$ is triangular (either upper or lower, whichever is preferred). The algorithm randUTV is developed primarily to be a fast and easily parallelized alternative to algorithms for computing the Singular Value Decomposition (SVD). randUTV provides accuracy very close to that of the SVD for problems such as low-rank approximation, solving ill-conditioned linear systems, determining bases for various subspaces associated with the matrix, etc. Moreover, randUTV produces highly accurate approximations to the singular values of $A$. Unlike the SVD, the randomized algorithm proposed builds a UTV factorization in an incremental, single-stage, and non-iterative way, making it possible to halt the factorization process once a specified tolerance has been met. Numerical experiments comparing the accuracy and speed of randUTV to the SVD are presented. These experiments demonstrate that in comparison to column pivoted QR, which is another factorization that is often used as a relatively economic alternative to the SVD, randUTV compares favorably in terms of speed while providing far higher accuracy.

We present high performance GPU implementations of matvec and compression operations for the H2-variant of hierarchical matrices. H2 matrices, an algebraic generalization of FMM, are space and time efficient representations of dense matrices that exploit the low rank structure of matrix blocks at different levels of granularity and employ both a hierarchical block partitioning and hierarchical bases for the block representations. These two operations are at the core of algebraic operations for hierarchical matrices, the matvec being a ubiquitous operation in numerical algorithms and compression representing a key building block for algebraic operations which require periodic recompression during execution. The difficulties in developing efficient GPU algorithms come primarily from the irregular tree data structures that underlie the hierarchical representations, and the key to performance is to expose fine grained parallelism by recasting the computations on flattened trees and marshaling the irregularly laid out data in ways that allow batched linear algebra operations to be performed. Our numerical results on covariance matrices from 2D and 3D problems from spatial statistics show the high efficiency our routines achieve: over 550 GB/s for the bandwidth-limited matrix-vector operation and over 850 GFLOPS/s for the compression operation on the P100 Pascal GPU.

Matrices coming from elliptic Partial Differential Equations have been shown to have a low-rank property which can be efficiently exploited in multifrontal solvers to provide a substantial reduction of their complexity. Among the possible low-rank formats, the Block Low-Rank format (BLR) is easy to use in a general purpose multifrontal solver and its potential compared to standard (full-rank) solvers has been demonstrated. Recently, new variants have been introduced and it was proved that they can further reduce the complexity but their performance has never been analyzed. In this paper, we present a multithreaded BLR factorization, and analyze its efficiency and scalability in shared-memory multicore environments. We identify the challenges posed by the use of BLR approximations in multifrontal solvers and put forward several algorithmic variants of the BLR factorization that overcome these challenges by improving its efficiency and scalability. We illustrate the performance analysis of the BLR multifrontal factorization with numerical experiments on a large set of problems coming from a variety of real-life applications.

#### Algorithm xxx: The 2D Tree Sliding Window Discrete Fourier Transform

Lee Richardson (Carnegie Mellon University); William Eddy (Carnegie Mellon University)The recent version of the PLASMA (Parallel Linear Algebra Software for Multicore Architectures) library is based on tasks with dependencies from the OpenMP standard. The main functionality of the library is presented. Extensive benchmarks are targeted on three recent multicore and manycore architectures, namely an Intel Xeon, Intel Xeon Phi, and IBM POWER 8 processors.

Batched dense linear algebra kernels are becoming ubiquitous in scientific applications, ranging from tensor contractions in deep learning to data compression in hierarchical low rank matrix approximation. Within a single API call, these kernels are capable of simultaneously launching up to thousands of similar matrix computations, removing the expensive overhead of multiple API calls while increasing the occupancy of the underlying hardware. A challenge is that for the existing hardware landscape (x86, GPUs, etc.) only a subset of the required batched operations is implemented by the vendors, with limited support for very small problem sizes. We describe the design and performance of a new class of batched triangular dense linear algebra kernels on very small data sizes using single and multiple GPUs. By deploying recursive formulations, stressing the register usage, maintaining data locality, reducing threads synchronization and fusing successive kernel calls, the new batched kernels outperform existing state-of-the-art implementations.