# Browsing by Subject "Error-correcting codes (information theory)"

Now showing 1 - 1 of 1

- Results Per Page
1 5 10 20 40 60 80 100

- Sort Options
Ascending Descending

Item A systematic approach to the design and analysis of linear algebra algorithms(2001-12) Gunnels, Joseph Andrew; Van de Geijn, Robert A.Show more Over the last two decades, much progress has been made in the area of the high-performance sequential and parallel implementation of dense linear algebra operations. At what time can we confidently state that we truly understand this problem area and what form might evidence in support of this assertion take? It is our thesis that if we focus this question on the software architecture of libraries for dense linear algebra operations, we can claim to have reached the point where, for a restricted class of problems, we understand this area. In this dissertation, we provide evidence in support of this assertion by outlining a systematic and partially automated approach to the derivation and high-performance implementation of a large class of dense linear algebra operations. We have arrived at a conclusion that the answer is to apply formal derivation techniques from Computing Science to the development of highperformance linear algebra libraries. The resulting approach has resulted in an aesthetically pleasing, coherent code that facilitates performance analysis, intelligent modularity, and the enforcement of program correctness via assertions. In this dissertation, we illustrate this observation by looking at the development of the Formal Linear Algebra Methods Environment (FLAME) for implementing linear algebra algorithms. We believe that traditional methods of implementation do not reflect the natural manner in which an algorithm is either classified or derived. To remedy this discrepancy, we propose the use of a small set of abstractions that can be used to design and implement linear algebra algorithms in a simple and straightforward manner. These abstractions may be expressed in a script language that can be compiled into efficient executable code. We extend this approach to parallel implementations without adding substantial complexity. It should also be possible to translate these scripts into analytical equations that reflect their performance profiles. These profiles may allow software designers to systematically optimize their algorithms for a given machine or to meet a particular resource goal. Given the more systematic approach to deriving and implementing algorithms that is facilitated by better abstraction and classification techniques, this sort of analysis can be shown to be systematically derivable and automated.Show more