A reliable linear algebra library for transputer networks (Q1899466)

The authors present a library for high accuracy basic linear algebra computations on T800 transputer networks. The library is implemented in OCCAM-2. Since some operations -- like directed rounding for floating point operations -- become significantly slower in OCCAM-2, transputer assembler language is used in such critical cases. The library implements an optimal (one rounding only) scalar product as an operation on one processor. Distributed computations of the optimal scalar product are not considered for efficiency reasons. Real interval arithmetic (for the 4 basic operations and the optimal scalar product) is also provided by the library. Furthermore, the library contains a collection of communication routines some of which are implemented in a topology-independent manner. Built upon these elementary operations, the library contains several modules for matrix-vector and matrix-matrix multiplication, matrix inversion and factorization, but also for typical computations needed in validation methods like \(\varepsilon\)-inflation or enclosures for a residual matrix. For researchers, the library is available for free from the authors.