scientific article; zbMATH DE number 194119

From MaRDI portal
Publication:4040908





The present book is based, mainly, on the results of the author's PhD thesis. The premises of the current approach are that very high level syntax is more suitable for optimization because it leaves the automatic discovery of program parallelism to the compiler, independently of the (super)computer various architectures. Chapter 2 is dealing with data dependence (DD) analysis (the main topic of the book). There are (in Fortran, Pascal, C) three kinds of DD: flow- dependence, anti-dependence, and output-dependence. It is explained how compilers can build a dependence graph: nodes are statements of the program, and directed arcs are dependence (precedence) relations. Each DD arc is augmented with a DD direction vector, containing informations about the do loop iterations, at the head and the tail of the DD arc. Computing a dependence graph with the associated DD direction vectors is time consuming but an important task for optimizing compilers. A program for multiple index exact test is given. Chapter 3 considers the task of automatic vectorization of serial computer programming languages constructions into vector machine instructions. Chapter 4 refers to the process of translating serial loops into code for (concurrent of parallel) execution on (shared-memory) multiprocessor computers. Chapter 5 analyses the inverse problem of loop distribution: loop fusion. This is important for compilers used in virtual memory machines, vector register machines etc. A validity test for loop fusion is discussed. Chapter 6 attacks one of the most powerful program transformations: loop interchanging (the process of switching inner and outer loops). DD tests are exposed for different machine architectures, and different kinds of loops. Chapter 7 faces two problems: (1) The roundoff error accumulation when dealing with reduction operations; (2) The fast execution of the (first-order linear) recurrence relations. Chapter 8 gives an alternate formulation of the wavefront method based on loop skewing and loop interchanging. Chapter 9 discusses strategies to manage the compiler temporary arrays for different architectures, while Chapter 10 shows that vectorizing the while loops is difficult, and often impossible.




Cited in
(33)








This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4040908)