The thesis studies theory and methods for generating automatically centralized and distributed implementations from a high-level model of an application software in BIP. BIP (Behavior, Interaction, Priority) is a component framework with formal operational semantics. Coordination between components is achieved by using multiparty interactions and dynamic priorities for scheduling interactions. A key idea is to use a set of correct source-to-source transformations preserving the functional properties of a given application software. By application of these transformations we can generate a full range of implementations from centralized to fully distributed. Centralized Implementation: the implementation method transforms the interactions of an application software described in BIP and generates a functionally equivalent program. The method is based on the successive application of three types of source-to-source transformations: flattening of components, flattening of connectors and composition of atomic components. We shown that the system of the transformations is confluent and terminates. By exhaustive application of the transformations, any BIP component can be transformed into an equivalent monolithic component. From this component, efficient standalone C++ code can be generated. Distributed Implementation: the implementation method transforms an application software described in BIP for a given partition of its interactions, into a Send/Receive BIP model. Send/Receive BIP models consist of components coordinated by using asynchronous message passing (Send/Receive primitives). The method leads to 3-layer architectures. The bottom layer includes the components of the application software where atomic strong synchronization is implemented by sequences of Send/Receive primitives. The second layer includes a set of interaction protocols. Each protocol handles the interactions of a class of the given partition. The third layer implements a conflict resolution protocol used to resolve conflicts between conflicting interactions of the second layer. Depending on the given partition, the execution of obtained Send/Receive BIP model range from centralized (all interactions in the same class) to fully distributed (each class has a single interaction). From Send/Receive BIP models and a given mapping of their components on a platform providing Send/Receive primitives, an implementation is automatically generated. For each class of the partition we generate C++ code implementing the global behavior of its components. The transformations have been fully implemented and integrated into BIP tool-set. The experimental results on non trivial examples and case studies show the novelty and the efficiency of our approach.

This thesis presents SkePU 2, the next generation of the SkePU C++ framework for programming of heterogeneous parallel systems using the skeleton programming concept. SkePU 2 is presented after a thorough study of the state of parallel programming models, frameworks and tools, including other skeleton programming systems. The advancements in SkePU 2 include a modern C++11 foundation, a native syntax for skeleton parameterization with user functions, and an entirely new source-to-source translator based on Clang compiler front-end libraries. SkePU 2 extends the functionality of SkePU 1 by embracing metaprogramming techniques and C++11 features, such as variadic templates and lambda expressions. The results are improved programmability and performance in many situations, as shown in both a usability survey and performance evaluations on high-performance computing hardware. SkePU’s skeleton programming model is also extended with a new construct, Call, unique in the sense that it does not impose any predefined skeleton structure and can encapsulate arbitrary user-defined multi-backend computations. We conclude that SkePU 2 is a promising new direction for the SkePU project, and a solid basis for future work, for example in performance optimization.

Logic programming is appealing since it allows the programmer to concentrate on the meaning of the problem to be solved. Unfortunately, for efficiency reasons, the declarative and operational natures of Prolog do not coincide. Prolog uses an incomplete depth-first search rule, unifications and negations may be unsound, and there are extralogical features like the cut or dynamic predicates. Methodologies have been proposed to construct operationally correct and efficient Prolog code. Researchers have designed methods to automate the verification of operational properties on which optimisation of logic programs can be based. A few tools have been implemented but there is a lack of a unified framework. <P> The goal and topic of this thesis is the design, implementation and evaluation of an abstract interpretation framework of Prolog to integrate state-of-the-art techniques. The analyser is based on an original proposal that defines the notion of abstract sequence, which allows one to verify many desirable operational properties of a logic procedure. The properties include types, modes, sharing of terms, proving termination, linear relations between the size of input/output terms and the number of solutions to a call. A single global analysis is performed, and abstract sequences are derived at each program point. <P> In this thesis, we implement and evaluate the original framework, and, more importantly, we overcome its limitations to make it accurate and usable in practice: the improved framework accepts any Prolog code with modules, new abstract domains and operations are added, and the language of specifications is more expressive. We also design and implement an optimiser that generates specialised code. The optimiser uses the abstract information to safely apply source-to-source transformations. Code transformations include clause and literal reordering, introduction of cuts, and removal of redundant literals. The optimiser follows a precise strategy to choose the most rewarding transformations in best order.