Electronic structure calculations of organic molecules are an important set of tools to gain understanding of molecular structures. This thesis presents two separate contributions to applying quantum chemistry to organic molecules. In the first section, the computational cost of a post-Hartree-Fock method is improved for large molecules by using graphical processing units. In this work, the resolution-of-the-identity second-order Møller-Plesset perturbation theory (RI-MP2) algorithm was adapted to send the large matrix multiplication steps to be run on a graphics co-processor. As a result, the calculations were performed up to 15x faster than a standard implementation for large molecules such as taxol. In the second section of the thesis, density functional theory is used to predict the molecular dipole moments of molecules that form self-assembled monolayers (SAMs) on metal surfaces. The dipole moment of the molecule that is aligned perpendicular to the surface in a SAM changes the work function of the surface. The calculated dipole moments correlate with the current density measured for the junctions by experimental collaborators. This result holds for a series of alkane chains with even and odd numbers of carbons and for molecules that have an amide group substituted for an ethylene unit. / Chemistry and Chemical Biology

The main benefit of the work is the optimization of the hardware configuration for the calculation of neural networks. The theoretical part describes neural networks, deep learning frameworks and hardware options. The next part of the thesis deals with implementation of performance tests, which include application of Inception V3 and ResNet models. Network models are applied to various graphics cards and computing hardware. The output of the thesis is the implemented model of the network Inception V3, which examines the graphics cards and their performance, time-consuming calculations and their efficiency. The ResNet model is applied to a section that examines other impacts on neural network computing such as used disk, operating memory, and so on. Each practical part contains a discussion where the knowledge of the given part is explained. In the case of consumption measurement, a mismatch between the declaration by the manufacturer and the measured values was identified.

Massively Parallel Processors provide significantly higher peak performance figures than other forms of general purpose processors. However, this comes at a cost to the developer, who needs to deal with an increasingly complicated piece of hardware, for which applications need to be tweaked and optimised to achieve high performance. Domain-specific languages have been proposed as a potential solution to this complexity problem: generating GPU applications from high-level, declarative specifications. This thesis explores two related ideas: firstly, is it practical to synthesise DSLs from high-level languages, and secondly, how can we simplify the creation of such DSLs? This thesis proposes a novel approach whereby rather than considering single domains, we consider collections of collaborative domains in order to share common features and thus reduce the cost of development. We achieve this using a DSLs-within-a-DSL approach: a custom designed host language, into which extensions may be embedded. In order to ground our approach in a real case-study, we propose, design and develop a DSLs-within-a-DSL framework for bioinformatics. We use a restricted recursive functional language as the host language, and embed new DSLs into this language. Importantly, we describe how we can use a combination of novel and adopted automatic parallelisation techniques to synthesise a massively-parallel program for a GPU. This automatic parallelisation, achieved through the discovery of a schedule, and program synthesis techniques using the polyhedral model, interacts productively with our embedded extensions. To further simplify development, we provide a series of customisable heuristics for defining GPU parameters such as the block size (number of threads), grid size and location in the memory hierarchy of data-structures. This encapsulates GPU expertise within the compiler itself. We finally demonstrate that the total combination of these techniques results in applications with competitive performance, at much lower development cost and greater flexibility than comparable hand-coded applications.

Afin de créer des productions toujours plus réalistes, les industries du jeu vidéo et du cinéma cherchent à générer des environnements de plus en plus larges et complexes. Cependant, la modélisation manuelle des objets 3D dans de tels décors se révèle très coûteuse. A l’inverse, les méthodes de génération procédurale permettent de créer facilement une grande variété d’objets, tels que les plantes et les bâtiments. La modélisation par règles de grammaire offre un outil de haut niveau pour décrire ces objets, mais utiliser correctement ces règles s’avère très souvent compliqué. De plus, aucune solution de modélisation basée grammaire ne supporte l’édition et la visualisation d’environnements massifs en temps interactif. Dans un tel scénario, les artistes doivent modifier les objets en dehors de la scène avant de voir le résultat intégré.Dans ces travaux de recherche, nous nous intéressons à la génération procédurale et au rendu d’environnements à grande échelle. Nous voulons aussi faciliter la tâche des artistes avec des outils intuitifs de contrôle de grammaires. Tout d’abord nous proposons un système permettant la génération procédurale en parallèle sur le GPU en temps interactif. Pour cela, nous adoptons une approche d’expansion indépendante par segment, permettant une amplification des données en parallèle. Nous étendons ce système pour générer des modèles basés sur une structure interne, tels que les toits. Nous présentons aussi une solution utilisant des contextes externes pour contrôler facilement les grammaires par le biais de surface ou de texture. Pour finir nous intégrons un système de niveaux de détails et des techniques d’optimisation permettant la génération, l’édition et la visualisation interactives d’environnements à grande échelle. Grâce à notre système il est possible de générer et d’afficher interactivement des scènes comprenant des milliers de bâtiments et d’arbres, représentant environ 2 téraoctets de données. / With the increasing computing and storage capabilities of recent hardware, movie and video games industries desire huger realistic environments. However, modeling such sceneries by hand turns out to be highly time consuming and costly. On the other hand, procedural modeling provides methods to easily generate high diversity of elements such as vegetation and architecture. While grammar rules bring a high-level powerful modeling tool, using these rules is often a tedious task, necessitating frustrating trial and error process. Moreover, as no solution proposes real-time generation and rendering for massive environments, artists have to work on separate parts before integrating the whole and see the results.In this research, we aim to provide interactive generation and rendering of very large sceneries, while offering artist-friendly methods for controlling grammars behavior. We first introduce a GPU-based pipeline providing parallel procedural generation at render time. To this end we propose a segment-based expansion method working on independent elements, thus allowing for parallel amplification. We then extend this pipeline to permit the construction of models relying on internal contexts, such as roofs. We also present external contexts to control grammars with surface and texture data. Finally, we integrate a LOD system with optimization techniques within our pipeline providing interactive generation, edition and visualization of massive environments. We demonstrate the efficiency of our pipeline with a scene comprising hundred thousand trees and buildings each, representing 2 terabytes of data.