Les applications multimédia se composent généralement de blocs numériques exhibant des schémas de calcul flottant réguliers. Sur les processeurs sans support architectural pour l'arithmétique flottante, ils peuvent être profitablement transformés en opérateurs dédiés, s'ajoutant aux 5 opérateurs élémentaires (+, -, X, / et √) : en traitant plus d'opérations simultanément, ils permettent d'obtenir de meilleures performances. Cette thèse porte sur la conception de tels opérateurs, et les techniques de compilation mises en œuvre pour les sélectionner. Nous avons réalisé des implémentations optimisées pour un ensemble d'opérateurs dédiés : élévation au carré, mise à l'échelle, fused multiply-add, produit scalaire en dimension deux (DP2), addition/soustraction simultané et sinus/cosinus simultanés. En proposant de nouveaux algorithmes cherchant à maximiser le parallélisme d'instructions et détaillés ici, nous obtenons des accélérations d'un facteur allant jusqu'à 4.2 par appel. Nous détaillons également les changements apportés dans le compilateur pour effectuer la sélection. La plupart des opérateurs sont sélectionnés au niveau syntaxique. Cependant, pour certains opérateurs, nous avons dû améliorer l'analyse d'intervalles entiers pour prendre en compte les variables de type flottant, afin de prouver certaines conditions de positivité requises à leur sélection. Enfin, nous apportons la preuve en pratique de la pertinence de cette approche : sur des noyaux typiques du traitement du signal et sur certaines applications, nous mesurons une amélioration de performance allant jusqu'à 1.59x en comparaison avec la performance obtenue avec les seuls opérateurs élémentaires. / Media processing applications typically involve numerical blocks that exhibit regular floating-point computation patterns. For processors whose architecture supports only integer arithmetic, these patterns can be profitably turned into custom operators, coming in addition to the five basic ones (+, -, X, / and √), but achieving better performance by treating more operations. This thesis addresses the design of such custom operators as well as the techniques developed in the compiler to select them in application codes. We have designed optimized implementations for a set of custom operators which includes squaring, scaling, adding two nonnegative terms, fused multiply-add, fused square-add (x*x+z, with z>=0), two-dimensional dot products (DP2), sums of two squares, as well as simultaneous addition/subtraction and sine/cosine. With novel algorithms targeting high instruction-level parallelism and detailed here for squaring, scaling, DP2, and sin/cos, we achieve speedups of up to 4.2x for individual custom operators even when subnormal numbers are fully supported. Furthermore, we introduce the optimizations developed in the ST231 C/C++ compiler for selecting such operators. Most of the selections are achieved at high level, using syntactic criteria. However, for fused square-add, we also enhance the framework of integer range analysis to support floating-point variables in order to prove the required positivity condition z>= 0. Finally, we provide quantitative evidence of the benefits to support this selection of custom operations: on DSP kernels and benchmarks, our approach allows us to be up to 1.59x faster compared to the sole usage of basic ones.

Modern embedded processors with dedicated address generation unit support memory accesses using indirect addressing mode with auto-increment and auto-decrement. The auto-increment/decrement mode, if properly utilized, can save address arithmetic instructions, reduce static and dynamic footprint of the program and speed up the execution as well. We propose an optimization framework for embedded processors based on the auto-increment and decrement addressing modes for address registers. Existing work on this class of optimizations focuses on using an access graph and finding the maximum weight path cover to find an optimized stack variables layout. We take this further by using coalescing, addressing mode selection and offset registers to find further opportunities for reducing the number of load-address instructions required. We also propose an algorithm for building the layout with considerations for memory accesses across basic blocks, because existing work mainly considers intra-basic-block information. We then use the available offset registers to try to further reduce the number of address arithmetic instructions after layout assignment.

Green computing techniques aim to reduce the power foot print of modern embedded devices with particular emphasis on processors, the power hot-spots of these devices. In this thesis we propose compiler-driven and profile-driven optimizations that reduce power consumption in a modern embedded processor. We show that these optimizations reduce power consumption in functional units and memory subsystems with very low performance loss. We present three new techniques to reduce power consumption in processors, namely, transition aware scheduling, leakage reduction in data caches using criticality analysis, and dynamic power reduction in data caches using locality analysis of data regions. A novel instruction scheduling technique to address leakage power consumption in functional units is proposed. This scheduling technique, transition aware scheduling, is motivated by idle periods that arise in the utilization of functional units during program execution. A continuously large idle period in a functional unit can be exploited to place the unit in low power state. This novel scheduling algorithm increases the duration of idle periods without hampering performance and drives power gating in these periods. A power model defined with idle cycles as a parameter shows that this technique saves up to 25% of leakage power with very low performance impact. In modern embedded programs, data regions can be classified as critical and non-critical. Critical data regions significantly impact the performance. A new technique to identify such data regions through profiling is proposed. This technique along with a new criticality based cache policy is used to control the power state of the data cache. This scheme allocates non-critical data regions to low-power cache regions, thereby reducing leakage power consumption by up to 40% without compromising on the performance. This profiling technique is extended to identify data regions that have low locality. Some data regions have high data reuse. A locality based cache policy based on cache parameters like size and associativity is proposed. This scheme reduces dynamic as well as static power consumption in the cache subsystem. This optimization reduces 25% of the total power consumption in the data caches without hampering the execution time. In this thesis, the problem of power consumption of a program is decoupled from the number of processor cores. The underlying architecture model is simplified to abstract away a variety of processor scenarios. This simplified model can be scaled up to be implemented in various multi-core architecture models like Chip Multi-Processors, Simultaneous Multi-Threaded Processors, Chip Multi-Threaded Processors, to name a few. The three techniques proposed in this thesis leverage underlying hardware features like low power functional units, drowsy caches and split data caches. These techniques reduce power consumption of a wide range of benchmarks with low performance loss.