<p>In this thesis we address the problem of integrated software pipelining for clustered VLIW architectures. The phases that are integrated and solved as one combined problem are: cluster assignment, instruction selection, scheduling, register allocation and spilling.</p><p>As a first step we describe two methods for integrated code generation of basic blocks. The first method is optimal and based on integer linear programming. The second method is a heuristic based on genetic algorithms.</p><p>We then extend the integer linear programming model to modulo scheduling. To the best of our knowledge this is the first time anybody has optimally solved the modulo scheduling problem for clustered architectures with instruction selection and cluster assignment integrated.</p><p>We also show that optimal spilling is closely related to optimal register allocation when the register files are clustered. In fact, optimal spilling is as simple as adding an additional virtual register file representing the memory and have transfer instructions to and from this register file corresponding to stores and loads.</p><p>Our algorithm for modulo scheduling iteratively considers schedules with increasing number of schedule slots. A problem with such an iterative method is that if the initiation interval is not equal to the lower bound there is no way to determine whether the found solution is optimal or not. We have proven that for a class of architectures that we call transfer free, we can set an upper bound on the schedule length. I.e., we can prove when a found modulo schedule with initiation interval larger than the lower bound is optimal.</p><p>Experiments have been conducted to show the usefulness and limitations of our optimal methods. For the basic block case we compare the optimal method to the heuristic based on genetic algorithms.<em></em></p><p><em>This work has been supported by The Swedish national graduate school in computer science (CUGS) and Vetenskapsrådet (VR).</em></p>

Instruction scheduling and Software pipelining are important compilation techniques which reorder instructions in a program to exploit instruction level parallelism. They are essential for enhancing instruction level parallelism in architectures such as very Long Instruction Word and tiled processors. This thesis addresses two important problems in the context of these instruction reordering techniques. The first problem is for general purpose applications and architectures, while the second is for media and graphics applications for tiled and multi-core architectures. The first problem deals with software pipelining which is an instruction scheduling technique that overlaps instructions from multiple iterations. Software pipelining increases the register pressure and hence it may be required to introduce spill instructions. In this thesis, we model the problem of register allocation with optimal spill code generation and scheduling in software pipelined loops as a 0-1 integer linear program. By minimizing the amount of spill code produced, the formulation ensures that the initiation interval (II) between successive iterations of the loop is not increased unnecessarily. Experimental results show that our formulation performs better than the existing heuristics by preventing an increase in the II and also generating less spill code on average among loops extracted from Perfect Club and SPEC benchmarks. The second major contribution of the thesis deals with the code size aware scheduling of stream programs. Large scale synchronous dataflow graphs (SDF’s) and StreamIt have emerged as powerful programming models for high performance streaming applications. In these models, a program is represented as a dataflow graph where each node represents an autonomous filter and the edges represent the channels through which the nodes communicate. In constructing static schedules for programs in these models, it is important to optimize the execution time buffer requirements of the data channel and the space required to store the encoded schedule. Earlier approaches have either given priority to one of the requirements or proposed ad-hoc methods for generating schedules with good trade-offs. In this thesis, we propose a genetic algorithm framework based on non-dominated sorting for generating serial schedules which have good trade-off between code size and buffer requirement. We extend the framework to generate software pipelined schedules for tiled architectures. From our experiments, we observe that the genetic algorithm framework generates schedules with good trade-off and performs better than the earlier approaches.

The main resources for computation are time, space and energy. Reducing them is the main challenge in the field of processor performance.In this thesis, we are interested in a fourth factor which is information. Information has an important and direct impact on these three resources. We show how it contributes to performance optimization. Landauer has suggested that independently on the hardware where computation is run information erasure generates dissipated energy. This is a fundamental result of thermodynamics in physics. Therefore, under this hypothesis, only reversible computations where no information is ever lost, are likely to be thermodynamically adiabatic and do not dissipate power. Reversibility means that data can always be retrieved from any point of the program. Information may be carried not only by the data but also by the process and input data that generate it. When a computation is reversible, information can also be retrieved from other already computed data and reverse computation. Hence reversible computing improves information locality.This thesis develops these ideas in two directions. In the first part, we address the issue of making a computation DAG (directed acyclic graph) reversible in terms of spatial complexity. We define energetic garbage as the additional number of registers needed for the reversible computation with respect to the original computation. We propose a reversible register allocator and we show empirically that the garbage size is never more than 50% of the DAG size. In the second part, we apply this approach to the trade-off between recomputing (direct or reverse) and storage in the context of supercomputers such as the recent vector and parallel coprocessors, graphical processing units (GPUs), IBM Cell processor, etc., where the gap between processor cycle time and memory access time is increasing. We show that recomputing in general and reverse computing in particular helps reduce register requirements and memory pressure. This approach of reverse rematerialization also contributes to the increase of instruction-level parallelism (Cell) and thread-level parallelism in multicore processors with shared register/memory file (GPU). On the latter architecture, the number of registers required by the kernel limits the number of running threads and affects performance. Reverse rematerialization generates additional instructions but their cost can be hidden by the parallelism gain. Experiments on the highly memory demanding Lattice QCD simulation code on Nvidia GPU show a performance gain up to 11%.

In this thesis we address the problem of integrated software pipelining for clustered VLIW architectures. The phases that are integrated and solved as one combined problem are: cluster assignment, instruction selection, scheduling, register allocation and spilling. As a first step we describe two methods for integrated code generation of basic blocks. The first method is optimal and based on integer linear programming. The second method is a heuristic based on genetic algorithms. We then extend the integer linear programming model to modulo scheduling. To the best of our knowledge this is the first time anybody has optimally solved the modulo scheduling problem for clustered architectures with instruction selection and cluster assignment integrated. We also show that optimal spilling is closely related to optimal register allocation when the register files are clustered. In fact, optimal spilling is as simple as adding an additional virtual register file representing the memory and have transfer instructions to and from this register file corresponding to stores and loads. Our algorithm for modulo scheduling iteratively considers schedules with increasing number of schedule slots. A problem with such an iterative method is that if the initiation interval is not equal to the lower bound there is no way to determine whether the found solution is optimal or not. We have proven that for a class of architectures that we call transfer free, we can set an upper bound on the schedule length. I.e., we can prove when a found modulo schedule with initiation interval larger than the lower bound is optimal. Experiments have been conducted to show the usefulness and limitations of our optimal methods. For the basic block case we compare the optimal method to the heuristic based on genetic algorithms. This work has been supported by The Swedish national graduate school in computer science (CUGS) and Vetenskapsrådet (VR).

Les trois principales ressources du calcul sont le temps, l'espace et l'énergie, les minimiser constitue un des défis les plus importants de la recherche de la performance des processeurs.Dans cette thèse, nous nous intéressons à un quatrième facteur qui est l'information. L'information a un impact direct sur ces trois facteurs, et nous montrons comment elle contribue ainsi à l'optimisation des performances. Landauer a montré que c’est la destruction - logique - d’information qui coûte de l’énergie, ceci est un résultat fondamental de la thermodynamique en physique. Sous cette hypothèse, un calcul ne consommant pas d’énergie est donc un calcul qui ne détruit pas d’information. On peut toujours retrouver les valeurs d’origine et intermédiaires à tout moment du calcul, le calcul est réversible. L'information peut être portée non seulement par une donnée mais aussi par le processus et les données d’entrée qui la génèrent. Quand un calcul est réversible, on peut aussi retrouver une information au moyen de données déjà calculées et du calcul inverse. Donc, le calcul réversible améliore la localité de l'information. La thèse développe ces idées dans deux directions. Dans la première partie, partant d'un calcul, donné sous forme de DAG (graphe dirigé acyclique), nous définissons la notion de « garbage » comme étant la taille mémoire – le nombre de registres - supplémentaire nécessaire pour rendre ce calcul réversible. Nous proposons un allocateur réversible de registres, et nous montrons empiriquement que le garbage est au maximum la moitié du nombre de noeuds du graphe.La deuxième partie consiste à appliquer cette approche au compromis entre le recalcul (direct ou inverse) et le stockage dans le contexte des supercalculateurs que sont les récents coprocesseurs vectoriels et parallèles, cartes graphiques (GPU, Graphics Processing Unit), processeur Cell d’IBM, etc., où le fossé entre temps d’accès à la mémoire et temps de calcul ne fait que s'aggraver. Nous montons comment le recalcul en général, et le recalcul inverse en particulier, permettent de minimiser la demande en registres et par suite la pression sur la mémoire. Cette démarche conduit également à augmenter significativement le parallélisme d’instructions (Cell BE), et le parallélisme de threads sur un multicore avec mémoire et/ou banc de registres partagés (GPU), dans lequel le nombre de threads dépend de manière importante du nombre de registres utilisés par un thread. Ainsi, l’ajout d’instructions du fait du calcul inverse pour la rematérialisation de certaines variables est largement compensé par le gain en parallélisme. Nos expérimentations sur le code de Lattice QCD porté sur un GPU Nvidia montrent un gain de performances atteignant 11%. / The main resources for computation are time, space and energy. Reducing them is the main challenge in the field of processor performance.In this thesis, we are interested in a fourth factor which is information. Information has an important and direct impact on these three resources. We show how it contributes to performance optimization. Landauer has suggested that independently on the hardware where computation is run information erasure generates dissipated energy. This is a fundamental result of thermodynamics in physics. Therefore, under this hypothesis, only reversible computations where no information is ever lost, are likely to be thermodynamically adiabatic and do not dissipate power. Reversibility means that data can always be retrieved from any point of the program. Information may be carried not only by the data but also by the process and input data that generate it. When a computation is reversible, information can also be retrieved from other already computed data and reverse computation. Hence reversible computing improves information locality.This thesis develops these ideas in two directions. In the first part, we address the issue of making a computation DAG (directed acyclic graph) reversible in terms of spatial complexity. We define energetic garbage as the additional number of registers needed for the reversible computation with respect to the original computation. We propose a reversible register allocator and we show empirically that the garbage size is never more than 50% of the DAG size. In the second part, we apply this approach to the trade-off between recomputing (direct or reverse) and storage in the context of supercomputers such as the recent vector and parallel coprocessors, graphical processing units (GPUs), IBM Cell processor, etc., where the gap between processor cycle time and memory access time is increasing. We show that recomputing in general and reverse computing in particular helps reduce register requirements and memory pressure. This approach of reverse rematerialization also contributes to the increase of instruction-level parallelism (Cell) and thread-level parallelism in multicore processors with shared register/memory file (GPU). On the latter architecture, the number of registers required by the kernel limits the number of running threads and affects performance. Reverse rematerialization generates additional instructions but their cost can be hidden by the parallelism gain. Experiments on the highly memory demanding Lattice QCD simulation code on Nvidia GPU show a performance gain up to 11%.