Contributions to the Scalability of Automatic Precision Tuning

Vous êtes cordialement invités à venir assister à la soutenance de thèse de HA Van-Phu, équipe TARAN, le vendredi 10 mars 2023 à 10h00 en salle Pétri/Turing.

Contributions to the Scalability of Automatic Precision Tuning

Energy consumption is one of the major issues in computing today, shared by all domains of computer science, from high-performance computing to embedded systems. In recent years, approximation during computation has received renewed interest to improve energy efficiency. Many applications do not require high precision, thus hardware designers often trade-off the accuracy for cost reduction and speed-up. Various techniques for approximate computing augment the design space by providing another set of design knobs for performance-accuracy trade-off. This thesis focuses on developing methods for systematic exploration of this design space, including performance and accuracy modeling and design automation. We optimize the word length of each data to get the good balance between the cost and the accuracy of the final design. This problem is called Word length Optimization (WLO) or automatic precision tuning. The thesis contributes to three research directions. First, a method is proposed to improve the scalability of WLO for large applications. To reduce exponential complexity in the nature of WLO, the input application is decomposed into smaller kernels, which are then solved independently using noise budgets to reduce the exploration time. To allocate noise budgets to each kernel, the main idea is to characterize the impact of approximating each kernel on accuracy and cost through simulation and regression to construct the empirical models. These models are then used to obtain the noise budgets. The second research direction is a hybrid algorithm combining Bayesian optimization (BO) and a fast local search to speed up the WLO procedure. An efficient mechanism is proposed to switch between the BO and the local search to obtain good designs in a short time. The last contribution opens a new research direction on resource-constrained WLO. In this study, a Bayesian optimization based algorithm is proposed to maximize the quality of computations constrained by a cost budget.