University of Patras
Τεχνικές μεταγλωττιστών και αρχιτεκτονικές επεξεργαστών για στατικές και δυναμικές εφαρμογές
Abstract
dc:descriptionModern applications have indence needs in processing power in order to be executed in short time. For satisfying the time limits, there have to be generated new techniques for optimizing the designs. The object of the present thesis is about developing new compiler techniques and hardware architectures which aim to transfer data faster, from the memory hierarchy to the CPU. a) Methdology for accelerating the execution of matrix multiplications A new methodology using the standard MMM algorithm is presented, achieving improved performance by focusing on data locality (both temporal and spatial). This methodology finds the scheduling which conforms with the optimum memory management. The scheduling used for the tile level is different from the element level’s one, having better data locality, suited to the sizes of memory hierarchy. Its exploration time is short, because it searches only for the number of the level of tiling used for finding the best tile size for each cache level. Compared with the best existing related work, which we implemented, better performance up to 55% β)Methodology for increasing performance on Fast Fourier Transform (FFT) A new methodology is presented based on minimizing the memory accesses for FFT. It exploits, the production and comsumption of the FFT batterfly results and the reuse of data. The optimum scheduling solution is found taking into account the number of registers and the cache memory size. The compile time and source code size are short comparing to SOA library. The methodology performance gains are up to 63% comparing to FFTW library. γ)Ανάπτυξη Αρχιτεκτονικών για Διαχείριση Μνήμης A decoupled processors architecture with a memory hierarchy is presented consisting only of scratch–pad memories, and a main memory. This architecture exploits both the benefits of scratch-pad memories and the parallelism between address computation and application data processing. The architecture is compared in performance with the MIPS architecture with cache and with scratch-pad memory hierarchies and with the existing decoupled architectures showing its higher normalized performance. Experimental results show that the performance is increased up to 3.7 times. Continuing, more research is done on Scratch-pad memories. We present an architecture that is able to provide information about the exact data contents of scratch-pad during execution and can also do all the necessary operations for placing the new data blocks in scratch-pad. Thereby, the temporal locality which occurs randomly and can not be identified by the compiler is exploited. It is compared with the MIPS architecture with cache and with scratch-pad memories showing its higher normalized performance. Experimental results show that the performance is increased up to 5 times compared to cache architectures and 2,5 times compared to existing scratch-pad architectures
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2009
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Alachiotis, Nikolaos
- Αλαχιώτης, Νικόλαος
Subjects
dc:subject × 19- Τεχνικές μεταγλωττιστών
- Αρχιτεκτονικές υπολογιστών
- Ενσωματωμένα συστήματα
- Στατικές και δυναμικές εφαρμογές
- Χαμηλή κατανάλωση ισχύος
- Υψηλή απόδοση
- Διαχείριση μνήμης
- Compiler optimizations
- Computer architecture
- Embedded systems
- Low power
- High performance
- Memory management
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/27970