Escalonamento de instruções em arquiteturas VLIW particionadas explorando Bypassing de operandos

AUTOR(ES)
DATA DE PUBLICAÇÃO

2001

RESUMO

The untiring search for faster machines, alIied to the great technological advances in the field of integrated circuits conception, brought out the Very Long Instruction Word architectures from an amorphous status to reality. Although they have appeared recently as real chips [1], the VLIW machines were idealized some decades ago [13, 16, 22, 23]. The microprocessors that define this processing model no longer obey classical rules of execution: instructions coming from one of the possible control flows resulted of a branch instruction are executed even before the finish of the evaluation condition. This evaluation condition will determine if the control transfer should occur or noto Also, these architectures execute simultaneously many instructions, of different kinds, issued from the same programo Moreover, these processors compute programs that were compiled through a revolutionary way: alI the program is analized to search for paralelizable operations. As an attempt to contribute to this research field, this work aim the development of a methodology to detect and exploit the paralelism "hided" in sequential-written programs. The results generated by this search are analized and quantified in order to find a targetarchitecture for a specific application. This work is inserted in the context of an area calIed Embedded Systems. This research field worry about the maximum optimization of an application class or even only one key-application of a embedded system. The architecture model considered in this work is denoted as "Partitioned VLIW Architecture". This model is slightly different of the ideal VLIW architecture model. In the ideal model, there must be only one centralized register file, in order to guarantee the maximum Instruction Levei ParalIelism (ILP). AlI the functional units share the same register file. On the other hand, the architecture model being considered here presents many distributed register files, which have an special bus to communicate data among them. With this architecture model in mind, the work developed in this thesis investigates some of the problems related to mapping one specific application to an embedded VLIW architecture. Roughly speaking, this work tries to answer the following question: "What is the ideal VLIW architecture for a given application 1" or "How many register files and how many functional units the processor for that application should have 1"

ASSUNTO(S)

compiladores (computadores) arquitetura de computador

ACESSO AO ARTIGO

Documentos Relacionados