文档介绍：Path Grouping and Data Dependence Relaxation for Software
Pipelining
RONG Hong-bo TANG Zhi-zhong
Department puter Science and Technology, Tsinghua University, Beijing 100084
Email: ******@.
tzz-******@tsinghua.
well the loop has been parallelized. Finding the
Abstract smallest possible II is the aim of software pipelining.
II is affected by two factors: (1) Resource
Software pipelining of loops with conditional constraints. When II is too small, concurrent
branches remains a challenge. There are two pete for the limited resources. (2)
antagonistic problems, transfer time minimization and Dependence constraints. When II is too small, control
the worst-case constraints. The paper presents a novel and data dependencies between operations may be
framework to address them without conflict. The key violated.
ideas are: (1) Path grouping, which splits paths into For a loop without conditional statements, all
distinct groups based on their execution and transfer iterations have the same operations. Considering the
probabilities, so as to minimize the transfer time. (2) above constraints is enough for a good schedule,
Data dependence relaxation, which avoids the worst- which may be formed by a fixed II [1-3].
case constraints by respecting data dependencies only For a loop with conditional statements (with
when they have instances. Analysis sugg