Dynamic Conditional Branch Balancing during the High-Level Synthesis of Control-Intensive Designs

摘要

We present two novel strategies to increase the scope for application of speculative code motions: (1) Adding scheduling steps dynamically during scheduling to conditional branches with fewer scheduling steps. This increases the opportunities to apply code motions such as conditional speculation that duplicate operations into the branches of a conditional block. (2) Determining if an operation can be conditionally speculated into multiple basic blocks either by using existing idle resources or by creating new scheduling steps. These strategies lead to balancing of the number of steps in the conditional branches without increasing the longest path through the conditional block. Algorithms for these strategies have been implemented within the Spark high-level synthesis framework that accepts a behavioral description in ANSI-C as input and produces synthesizable register-transfer level VHDL. Experiments on two moderately complex industrial-strength applications, namely, MPEG-1 and the GIMP image processing tool, demonstrate that conditional speculation is ineffective without using these strategies.