Abstract:

In order to automatically map algorithms onto reconfigurable multimedia processor and improve the parallel efficiency of algorithms, a task compiler frontend is designed. The kernel loop unrolling is introduced to improve the degree of parallelism; and the scalar replacement technique based on data dependence analysis is used to optimize the cost of data transmission. Experiments show that the task compiler frontend improves the degree of parallelism effectively, its performance can be, compared with the frontend of Garp C compiler, sharply increased up to 2~4 times in the whole system.

Key words: reconfigurable computing, task compiler, loop unrolling, scalar replacement