Research On Parallel Optimization Method Of Source Code Level For Coarse Grained Reconfigurable Processor

As reconfigurable processor has the flexibility of general-purpose processor processing various applications and efficient processing capability like ASIC,it has become a hot research direction in the field of science and technology.However,due to the lack of corresponding compiler architecture and automatic parallelizing method,the development of compiler for reconfigurable processor has been a major difficulty in the development of reconfigurable computing.Aiming at the hardware architecture of coarse-grained reconfigurable processor,we design an extended language GR-C based on standard C and a compiler called GRVM.In order to fully exploit the parallelism,we designed an automatic parallelization module through transforming source code to target code in preprocessing end of the compiler,and it can convert the program with single configuration package(single task)into that with the multiple configuration packages(multiple tasks)according to the dependency of the source code.For the multi-tasks program,we design a multi-tasks parallel scheduling strategy.If the number of configuration package is greater than 4,a parallel strategy of long pipeline is proposed to fully explore the parallelism among multiple reconfigurable processing arrays.In order to provide the programmer good programming interface,we design a pipelined parallel programming mode.For the multi-tasks model based on data partition,we design the task-level parallel programming mode,which greatly lowers the difficulty of programming.In view of non-perfect loop and uncertain-length loop,we adopt various loop transformation methods such as loop unrolling,loop separation,loop fusion and code redivision according to different situations,to enable the compiler to produce configuration information for different loops.We have selected a number of computing and data intensive algorithms to perform the verification on the ESL platform.Through the contrast between the reconfigurable processor and general purpose processor,between GRVM with automatic parallelization module and without automatic parallelization module,and between automatic compilation and handwritten configuration information,we verified the accuracy and the superiority in performance of the parallel optimization method of source code level.