Generic Loop Parallelization for Reconfigurable Architectures

Reconfigurable Computing (RC) is one of the most intensively studied research areas nowadays due to its potential to dramatically increase application performance. RC combines a general purpose processor (GPP) and a Field Programmable Gate Array (FPGA), having the advantages of both hardware performance and software flexibility. Modern real-life applications (such as audio, video, image processing, etc) spend most of the execution time in loops, which represent or include the application kernels. These loops are an important source of performance improvement. In our work, we target loops that contain in their bodies code for the GPP (software functions) and also for the FPGA (hardware functions). We assume there are data dependencies between consecutive tasks in the loop body, but not between different loop iterations. Assuming the Molen machine organization as our framework, we focus on applying existing loop optimizations to such loops, with the purpose of parallelizing applications such that multiple kernel instances run in parallel on the reconfigurable hardware, while concurrently executing code on the GPP. In this paper, we focus on loop transformations that are suitable for loops containing an arbitrary number of software and hardware functions. The extended shifting consists of relocating the functions placed in the beginning and in the end of one loop iteration, in order to eliminate the data dependencies and allow certain software and hardware functions to be executed in parallel. The loop distribution consists of splitting the loop into small loops (e.g., with only one kernel) allowing in some cases a larger degree of parallelism when applying the loop unrolling and shifting techniques. We estimate the performance achieved by applying the extended shifting technique in conjunction with loop unrolling and compare it to the performance achieved when applying the loop unrolling and shifting techniques to smaller loops obtained by distributing the original loop. For the experimental results we used randomly generated tests, for loops containing a variable number of kernels (between 2 and 8 kernels).