Study of parallel processing area extraction and data transfer number reduction for automatic GPU offloading of IoT applications

Study and evaluation of improved automatic GPU offloading method

“…There are two main features for applications offloading. The first one is the loop statements suitable for a GPU are extracted using evolutionary computation [21] of the genetic algorithm (GA) [28], and the variables used in nested loop statements are transferred between the CPU and GPU in the outer-most loop possible [20].…”

“…In addition to the appropriate extraction of loop statements, I previously proposed to transfer as many variables used in the nested loop statements as possible to the upper level loop to reduce the number of transfers between CPU and GPU [20]. When offloading a loop statement to a GPU with a naive method, if CPU-GPU transfer is executed at the lower level of nesting, transfer is done at each lower-level loop, which is not efficient.…”

“…In my previous studies, I attempted to automate the trial and error of parallel processing areas [20,21]. Loop statements suitable for GPU offloading were extracted appropriately by repeating the performance measurements in the verification environment using evolutionary computation, and as many variables in the nested loop statements as possible were transferred in the upper loop.…”

“…I previously proposed environment adaptive software that effectively runs once-written applications by automatically executing code conversion and configurations so that GPUs, FPGAs, and IoT devices can be effectively used on deployment environments [20]. Even if the performance is not so high compared to manual tuning of highly skilled engineers, automation is important, I think.…”

“…As part of environment adaptive software, I also proposed a method to offload loop statements of applications to GPUs [21] and FPGAs [22] automatically. In this paper, I improved upon the GPU offloading method, which I previously proposed using evolutionary computation [20,21], to be applied to more applications with higher performance. I implemented this method and evaluated its effectiveness for several applications.…”