Identification and Elimination of the Overhead of Accelerate with a Super-resolution Application

This paper supplements the evaluation of libraries for parallel computing in Haskell by Matsumoto and Matsuzaki. The previous work implemented an image processing program by using two different libraries and compared expressiveness and parallel execution performance. In particular, it found that a parallel execution of the Accelerate program on a GPU is slower than the execution of the handwritten CUDA-C program by a factor of more than 70. We analyzed the execution of the Accelerate program and identified that the major sources of the overhead are in the process of the syntax tree before execution of a GPU kernel. Since the program realizes repeated execution by constructing a large syntax tree whose size is proportional to the number of repetitions, we rewrote this part by using a while-construct, which significantly reduces the overhead as it runs half as fast as the CUDA-C program. We also discuss language design alternatives for avoiding this type of overhead.------------------------------This is a preprint of an article intended for publication Journal ofInformation Processing(JIP). This preprint should not be cited. Thisarticle should be cited as: Journal of Information Processing Vol.26(2018) (online)------------------------------

This paper supplements the evaluation of libraries for parallel computing in Haskell by Matsumoto and Matsuzaki. The previous work implemented an image processing program by using two different libraries and compared expressiveness and parallel execution performance. In particular, it found that a parallel execution of the Accelerate program on a GPU is slower than the execution of the handwritten CUDA-C program by a factor of more than 70. We analyzed the execution of the Accelerate program and identified that the major sources of the overhead are in the process of the syntax tree before execution of a GPU kernel. Since the program realizes repeated execution by constructing a large syntax tree whose size is proportional to the number of repetitions, we rewrote this part by using a while-construct, which significantly reduces the overhead as it runs half as fast as the CUDA-C program. We also discuss language design alternatives for avoiding this type of overhead.------------------------------This is a preprint of an article intended for publication Journal ofInformation Processing(JIP). This preprint should not be cited. Thisarticle should be cited as: Journal of Information Processing Vol.26(2018) (online)------------------------------