Impacts of code parallelization and GPU-implementation to acceleration of multi-image phase retrieval

Abstract

Multi-image phase retrieval is an iterative technique in coherent metrology which requires performing computations on large data sets. While previous studies used a parallelization scheme and an alternative hardware for improving execution times, there was no clear breakdown of how much of the gain was solely due to code parallelization versus the hardware. In this study, we investigated the isolated contributions of the hardware, which is a graphics processing unit or GPU, and the code parallelization to the acceleration of phase retrieval. A serialized implementation of the phase retrieval technique was adapted. The execution times between the serial and parallel schemes performed in different hardwares were compared. We found that most of the acceleration was due to the GPU, resulting in shortened runtime from twenty minutes to less than one minute. The improvement due code parallelization was marginal, resulting in only 2.46 s lesser computational time. In future works, an improved code parallelization scheme could be developed to achieve much faster execution times together with a GPU-based implementation.