Array relocation approach for radial scanning algorithms on multi-GPU systems: total viewshed problem as a case study
share
In geographic information systems, Digital Elevation Models (DEMs) are commonly processed using radial scanning based algorithms. These algorithms are particularly popular when calculating parameters whose magnitudes decrease with the distance squared such as those related to radio signals, sound waves, and human eyesight. However, radial scanning algorithms imply a large number of accesses to 2D arrays, which despite being regular, results in poor data locality. This paper proposes a new methodology, termed sDEM, which substantially improves the locality of memory accesses and largely increases the inherent parallelism involved in the computation of radial scanning algorithms. In particular, sDEM applies a data restructuring technique prior to accessing the memory and performing the computation. In order to demonstrate the high efficiency of sDEM, we use the problem of total viewshed computation as a case study. Sequential, parallel, single-GPU and multi-GPU implementations are analyzed and compared with the state-of-the-art total viewshed computation algorithm. Experiments show that sDEM achieves an acceleration rate of up to 827.3 times for the best multi-GPU execution approach with respect to the best multi-core implementation.
READ FULL TEXT
