Parallel Embedded Computing Architectures
Michael Schmidt, Dietmar Fey, Marc Reichenbach · InTech eBooks · 2012
Will-be-set-by-IN-TECHThis situation is different in the case of a given problem which can be decomposed according to geometric principles.That means, we have given a 2D or 3D problem space which is divided in sub regions.In each sub region the same function is carried out.Each sub region is further subdivided in grid points and also on each grid point the same function is applied to.Often this function requires also input from grid points located in the nearest neighbourhood of the grid point.A common parallelization strategy for such problems is to process the grid points of one sub region in a serial manner and to process all sub regions simultaneously, e.g. on different cores.Also this function can be denoted as a task.As mentioned, all these tasks are identical and are applied to different data, whereas the tasks in task parallelism carry out different tasks usually.Furthermore, data parallel tasks can be processed in a complete synchronous way.That means, there are only geometric dependencies between these tasks and no casual time dependencies among the tasks, what is once again contrary to the case of task parallelism.If there are time dependencies then they hold for all tasks.That is why they are synchronous in the sense that all grid points are updated in a time serial loop.