Bringing the loop nests to titly-nested loops
The DVM system allows you to get a significant win from parallelizing tightly nested loops (perfectly nested loops). In such loop nests, the body of the outer loop is the inner loop, no statements must be placed between these two loops. Multidimensional loop nests can be mapped to a multidimensional grid of processors according to the data distribution. This possibility is in good agreement with the structure of the solved physical problems, which use grid methods, and allows to obtain the maximum parallelism in the calculations for each dimension of the grid. Bringing the nests of the loops to the tightly-nested ones can be implemented in the SAPFOR system.
The paper is not published yet.