Loop tiling is one of the most important compiler optimization techniques. A multilevel tiled code can improve the running time of the program by maximizing data reuse.PrimeTile and DynTile are tools that are used to generate parametrized multi-level tiledcode.
The tile size parameters need not be specified during compilation. This enables runtime optimization of the program The values assigned to the tile size parameters influence the running time of the program. The choice of the tile size parameters play an important role in the running time of the program.
After generating the multi-level tiled code, it would be ideal to set the tile size parameters to the values that would give the best running time for all the valid permutations of the tile size parameters. In a multi-level tiled code there are several valid permutations for the tile size parameters and hence the search space is huge. An exhaustive search through all the valid permutations defeats the purpose of the search because the time taken for such a search would be much longer than the execution time of the program.
This calls for redefining the problem. After generating the multi-level tiled code, the tile size parameters should be set to the values that give an execution time that is very close to the best running time for all the valid permutations for the tile sizes. To suggest a good search method to find the required tile size parameters, it is necessary to perform an exhaustive study over the effects on the running time of the program by changing the tile size parameters. This study must be done over different benchmarks.
The tests are done on tiled loops generated by PrimeTile and DynTile. The characteristics of the tiled loops generated by PrimeTile are compared with the tiled loops generated by DynTile. The effects of vectorization on these tiled loops are also studied.