Section: New Results

Parametric Inter-Tile Reuse for Kernel Offloading

Participants : Alain Darte, Alexandre Isoard.

The method described in Section  6.4 is not parametric in terms of the tile size, i.e., the tile size needs to be fixed before compiling the program. Furthermore, the size of the required local memory depends on the tile size and is available only after program analysis. As a result, to select the tile size with respect to the size of the local memory, the program first needs to be compiled (actually analyzed) for all tile sizes. A parametric program analysis would be much more convenient. The situation is even worse to get the runtime performances in terms of the tile size. Indeed, so far, Chuba generates a C code that is generic and cannot be immediately compiled by C2H. A few modifications by hand are still needed, such as inserting the adequate pragmas for C2H, transforming array accesses to linearized addresses with the right base addresses, changing some arrays into non-aliasing pointers so that C2H, whose dependence analyzer and software pipeliner are weak, can generate codes with the right initiation intervals, etc. These changes are minor and systematic and take time when performed by hand. A fully-parametric compilation scheme would be a plus.

The formulation proposed in Section  6.4 is unfortunately quadratic in terms of the tile size, which prevents to parameterize it. Indeed, it relies on parametric linear programming, which works only with a linear use of parameters. As part of the Master internship of Alexandre Isoard, we nevertheless succeeded to design a fully-parametric scheme for inter-tile reuse and buffer size computation. The method is much more involved but is still compatible with approximations. These results have still to be implemented and submitted for publication.