New and Improved MapFusion

[philip-paul-mueller]

This PR introduces a new and improved version of MapFusion.
A summary of the changes can also be found here, it compares the resulting SDFGs generated by the old and new transformation of some unit tests.

Fixed Bugs and removed Limitations

• The subsets (not the .subset member of the Memlet; I mean the concept) of the new intermediate data descriptor were not computed correctly in some cases, especially in presence of offsets. See the test_offset_correction_range_read(), test_offset_correction_scalar_read() and the test_offset_correction_empty() tests.
• Upon the propagation of the subsets, due to the changed intermediate, was not handled properly. Essentially, the transformation only updated .subset and ignored .other_subset. Which is correct in most cases but not always. See the test_fusion_intrinsic_memlet_direction() for more.
• During the check if two maps could be fused the .dynamic property of the Memelts were fully ignored leading to wrong code.
• The read-write conflict checks were refined, before all arrays needed to be accessed the wrong way, i.e. before a fusion was rejected when one map accessed A[i, j] and the other map was accessing B[i + 1, j]. Now this is possible as long as every access is point wise. See the test_fusion_different_global_accesses() test for an example.
• The shape of the reduced intermediate is cleaned, i.e. unnecessary dimensions of size 1, are removed, except they were present in the original shape. To make an example, the intermediate array, T, had shape (10, 1, 20) and inside the map was accessed T[__i, 0, __j], then the old transformation would have created an reduced intermediate of shape (1, 1, 1), new its shape is (1). Note if the intermediate has shape (10, 20) instead and would be accessed as T[__i, __j] then a Scalar would have been created. See also the struct_dataflow flag below.

New Flags

• only_toplevel_maps: If True the transformation will only fuse maps that are located at the top level, i.e. maps inside maps will not be merged.
• only_inner_maps: If True then the transformation will only fuse maps that are inside other maps.
• assume_always_shared: If True` then the transformation will assume that every intermediate is shared, i.e. the referenced data is used somewhere else in the SDFG and has to become an output of the fused maps. This will create dead data flow, but avoids a scan of the full SDFG.
• strict_dataflow: This flag is enabled by default. It has two effects, first it will disable the cleaning of reduced intermediate storage. The second effect is more important as it will preserve a much stricter data flow. Most importantly, if the intermediate array is used downstream (this is not limited to the case that the array is the output of the second map) then the maps will not be fused together. This is mostly to work around some other bugs in DaCe, where other transformations failed to pink up the dependency. Note that the fused map would be correct, the problem are other transformations.

FullMapFusion

This PR also introduced the FullMapFusion pass, which makes use of the FindSingleUseData pass that was introduced in PR#1906.
The FullMapFusion applies MapFusion as long as possible, i.e. fuses all maps that can be fused.
But instead of scanning the SDFG every time an intermediate node has to be classified, i.e. can it be deleted or not, it is done once and then reused which will speed up fusion process as it will remove the need to traverse the full SDFG many times.
This new pass also replaced the direct application of MapFusion in auto_optimizer.

References

Collection of known issues in other transformation:

• RedundantArrayRemoval (auto optimize)
• Bug in RefineNestedAccess and `SDFGState._read_and_write_sets()
• Error in Composite Fusion (auto optimize)