progam.sdfg Does Not Lead to Reproducibility

[philip-paul-mueller]

When DaCe generates the code for an SDFG it also creates the program.sdfg file which is the SDFG that was actually compiled.
This is not the SDFG on which .compile() was called, but one that further transformed (such as CopyToMap or the expansion of library node).
However, for some SDFG it was noticed that this does not lead to the same code (this only refers to the .cu file).
Consider the following:

sdfg1 = make_sdfg()
sdfg1.name = "my_sdfg1"
csdfg1 = sdfg1.compile()

# Now load the SDFG from the folder
sdfg2 = dace.SDFG.from_file(".dacecache/my_sdfg1/program.sdfg")
sdfg2.name = "my_sdfg2"
csdfg2 = sdfg2.compile()

We would now expect that the two folders .dacecache/my_sdfg1/src/cuda/ and .dacecache/my_sdfg2/src/cuda/ contains the same code, beside the changes due to the renaming from my_sdfg1 to my_sdfg2.

However, we currently have an SDFG for which this is not true.
The code, especially the order is significantly different for the two version, see attached files.

However, if we would now do the same with sdfg2, i.e. if we would construct sdfg3 = dace.SDFG.from_file(".dacecache/my_sdfg2/program.sdfg") then we would get a stable behaviour.

sdfg_compile_instability.tar.gz

I am not 100% sure, but I am pretty sure that there is no relation to issue #1946