arch: support rocm for gpu info

[mloubout]

currently removes the Xcompile flag needed for host side as a duplicate flag

Fixes #2260

[codecov]

Codecov Report

Merging #2261 (d0002b2) into master (e707027) will decrease coverage by 0.07%.
The diff coverage is 19.44%.

@@            Coverage Diff             @@
##           master    #2261      +/-   ##
==========================================
- Coverage   86.95%   86.88%   -0.07%     
==========================================
  Files         229      229              
  Lines       41970    42034      +64     
  Branches     7752     7760       +8     
==========================================
+ Hits        36495    36522      +27     
- Misses       4838     4871      +33     
- Partials      637      641       +4     

Files	Coverage Δ
devito/mpi/distributed.py	92.92% <100.00%> (+0.02%)	⬆️
devito/arch/compiler.py	44.42% <85.71%> (+4.34%)	⬆️
tests/test_gpu_common.py	1.40% <0.00%> (-0.02%)	⬇️
devito/arch/archinfo.py	40.15% <2.12%> (-3.72%)	⬇️

[georgebisbas]

What was the mpi problem?

[deckerla]

What was the mpi problem?

This is the related issue:

#2262

[FabioLuporini]

After playing with the test a bit, I'm not sure this is the right fix.

The test does need a halo exchange (u needs to be up-to-date before getting interpolated), but what this patch is essentially: "well, if the halo_scheme is empty, drop it and don't generate anything"

However, empty halo schemes should have been dropped earlier on, to avoid clusters-level pollution. In particular, here: https://github.com/devitocodes/devito/blob/master/devito/ir/clusters/algorithms.py#L395

So, it turns out that is_void gives False here because the fmapper is non-empty while the distributed-aindices are empty. And the latter is caused by the fact that no dimensions are used to index into u -- just plain symbols, AFAICT

So we might have to:

• fix this somewhere else
• make the test more robust, e.g. by placing the receivers at the MPI borders

About the fix: it might need to go somewhere here https://github.com/devitocodes/devito/blob/master/devito/mpi/halo_scheme.py#L470

by the looks of it, if f.grid.is_distributed(d) then:

• if there's an aindex, we good (most of the cases)
• if i[d] is a number, we should throw a user exception, since it's illegal MPI code (can't use pure numbers to index into distributed dimensions
• if i[d] isn't a number (e.g., like in our test, an expression of pure symbols), then we should use a dummy dimension placeholder, capturing the fact that such dimension does require a halo exchange

[mloubout]

It doesn't drop it it just keeps it where it is in the list of clusters. It's dropped somewhere else probably in the mpi pass but that's a different issue.

[FabioLuporini]

whatever happens here, and regardless of how we end up doing it, we should always do it, irrespective of whether MPI is enabled or not. This will confer robustness to our compiler and, in particular, it will avoid "bugs to only pop up with MPI enabled"