Merge pull request #2589 from devitocodes/nvc-red

compiler: Avoid nvc++ array reductions and add MPI fallback

Author: FabioLuporini

devito/passes/iet/languages/openmp.py

@@ -5,7 +5,7 @@
 from sympy import And, Ne, Not
 
 from devito.arch import AMDGPUX, NVIDIAX, INTELGPUX, PVC
-from devito.arch.compiler import GNUCompiler
+from devito.arch.compiler import GNUCompiler, NvidiaCompiler
 from devito.ir import (Call, Conditional, DeviceCall, List, Pragma, Prodder,
                        ParallelBlock, PointerCast, While, FindSymbols)
 from devito.passes.iet.definitions import DataManager, DeviceAwareDataManager
@@ -224,7 +224,11 @@ def _support_array_reduction(cls, compiler):
         if isinstance(compiler, GNUCompiler) and \
            compiler.version < Version("6.0"):
             return False
-        return True
+        elif isinstance(compiler, NvidiaCompiler):
+            # NVC++ does not support array reduction and leads to segfault
+            return False
+        else:
+            return True
 
 
 class DeviceOmpizer(PragmaDeviceAwareTransformer):

devito/passes/iet/mpi.py

@@ -264,14 +264,14 @@ def _mark_overlappable(iet):
 
 
 @iet_pass
-def make_halo_exchanges(iet, mpimode=None, **kwargs):
+def make_halo_exchanges(iet, mpimode=None, fallback='basic', **kwargs):
     """
     Lower HaloSpots into halo exchanges for distributed-memory parallelism.
     """
     # To produce unique object names
     generators = {'msg': generator(), 'comm': generator(), 'comp': generator()}
 
-    sync_heb = HaloExchangeBuilder('basic', generators, **kwargs)
+    sync_heb = HaloExchangeBuilder(fallback, generators, **kwargs)
     user_heb = HaloExchangeBuilder(mpimode, generators, **kwargs)
     mapper = {}
     for hs in FindNodes(HaloSpot).visit(iet):