Merge pull request #2555 from devitocodes/tweak-idg-construction-2

compiler: Enhance IR for GPU codegen

Author: FabioLuporini

devito/ir/support/basic.py

@@ -1266,6 +1266,7 @@ def translated(self, other, dims=None):
         dims = set(as_tuple(dims))
 
         # Check bases and offsets
+        distances = {}
         for i in ['Tbases', 'Toffsets']:
             Ti0 = getattr(self, i)
             Ti1 = getattr(other, i)
@@ -1288,13 +1289,16 @@ def translated(self, other, dims=None):
 
                 distance = set(o0 - o1)
                 if len(distance) != 1:
-                    return False
+                    return {}
+                v = distance.pop()
 
                 if not d._defines & dims:
-                    if distance.pop() != 0:
-                        return False
+                    if v != 0:
+                        return {}
+
+                distances[d] = v
 
-        return True
+        return distances
 
     @cached_property
     def iinstances(self):

devito/types/array.py

@@ -90,16 +90,18 @@ class Array(ArrayBasic):
         to 'local'. Used to override `_mem_local` and `_mem_mapped`.
     scope : str, optional
         The scope in the given memory space. Allowed values: 'heap', 'stack',
-        'static', 'constant', 'shared', 'shared-remote'. 'static' refers to a
-        static array in a C/C++ sense. 'constant' and 'shared' mean that the
-        Array represents an object allocated in so called constant and shared
-        memory, respectively, which are typical of device architectures. If
-        'shared' is specified but the underlying architecture doesn't have
-        something akin to shared memory, the behaviour is unspecified. If
-        'constant' is specified but the underlying architecture doesn't have
-        something akin to constant memory, the Array falls back to a global,
-        const, static array in a C/C++ sense.  Note that not all scopes make
-        sense for a given space.
+        'static', 'constant', 'shared', 'shared-remote', 'registers'.
+        'static' refers to a static array in a C/C++ sense. 'constant' and
+        'shared' mean that the Array represents an object allocated in so
+        called constant and shared memory, respectively, which are typical of
+        device architectures. If 'shared' is specified but the underlying
+        architecture doesn't have something akin to shared memory, the
+        behaviour is unspecified. If 'constant' is specified but the underlying
+        architecture doesn't have something akin to constant memory, the Array
+        falls back to a global, const, static array in a C/C++ sense.
+        'registers' is used to indicate that the Array has a small static size
+        and, as such, it could be allocated in registers. Defaults to 'heap'.
+        Note that not all scopes make sense for a given space.
     grid : Grid, optional
         Only necessary for distributed-memory parallelism; a Grid contains
         information about the distributed Dimensions, hence it is necessary
@@ -133,7 +135,7 @@ def __init_finalize__(self, *args, **kwargs):
 
         self._scope = kwargs.get('scope', 'heap')
         assert self._scope in ['heap', 'stack', 'static', 'constant', 'shared',
-                               'shared-remote']
+                               'shared-remote', 'registers']
 
         self._initvalue = kwargs.get('initvalue')
         assert self._initvalue is None or self._scope != 'heap'
@@ -166,7 +168,7 @@ def _C_ctype(self):
 
     @property
     def _mem_stack(self):
-        return self._scope in ('stack', 'shared')
+        return self._scope in ('stack', 'shared', 'registers')
 
     @property
     def _mem_heap(self):
@@ -180,6 +182,10 @@ def _mem_shared(self):
     def _mem_shared_remote(self):
         return self._scope == 'shared-remote'
 
+    @property
+    def _mem_registers(self):
+        return self._scope == 'registers'
+
     @property
     def _mem_constant(self):
         return self._scope == 'constant'