Merge branch 'master' into mpi4py_dev

dace/codegen/compiled_sdfg.py

@@ -452,9 +452,10 @@ def _construct_args(self, kwargs) -> Tuple[Tuple[Any], Tuple[Any]]:
                 # GPU scalars are pointers, so this is fine
                 if atype.storage != dtypes.StorageType.GPU_Global:
                     raise TypeError('Passing an array to a scalar (type %s) in argument "%s"' % (atype.dtype.ctype, a))
-            elif not isinstance(atype, dt.Array) and not isinstance(atype.dtype, dtypes.callback) and not isinstance(
-                    arg,
-                (atype.dtype.type, sp.Basic)) and not (isinstance(arg, symbolic.symbol) and arg.dtype == atype.dtype):
+            elif (not isinstance(atype, (dt.Array, dt.Structure)) and
+                  not isinstance(atype.dtype, dtypes.callback) and
+                  not isinstance(arg, (atype.dtype.type, sp.Basic)) and
+                  not (isinstance(arg, symbolic.symbol) and arg.dtype == atype.dtype)):
                 if isinstance(arg, int) and atype.dtype.type == np.int64:
                     pass
                 elif isinstance(arg, float) and atype.dtype.type == np.float64:
@@ -472,7 +473,7 @@ def _construct_args(self, kwargs) -> Tuple[Tuple[Any], Tuple[Any]]:
                 else:
                     warnings.warn(f'Casting scalar argument "{a}" from {type(arg).__name__} to {atype.dtype.type}')
                     arglist[i] = atype.dtype.type(arg)
-            elif (isinstance(atype, dt.Array) and isinstance(arg, np.ndarray)
+            elif (isinstance(atype, dt.Array) and isinstance(arg, np.ndarray) and not isinstance(atype, dt.StructArray)
                   and atype.dtype.as_numpy_dtype() != arg.dtype):
                 # Make exception for vector types
                 if (isinstance(atype.dtype, dtypes.vector) and atype.dtype.vtype.as_numpy_dtype() == arg.dtype):
@@ -521,7 +522,7 @@ def _construct_args(self, kwargs) -> Tuple[Tuple[Any], Tuple[Any]]:
         # Construct init args, which only consist of the symbols
         symbols = self._free_symbols
         initargs = tuple(
-            actype(arg) if (not isinstance(arg, ctypes._SimpleCData)) else arg
+            actype(arg) if not isinstance(arg, ctypes._SimpleCData) else arg
             for arg, actype, atype, aname in callparams if aname in symbols)
 
         # Replace arrays with their base host/device pointers
@@ -531,7 +532,8 @@ def _construct_args(self, kwargs) -> Tuple[Tuple[Any], Tuple[Any]]:
 
         try:
             newargs = tuple(
-                actype(arg) if (not isinstance(arg, ctypes._SimpleCData)) else arg for arg, actype, atype in newargs)
+                actype(arg) if not isinstance(arg, (ctypes._SimpleCData)) else arg
+                for arg, actype, atype in newargs)
         except TypeError:
             # Pinpoint bad argument
             for i, (arg, actype, _) in enumerate(newargs):

dace/codegen/control_flow.py

@@ -82,6 +82,9 @@ class ControlFlow:
     # a string with its generated code.
     dispatch_state: Callable[[SDFGState], str]
 
+    # The parent control flow block of this one, used to avoid generating extraneous ``goto``s
+    parent: Optional['ControlFlow']
+
     @property
     def first_state(self) -> SDFGState:
         """ 
@@ -222,11 +225,18 @@ def as_cpp(self, codegen, symbols) -> str:
                 out_edges = sdfg.out_edges(elem.state)
                 for j, e in enumerate(out_edges):
                     if e not in self.gotos_to_ignore:
-                        # If this is the last generated edge and it leads
-                        # to the next state, skip emitting goto
+                        # Skip gotos to immediate successors
                         successor = None
-                        if (j == (len(out_edges) - 1) and (i + 1) < len(self.elements)):
-                            successor = self.elements[i + 1].first_state
+                        # If this is the last generated edge
+                        if j == (len(out_edges) - 1):
+                            if (i + 1) < len(self.elements):
+                                # If last edge leads to next state in block
+                                successor = self.elements[i + 1].first_state
+                            elif i == len(self.elements) - 1:
+                                # If last edge leads to first state in next block
+                                next_block = _find_next_block(self) 
+                                if next_block is not None:
+                                    successor = next_block.first_state
 
                         expr += elem.generate_transition(sdfg, e, successor)
                     else:
@@ -478,13 +488,14 @@ def children(self) -> List[ControlFlow]:
 
 def _loop_from_structure(sdfg: SDFG, guard: SDFGState, enter_edge: Edge[InterstateEdge],
                          leave_edge: Edge[InterstateEdge], back_edges: List[Edge[InterstateEdge]],
-                         dispatch_state: Callable[[SDFGState], str]) -> Union[ForScope, WhileScope]:
+                         dispatch_state: Callable[[SDFGState],
+                                                  str], parent_block: GeneralBlock) -> Union[ForScope, WhileScope]:
     """ 
     Helper method that constructs the correct structured loop construct from a
     set of states. Can construct for or while loops.
     """
 
-    body = GeneralBlock(dispatch_state, [], [], [], [], [], True)
+    body = GeneralBlock(dispatch_state, parent_block, [], [], [], [], [], True)
 
     guard_inedges = sdfg.in_edges(guard)
     increment_edges = [e for e in guard_inedges if e in back_edges]
@@ -535,10 +546,10 @@ def _loop_from_structure(sdfg: SDFG, guard: SDFGState, enter_edge: Edge[Intersta
             # Also ignore assignments in increment edge (handled in for stmt)
             body.assignments_to_ignore.append(increment_edge)
 
-            return ForScope(dispatch_state, itvar, guard, init, condition, update, body, init_edges)
+            return ForScope(dispatch_state, parent_block, itvar, guard, init, condition, update, body, init_edges)
 
     # Otherwise, it is a while loop
-    return WhileScope(dispatch_state, guard, condition, body)
+    return WhileScope(dispatch_state, parent_block, guard, condition, body)
 
 
 def _cases_from_branches(
@@ -617,6 +628,31 @@ def _child_of(node: SDFGState, parent: SDFGState, ptree: Dict[SDFGState, SDFGSta
     return False
 
 
+def _find_next_block(block: ControlFlow) -> Optional[ControlFlow]:
+    """
+    Returns the immediate successor control flow block.
+    """
+    # Find block in parent
+    parent = block.parent
+    if parent is None:
+        return None
+    ind = next(i for i, b in enumerate(parent.children) if b is block)
+    if ind == len(parent.children) - 1 or isinstance(parent, (IfScope, IfElseChain, SwitchCaseScope)):
+        # If last block, or other children are not reachable from current node (branches),
+        # recursively continue upwards
+        return _find_next_block(parent)
+    return parent.children[ind + 1]
+
+
+def _reset_block_parents(block: ControlFlow):
+    """
+    Fixes block parents after processing.
+    """
+    for child in block.children:
+        child.parent = block
+        _reset_block_parents(child)
+
+
 def _structured_control_flow_traversal(sdfg: SDFG,
                                        start: SDFGState,
                                        ptree: Dict[SDFGState, SDFGState],
@@ -645,7 +681,7 @@ def _structured_control_flow_traversal(sdfg: SDFG,
     """
 
     def make_empty_block():
-        return GeneralBlock(dispatch_state, [], [], [], [], [], True)
+        return GeneralBlock(dispatch_state, parent_block, [], [], [], [], [], True)
 
     # Traverse states in custom order
     visited = set() if visited is None else visited
@@ -657,7 +693,7 @@ def make_empty_block():
         if node in visited or node is stop:
             continue
         visited.add(node)
-        stateblock = SingleState(dispatch_state, node)
+        stateblock = SingleState(dispatch_state, parent_block, node)
 
         oe = sdfg.out_edges(node)
         if len(oe) == 0:  # End state
@@ -708,23 +744,25 @@ def make_empty_block():
             if (len(oe) == 2 and oe[0].data.condition_sympy() == sp.Not(oe[1].data.condition_sympy())):
                 # If without else
                 if oe[0].dst is mergestate:
-                    branch_block = IfScope(dispatch_state, sdfg, node, oe[1].data.condition, cblocks[oe[1]])
+                    branch_block = IfScope(dispatch_state, parent_block, sdfg, node, oe[1].data.condition,
+                                           cblocks[oe[1]])
                 elif oe[1].dst is mergestate:
-                    branch_block = IfScope(dispatch_state, sdfg, node, oe[0].data.condition, cblocks[oe[0]])
+                    branch_block = IfScope(dispatch_state, parent_block, sdfg, node, oe[0].data.condition,
+                                           cblocks[oe[0]])
                 else:
-                    branch_block = IfScope(dispatch_state, sdfg, node, oe[0].data.condition, cblocks[oe[0]],
-                                           cblocks[oe[1]])
+                    branch_block = IfScope(dispatch_state, parent_block, sdfg, node, oe[0].data.condition,
+                                           cblocks[oe[0]], cblocks[oe[1]])
             else:
                 # If there are 2 or more edges (one is not the negation of the
                 # other):
                 switch = _cases_from_branches(oe, cblocks)
                 if switch:
                     # If all edges are of form "x == y" for a single x and
                     # integer y, it is a switch/case
-                    branch_block = SwitchCaseScope(dispatch_state, sdfg, node, switch[0], switch[1])
+                    branch_block = SwitchCaseScope(dispatch_state, parent_block, sdfg, node, switch[0], switch[1])
                 else:
                     # Otherwise, create if/else if/.../else goto exit chain
-                    branch_block = IfElseChain(dispatch_state, sdfg, node,
+                    branch_block = IfElseChain(dispatch_state, parent_block, sdfg, node,
                                                [(e.data.condition, cblocks[e] if e in cblocks else make_empty_block())
                                                 for e in oe])
             # End of branch classification
@@ -739,11 +777,11 @@ def make_empty_block():
             loop_exit = None
             scope = None
             if ptree[oe[0].dst] == node and ptree[oe[1].dst] != node:
-                scope = _loop_from_structure(sdfg, node, oe[0], oe[1], back_edges, dispatch_state)
+                scope = _loop_from_structure(sdfg, node, oe[0], oe[1], back_edges, dispatch_state, parent_block)
                 body_start = oe[0].dst
                 loop_exit = oe[1].dst
             elif ptree[oe[1].dst] == node and ptree[oe[0].dst] != node:
-                scope = _loop_from_structure(sdfg, node, oe[1], oe[0], back_edges, dispatch_state)
+                scope = _loop_from_structure(sdfg, node, oe[1], oe[0], back_edges, dispatch_state, parent_block)
                 body_start = oe[1].dst
                 loop_exit = oe[0].dst
 
@@ -836,7 +874,8 @@ def structured_control_flow_tree(sdfg: SDFG, dispatch_state: Callable[[SDFGState
         if len(common_frontier) == 1:
             branch_merges[state] = next(iter(common_frontier))
 
-    root_block = GeneralBlock(dispatch_state, [], [], [], [], [], True)
+    root_block = GeneralBlock(dispatch_state, None, [], [], [], [], [], True)
     _structured_control_flow_traversal(sdfg, sdfg.start_state, ptree, branch_merges, back_edges, dispatch_state,
                                        root_block)
+    _reset_block_parents(root_block)
     return root_block

dace/codegen/dispatcher.py

@@ -505,11 +505,11 @@ def get_copy_dispatcher(self, src_node, dst_node, edge, sdfg, state):
             dst_is_data = True
 
         # Skip copies to/from views where edge matches
-        if src_is_data and isinstance(src_node.desc(sdfg), dt.View):
+        if src_is_data and isinstance(src_node.desc(sdfg), (dt.StructureView, dt.View)):
             e = sdutil.get_view_edge(state, src_node)
             if e is edge:
                 return None
-        if dst_is_data and isinstance(dst_node.desc(sdfg), dt.View):
+        if dst_is_data and isinstance(dst_node.desc(sdfg), (dt.StructureView, dt.View)):
             e = sdutil.get_view_edge(state, dst_node)
             if e is edge:
                 return None

dace/codegen/targets/cpp.py

@@ -370,6 +370,10 @@ def make_const(expr: str) -> str:
     # Register defined variable
     dispatcher.defined_vars.add(pointer_name, defined_type, typedef, allow_shadowing=True)
 
+    # NOTE: `expr` may only be a name or a sequence of names and dots. The latter indicates nested data and structures.
+    # NOTE: Since structures are implemented as pointers, we replace dots with arrows.
+    expr = expr.replace('.', '->')
+
     return (typedef + ref, pointer_name, expr)
 
 

dace/codegen/targets/cpu.py

@@ -55,10 +55,29 @@ def __init__(self, frame_codegen, sdfg):
         # Keep track of generated NestedSDG, and the name of the assigned function
         self._generated_nested_sdfg = dict()
 
-        # Keeps track of generated connectors, so we know how to access them in
-        # nested scopes
+        # NOTE: Multi-nesting with StructArrays must be further investigated.
+        def _visit_structure(struct: data.Structure, args: dict, prefix: str = ''):
+            for k, v in struct.members.items():
+                if isinstance(v, data.Structure):
+                    _visit_structure(v, args, f'{prefix}.{k}')
+                elif isinstance(v, data.StructArray):
+                    _visit_structure(v.stype, args, f'{prefix}.{k}')
+                elif isinstance(v, data.Data):
+                    args[f'{prefix}.{k}'] = v
+
+        # Keeps track of generated connectors, so we know how to access them in nested scopes
+        arglist = dict(self._frame.arglist)
         for name, arg_type in self._frame.arglist.items():
-            if isinstance(arg_type, data.Scalar):
+            if isinstance(arg_type, data.Structure):
+                desc = sdfg.arrays[name]
+                _visit_structure(arg_type, arglist, name)
+            elif isinstance(arg_type, data.StructArray):
+                desc = sdfg.arrays[name]
+                desc = desc.stype
+                _visit_structure(desc, arglist, name)
+
+        for name, arg_type in arglist.items():
+            if isinstance(arg_type, (data.Scalar, data.Structure)):
                 # GPU global memory is only accessed via pointers
                 # TODO(later): Fix workaround somehow
                 if arg_type.storage is dtypes.StorageType.GPU_Global:
@@ -195,9 +214,21 @@ def allocate_view(self, sdfg: SDFG, dfg: SDFGState, state_id: int, node: nodes.A
                                                         ancestor=0,
                                                         is_write=is_write)
         if not declared:
-            declaration_stream.write(f'{atype} {aname};', sdfg, state_id, node)
             ctypedef = dtypes.pointer(nodedesc.dtype).ctype
             self._dispatcher.declared_arrays.add(aname, DefinedType.Pointer, ctypedef)
+            if isinstance(nodedesc, data.StructureView):
+                for k, v in nodedesc.members.items():
+                    if isinstance(v, data.Data):
+                        ctypedef = dtypes.pointer(v.dtype).ctype if isinstance(v, data.Array) else v.dtype.ctype
+                        defined_type = DefinedType.Scalar if isinstance(v, data.Scalar) else DefinedType.Pointer
+                        self._dispatcher.declared_arrays.add(f"{name}.{k}", defined_type, ctypedef)
+                        self._dispatcher.defined_vars.add(f"{name}.{k}", defined_type, ctypedef)
+                # TODO: Find a better way to do this (the issue is with pointers of pointers)
+                if atype.endswith('*'):
+                    atype = atype[:-1]
+                if value.startswith('&'):
+                    value = value[1:]
+            declaration_stream.write(f'{atype} {aname};', sdfg, state_id, node)
         allocation_stream.write(f'{aname} = {value};', sdfg, state_id, node)
 
     def allocate_reference(self, sdfg: SDFG, dfg: SDFGState, state_id: int, node: nodes.AccessNode,
@@ -268,16 +299,19 @@ def allocate_array(self, sdfg, dfg, state_id, node, nodedesc, function_stream, d
         name = node.data
         alloc_name = cpp.ptr(name, nodedesc, sdfg, self._frame)
         name = alloc_name
+        # NOTE: `expr` may only be a name or a sequence of names and dots. The latter indicates nested data and
+        # NOTE: structures. Since structures are implemented as pointers, we replace dots with arrows.
+        alloc_name = alloc_name.replace('.', '->')
 
         if nodedesc.transient is False:
             return
 
         # Check if array is already allocated
-        if self._dispatcher.defined_vars.has(alloc_name):
+        if self._dispatcher.defined_vars.has(name):
             return
 
         # Check if array is already declared
-        declared = self._dispatcher.declared_arrays.has(alloc_name)
+        declared = self._dispatcher.declared_arrays.has(name)
 
         define_var = self._dispatcher.defined_vars.add
         if nodedesc.lifetime in (dtypes.AllocationLifetime.Persistent, dtypes.AllocationLifetime.External):
@@ -290,7 +324,18 @@ def allocate_array(self, sdfg, dfg, state_id, node, nodedesc, function_stream, d
         if not isinstance(nodedesc.dtype, dtypes.opaque):
             arrsize_bytes = arrsize * nodedesc.dtype.bytes
 
-        if isinstance(nodedesc, data.View):
+        if isinstance(nodedesc, data.Structure) and not isinstance(nodedesc, data.StructureView):
+            declaration_stream.write(f"{nodedesc.ctype} {name} = new {nodedesc.dtype.base_type};\n")
+            define_var(name, DefinedType.Pointer, nodedesc.ctype)
+            for k, v in nodedesc.members.items():
+                if isinstance(v, data.Data):
+                    ctypedef = dtypes.pointer(v.dtype).ctype if isinstance(v, data.Array) else v.dtype.ctype
+                    defined_type = DefinedType.Scalar if isinstance(v, data.Scalar) else DefinedType.Pointer
+                    self._dispatcher.declared_arrays.add(f"{name}.{k}", defined_type, ctypedef)
+                    self.allocate_array(sdfg, dfg, state_id, nodes.AccessNode(f"{name}.{k}"), v, function_stream,
+                                        declaration_stream, allocation_stream)
+            return
+        if isinstance(nodedesc, (data.StructureView, data.View)):
             return self.allocate_view(sdfg, dfg, state_id, node, function_stream, declaration_stream, allocation_stream)
         if isinstance(nodedesc, data.Reference):
             return self.allocate_reference(sdfg, dfg, state_id, node, function_stream, declaration_stream,
@@ -455,7 +500,7 @@ def deallocate_array(self, sdfg, dfg, state_id, node, nodedesc, function_stream,
                                               dtypes.AllocationLifetime.External)
             self._dispatcher.declared_arrays.remove(alloc_name, is_global=is_global)
 
-        if isinstance(nodedesc, (data.Scalar, data.View, data.Stream, data.Reference)):
+        if isinstance(nodedesc, (data.Scalar, data.StructureView, data.View, data.Stream, data.Reference)):
             return
         elif (nodedesc.storage == dtypes.StorageType.CPU_Heap
               or (nodedesc.storage == dtypes.StorageType.Register and symbolic.issymbolic(arrsize, sdfg.constants))):
@@ -1139,6 +1184,9 @@ def memlet_definition(self,
         if not types:
             types = self._dispatcher.defined_vars.get(ptr, is_global=True)
         var_type, ctypedef = types
+        # NOTE: `expr` may only be a name or a sequence of names and dots. The latter indicates nested data and
+        # NOTE: structures. Since structures are implemented as pointers, we replace dots with arrows.
+        ptr = ptr.replace('.', '->')
 
         if fpga.is_fpga_array(desc):
             decouple_array_interfaces = Config.get_bool("compiler", "xilinx", "decouple_array_interfaces")