do not force host transfers when computing norms

meshmode/dof_array.py

@@ -663,10 +663,32 @@ def _unflatten_from_numpy(subary):
 
 # {{{ flat_norm
 
+def _reduce_norm(actx, arys, ord):
+    from numbers import Number
+    from functools import reduce
+
+    # NOTE: these are ordered by an expected usage frequency
+    if ord == 2:
+        return actx.np.sqrt(sum(subary*subary for subary in arys))
+    elif ord == np.inf:
+        return reduce(actx.np.maximum, arys)
+    elif ord == -np.inf:
+        return reduce(actx.np.minimum, arys)
+    elif isinstance(ord, Number) and ord > 0:
+        return sum(subary**ord for subary in arys)**(1/ord)
+    else:
+        raise NotImplementedError(f"unsupported value of 'ord': {ord}")
+
+
 def flat_norm(ary, ord=None) -> Any:
     r"""Return an element-wise :math:`\ell^{\text{ord}}` norm of *ary*.
 
-    :arg ary: may be a :class:`DOFArray` or a
+    Unlike :attr:`arraycontext.ArrayContext.np`, this function handles
+    :class:`DOFArray`\ s by taking a norm of their flattened values
+    (in the sense of :func:`flatten`) regardless of how the group arrays
+    are stored.
+
+    :arg ary: may be a :class:`DOFArray` or an
         :class:`~arraycontext.ArrayContainer` containing them.
     """
 
@@ -678,24 +700,31 @@ def flat_norm(ary, ord=None) -> Any:
         ord = 2
 
     from arraycontext import is_array_container
+    actx = None
+
+    def _rec(_ary):
+        nonlocal actx
+
+        if isinstance(_ary, DOFArray):
+            if actx is None:
+                actx = _ary.array_context
+            else:
+                assert actx is _ary.array_context
+
+            assert actx is not None
+
+            return _reduce_norm(actx, [
+                actx.np.linalg.norm(actx.np.ravel(subary, order="A"), ord=ord)
+                for _, subary in serialize_container(_ary)
+                ], ord=ord)
+
+        elif is_array_container(_ary):
+            arys = [_rec(subary) for _, subary in serialize_container(_ary)]
+            return _reduce_norm(actx, arys, ord=ord)
+
+        raise TypeError(f"unsupported array type: '{type(_ary).__name__}'")
 
-    from arraycontext.fake_numpy import _scalar_list_norm
-    if isinstance(ary, DOFArray):
-        actx = ary.array_context
-        return _scalar_list_norm(
-                [
-                    actx.np.linalg.norm(actx.np.ravel(subary, order="A"), ord=ord)
-                    for _, subary in serialize_container(ary)],
-                ord=ord)
-
-    elif is_array_container(ary):
-        return _scalar_list_norm(
-                [flat_norm(subary, ord=ord)
-                    for _, subary in serialize_container(ary)],
-                ord=ord)
-
-    raise TypeError(
-            f"unsupported array type passed to flat_norm: '{type(ary).__name__}'")
+    return _rec(ary)
 
 # }}}