Misc Python/Cython optimizations

python/cudf/cudf/_lib/table.pyx

@@ -114,7 +114,9 @@ cdef class Table:
         for _ in column_names:
             data_columns.append(Column.from_unique_ptr(move(dereference(it))))
             it += 1
-        data = dict(zip(column_names, data_columns))
+        data = ColumnAccessor._init_unsafe(
+            dict(zip(column_names, data_columns))
+        )
 
         return Table(data=data, index=index)
 

python/cudf/cudf/core/buffer.py

@@ -42,6 +42,10 @@ def __init__(
             self.ptr = data.ptr
             self.size = data.size
             self._owner = owner or data._owner
+        elif isinstance(data, rmm.DeviceBuffer):
+            self.ptr = data.ptr
+            self.size = data.size
+            self._owner = data
         elif hasattr(data, "__array_interface__") or hasattr(
             data, "__cuda_array_interface__"
         ):

python/cudf/cudf/core/column/column.py

@@ -1017,7 +1017,9 @@ def distinct_count(
         return cpp_distinct_count(self, ignore_nulls=dropna)
 
     def astype(self, dtype: Dtype, **kwargs) -> ColumnBase:
-        if is_categorical_dtype(dtype):
+        if is_numerical_dtype(dtype):
+            return self.as_numerical_column(dtype)
+        elif is_categorical_dtype(dtype):
             return self.as_categorical_column(dtype, **kwargs)
         elif pd.api.types.pandas_dtype(dtype).type in {
             np.str_,
@@ -1548,6 +1550,16 @@ def build_column(
     """
     dtype = pd.api.types.pandas_dtype(dtype)
 
+    if is_numerical_dtype(dtype):
+        assert data is not None
+        return cudf.core.column.NumericalColumn(
+            data=data,
+            dtype=dtype,
+            mask=mask,
+            size=size,
+            offset=offset,
+            null_count=null_count,
+        )
     if is_categorical_dtype(dtype):
         if not len(children) == 1:
             raise ValueError(
@@ -1634,15 +1646,7 @@ def build_column(
             children=children,
         )
     else:
-        assert data is not None
-        return cudf.core.column.NumericalColumn(
-            data=data,
-            dtype=dtype,
-            mask=mask,
-            size=size,
-            offset=offset,
-            null_count=null_count,
-        )
+        raise TypeError(f"Unrecognized dtype: {dtype}")
 
 
 def build_categorical_column(

python/cudf/cudf/core/column_accessor.py

@@ -19,11 +19,7 @@
 
 import cudf
 from cudf.core import column
-from cudf.utils.utils import (
-    cached_property,
-    to_flat_dict,
-    to_nested_dict,
-)
+from cudf.utils.utils import cached_property, to_flat_dict, to_nested_dict
 
 if TYPE_CHECKING:
     from cudf.core.column import ColumnBase
@@ -84,6 +80,19 @@ def __init__(
             self.multiindex = multiindex
             self._level_names = level_names
 
+    @classmethod
+    def _init_unsafe(
+        cls,
+        data: Dict[Any, ColumnBase],
+        multiindex: bool = False,
+        level_names=None,
+    ) -> ColumnAccessor:
+        obj = cls()
+        obj._data = data
+        obj.multiindex = multiindex
+        obj._level_names = level_names
+        return obj
+
     def __iter__(self):
         return self._data.__iter__()
 
@@ -167,7 +176,7 @@ def _column_length(self):
             return 0
 
     def _clear_cache(self):
-        cached_properties = "columns", "names", "_grouped_data"
+        cached_properties = ("columns", "names", "_grouped_data")
         for attr in cached_properties:
             try:
                 self.__delattr__(attr)

python/cudf/cudf/core/frame.py

@@ -2408,7 +2408,9 @@ def _copy_type_metadata(
         for name, col, other_col in zip(
             self._data.keys(), self._data.values(), other._data.values()
         ):
-            self._data[name] = other_col._copy_type_metadata(col)
+            self._data.set_by_label(
+                name, other_col._copy_type_metadata(col), validate=False
+            )
 
         if include_index:
             if self._index is not None and other._index is not None:

python/cudf/cudf/utils/dtypes.py

@@ -144,16 +144,13 @@ def numeric_normalize_types(*args):
 
 
 def is_numerical_dtype(obj):
-    if is_categorical_dtype(obj):
+    # TODO: we should handle objects with a `.dtype` attribute,
+    # e.g., arrays, here.
+    try:
+        dtype = np.dtype(obj)
+    except TypeError:
         return False
-    if is_list_dtype(obj):
-        return False
-    return (
-        np.issubdtype(obj, np.bool_)
-        or np.issubdtype(obj, np.floating)
-        or np.issubdtype(obj, np.signedinteger)
-        or np.issubdtype(obj, np.unsignedinteger)
-    )
+    return dtype.kind in "biuf"
 
 
 def is_string_dtype(obj):