PERF: DataFrame.where for EA dtype mask

asv_bench/benchmarks/frame_methods.py

@@ -754,4 +754,19 @@ def time_memory_usage_object_dtype(self):
         self.df2.memory_usage(deep=True)
 
 
+class Where:
+    params = (
+        [True, False],
+        ["float64", "Float64", "float64[pyarrow]"],
+    )
+    param_names = ["dtype"]
+
+    def setup(self, inplace, dtype):
+        self.df = DataFrame(np.random.randn(100_000, 10), dtype=dtype)
+        self.mask = self.df < 0
+
+    def time_where(self, inplace, dtype):
+        self.df.where(self.mask, other=0.0, inplace=inplace)
+
+
 from .pandas_vb_common import setup  # noqa: F401 isort:skip

doc/source/whatsnew/v2.1.0.rst

@@ -100,6 +100,7 @@ Deprecations
 
 Performance improvements
 ~~~~~~~~~~~~~~~~~~~~~~~~
+- Performance improvement in :meth:`DataFrame.where` when ``cond`` is backed by an extension dtype (:issue:`51574`)
 - Performance improvement in :meth:`DataFrame.clip` and :meth:`Series.clip` (:issue:`51472`)
 -
 

pandas/core/internals/managers.py

@@ -359,6 +359,14 @@ def where(self: T, other, cond, align: bool) -> T:
             align_keys = ["cond"]
             other = extract_array(other, extract_numpy=True)
 
+        if isinstance(cond, ABCDataFrame) and cond._mgr.any_extension_types:
+            # GH51574: avoid object ndarray conversion later on
+            cond = cond._constructor(
+                cond.to_numpy(dtype=bool, na_value=False),
+                index=cond.index,
+                columns=cond.columns,
+            )
+
         return self.apply(
             "where",
             align_keys=align_keys,
@@ -390,6 +398,14 @@ def putmask(self, mask, new, align: bool = True):
             align_keys = ["mask"]
             new = extract_array(new, extract_numpy=True)
 
+        if isinstance(mask, ABCDataFrame) and mask._mgr.any_extension_types:
+            # GH51574: avoid object ndarray conversion later on
+            mask = mask._constructor(
+                mask.to_numpy(dtype=bool, na_value=False),
+                index=mask.index,
+                columns=mask.columns,
+            )
+
         return self.apply(
             "putmask",
             align_keys=align_keys,