From 37f588ff749e9617955e5e5f824f094b0b63eb1e Mon Sep 17 00:00:00 2001
From: Brock <jbrockmendel@gmail.com>
Date: Mon, 30 Aug 2021 13:48:40 -0700
Subject: [PATCH 1/2] PERF: nanops

---
 pandas/core/nanops.py | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/pandas/core/nanops.py b/pandas/core/nanops.py
index a80bd8ba76dac..c10efc7f379dd 100644
--- a/pandas/core/nanops.py
+++ b/pandas/core/nanops.py
@@ -449,6 +449,25 @@ def _na_for_min_count(values: np.ndarray, axis: int | None) -> Scalar | np.ndarr
         return np.full(result_shape, fill_value, dtype=values.dtype)
 
 
+def maybe_operate_rowwise(func):
+    """
+    NumPy operations on C-contiguous ndarrays with axis=1 can be
+    very slow. Operate row-by-row and concatenate the results.
+    """
+
+    @functools.wraps(func)
+    def newfunc(values: np.ndarray, *, axis: int | None = None, **kwargs):
+        if axis == 1:
+            if values.ndim == 2 and values.flags["C_CONTIGUOUS"]:
+                arrs = list(values)
+                results = [func(x, **kwargs) for x in arrs]
+                return np.array(results)
+
+        return func(values, axis=axis, **kwargs)
+
+    return newfunc
+
+
 def nanany(
     values: np.ndarray,
     *,
@@ -543,6 +562,7 @@ def nanall(
 
 @disallow("M8")
 @_datetimelike_compat
+@maybe_operate_rowwise
 def nansum(
     values: np.ndarray,
     *,
@@ -1111,6 +1131,7 @@ def nanargmin(
 
 
 @disallow("M8", "m8")
+@maybe_operate_rowwise
 def nanskew(
     values: np.ndarray,
     *,
@@ -1198,6 +1219,7 @@ def nanskew(
 
 
 @disallow("M8", "m8")
+@maybe_operate_rowwise
 def nankurt(
     values: np.ndarray,
     *,
@@ -1294,6 +1316,7 @@ def nankurt(
 
 
 @disallow("M8", "m8")
+@maybe_operate_rowwise
 def nanprod(
     values: np.ndarray,
     *,

From 5074138ef279e4c48922a53c466ef8e674cb4101 Mon Sep 17 00:00:00 2001
From: Brock <jbrockmendel@gmail.com>
Date: Mon, 30 Aug 2021 16:46:26 -0700
Subject: [PATCH 2/2] corner cases

---
 pandas/core/nanops.py | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/pandas/core/nanops.py b/pandas/core/nanops.py
index c10efc7f379dd..5d96e9bb6cd19 100644
--- a/pandas/core/nanops.py
+++ b/pandas/core/nanops.py
@@ -457,11 +457,21 @@ def maybe_operate_rowwise(func):
 
     @functools.wraps(func)
     def newfunc(values: np.ndarray, *, axis: int | None = None, **kwargs):
-        if axis == 1:
-            if values.ndim == 2 and values.flags["C_CONTIGUOUS"]:
-                arrs = list(values)
+        if (
+            axis == 1
+            and values.ndim == 2
+            and values.flags["C_CONTIGUOUS"]
+            and values.dtype != object
+        ):
+            arrs = list(values)
+            if kwargs.get("mask") is not None:
+                mask = kwargs.pop("mask")
+                results = [
+                    func(arrs[i], mask=mask[i], **kwargs) for i in range(len(arrs))
+                ]
+            else:
                 results = [func(x, **kwargs) for x in arrs]
-                return np.array(results)
+            return np.array(results)
 
         return func(values, axis=axis, **kwargs)