PERF: faster groupby diff

asv_bench/benchmarks/groupby.py

@@ -417,6 +417,7 @@ class GroupByMethods:
             "cumprod",
             "cumsum",
             "describe",
+            "diff",
             "ffill",
             "first",
             "head",

doc/source/whatsnew/v1.5.0.rst

@@ -160,6 +160,7 @@ Performance improvements
 ~~~~~~~~~~~~~~~~~~~~~~~~
 - Performance improvement in :meth:`.GroupBy.transform` for some user-defined DataFrame -> Series functions (:issue:`45387`)
 - Performance improvement in :meth:`DataFrame.duplicated` when subset consists of only one column (:issue:`45236`)
+- Performance improvement in :meth:`.GroupBy.diff` (:issue:`16706`)
 -
 
 .. ---------------------------------------------------------------------------

pandas/core/groupby/groupby.py

@@ -3484,6 +3484,47 @@ def shift(self, periods=1, freq=None, axis=0, fill_value=None):
         )
         return res
 
+    @final
+    @Substitution(name="groupby")
+    @Appender(_common_see_also)
+    def diff(self, periods=1, axis=0):
+        """
+        First discrete difference of element.
+
+        Calculates the difference of each element compared with another
+        element in the group (default is element in previous row).
+
+        Parameters
+        ----------
+        periods : int, default 1
+            Periods to shift for calculating difference, accepts negative values.
+        axis : axis to shift, default 0
+            Take difference over rows (0) or columns (1).
+
+        Returns
+        -------
+        Series or DataFrame
+            First differences.
+        """
+        if axis != 0:
+            return self.apply(lambda x: x.diff(periods=periods, axis=axis))
+
+        obj = self._obj_with_exclusions
+        shifted = self.shift(periods=periods, axis=axis)
+
+        # GH45562 - to retain existing behavior and match behavior of Series.diff(),
+        # int8 and int16 are coerced to float32 rather than float64.
+        dtypes_to_f32 = ["int8", "int16"]
+        if obj.ndim == 1:
+            if obj.dtype in dtypes_to_f32:
+                shifted = shifted.astype("float32")
+        else:
+            mask = obj.dtypes.astype(str).isin(dtypes_to_f32).values
+            if mask.any():
+                shifted.loc[:, mask] = shifted.loc[:, mask].astype("float32")
+
+        return obj - shifted
+
     @final
     @Substitution(name="groupby")
     @Appender(_common_see_also)