Add cudf python groupby.diff (#9446)

Fixes #5079
- [x] add cudf python groupby diff (implemented using groupby.shift)
- [x] add unit tests

Authors:
  - Karthikeyan (https://github.com/karthikeyann)

Approvers:
  - GALI PREM SAGAR (https://github.com/galipremsagar)
  - Ashwin Srinath (https://github.com/shwina)

URL: #9446

docs/cudf/source/api_docs/groupby.rst

@@ -77,6 +77,7 @@ application to columns of a specific data type.
    DataFrameGroupBy.cummin
    DataFrameGroupBy.cumsum
    DataFrameGroupBy.describe
+   DataFrameGroupBy.diff
    DataFrameGroupBy.ffill
    DataFrameGroupBy.fillna
    DataFrameGroupBy.idxmax

python/cudf/cudf/core/groupby/groupby.py

@@ -860,6 +860,38 @@ def last(self):
         """Get the last non-null value in each group."""
         return self.agg("last")
 
+    def diff(self, periods=1, axis=0):
+        """Get the difference between the values in each group.
+
+        Parameters
+        ----------
+        periods : int, default 1
+            Periods to shift for calculating difference,
+            accepts negative values.
+        axis : {0 or 'index', 1 or 'columns'}, default 0
+            Take difference over rows (0) or columns (1).
+            Only row-wise (0) shift is supported.
+
+        Returns
+        -------
+        Series or DataFrame
+            First differences of the Series or DataFrame.
+        """
+
+        if not axis == 0:
+            raise NotImplementedError("Only axis=0 is supported.")
+
+        # grouped values
+        value_columns = self.grouping.values
+        _, (data, index), _ = self._groupby.groups(
+            cudf.core.frame.Frame(value_columns._data)
+        )
+        grouped = self.obj.__class__._from_data(data, index)
+        grouped = self._mimic_pandas_order(grouped)
+
+        result = grouped - self.shift(periods=periods)
+        return result._copy_type_metadata(value_columns)
+
     def _scan_fill(self, method: str, limit: int) -> DataFrameOrSeries:
         """Internal implementation for `ffill` and `bfill`
         """

python/cudf/cudf/tests/test_groupby.py

@@ -1913,6 +1913,89 @@ def test_groupby_shift_row_zero_shift(nelem, fill_value):
     )
 
 
+@pytest.mark.parametrize("nelem", [2, 3, 100, 1000])
+@pytest.mark.parametrize("shift_perc", [0.5, 1.0, 1.5])
+@pytest.mark.parametrize("direction", [1, -1])
+def test_groupby_diff_row(nelem, shift_perc, direction):
+    pdf = make_frame(pd.DataFrame, nelem=nelem, extra_vals=["val2"])
+    gdf = cudf.from_pandas(pdf)
+    n_shift = int(nelem * shift_perc) * direction
+
+    expected = pdf.groupby(["x", "y"]).diff(periods=n_shift)
+    got = gdf.groupby(["x", "y"]).diff(periods=n_shift)
+
+    assert_groupby_results_equal(
+        expected[["val", "val2"]], got[["val", "val2"]]
+    )
+
+
+@pytest.mark.parametrize("nelem", [10, 50, 100, 1000])
+@pytest.mark.parametrize("shift_perc", [0.5, 1.0, 1.5])
+@pytest.mark.parametrize("direction", [1, -1])
+def test_groupby_diff_row_mixed_numerics(nelem, shift_perc, direction):
+    t = rand_dataframe(
+        dtypes_meta=[
+            {"dtype": "int64", "null_frequency": 0, "cardinality": 10},
+            {"dtype": "int64", "null_frequency": 0.4, "cardinality": 10},
+            {"dtype": "float32", "null_frequency": 0.4, "cardinality": 10},
+            {"dtype": "decimal64", "null_frequency": 0.4, "cardinality": 10},
+            {
+                "dtype": "datetime64[ns]",
+                "null_frequency": 0.4,
+                "cardinality": 10,
+            },
+            {
+                "dtype": "timedelta64[ns]",
+                "null_frequency": 0.4,
+                "cardinality": 10,
+            },
+        ],
+        rows=nelem,
+        use_threads=False,
+    )
+    pdf = t.to_pandas()
+    gdf = cudf.from_pandas(pdf)
+    n_shift = int(nelem * shift_perc) * direction
+
+    expected = pdf.groupby(["0"]).diff(periods=n_shift)
+    got = gdf.groupby(["0"]).diff(periods=n_shift)
+
+    assert_groupby_results_equal(
+        expected[["1", "2", "3", "4", "5"]], got[["1", "2", "3", "4", "5"]]
+    )
+
+
+@pytest.mark.parametrize("nelem", [10, 50, 100, 1000])
+def test_groupby_diff_row_zero_shift(nelem):
+    t = rand_dataframe(
+        dtypes_meta=[
+            {"dtype": "int64", "null_frequency": 0, "cardinality": 10},
+            {"dtype": "int64", "null_frequency": 0.4, "cardinality": 10},
+            {"dtype": "float32", "null_frequency": 0.4, "cardinality": 10},
+            {
+                "dtype": "datetime64[ns]",
+                "null_frequency": 0.4,
+                "cardinality": 10,
+            },
+            {
+                "dtype": "timedelta64[ns]",
+                "null_frequency": 0.4,
+                "cardinality": 10,
+            },
+        ],
+        rows=nelem,
+        use_threads=False,
+    )
+    gdf = cudf.from_pandas(t.to_pandas())
+
+    expected = gdf
+    got = gdf.groupby(["0"]).shift(periods=0)
+
+    assert_groupby_results_equal(
+        expected[["1", "2", "3", "4"]], got[["1", "2", "3", "4"]]
+    )
+
+
 # TODO: test for category columns when cudf.Scalar supports category type
 @pytest.mark.parametrize("nelem", [10, 100, 1000])
 def test_groupby_fillna_multi_value(nelem):