Allow for binops between two differently sized DecimalDtypes (#16638)

Currently cudf Python has some custom logic for determining the resulting dtype of a binop between 2 decimal dtypes since Python decimal dtype support `precision` and libcudf doesn't. But libcudf does require that the 2 operands have the same decimal type when calculating the binop, so we must ensure the inputs are cast to the same, resulting dtype.

Authors:
  - Matthew Roeschke (https://github.com/mroeschke)

Approvers:
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #16638

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

@@ -135,9 +135,15 @@ def _binaryop(self, other: ColumnBinaryOperand, op: str):
         # are computed outside of libcudf
         if op in {"__add__", "__sub__", "__mul__", "__div__"}:
             output_type = _get_decimal_type(lhs.dtype, rhs.dtype, op)
+            lhs = lhs.astype(
+                type(output_type)(lhs.dtype.precision, lhs.dtype.scale)
+            )
+            rhs = rhs.astype(
+                type(output_type)(rhs.dtype.precision, rhs.dtype.scale)
+            )
             result = libcudf.binaryop.binaryop(lhs, rhs, op, output_type)
-            # TODO:  Why is this necessary? Why isn't the result's
-            # precision already set correctly based on output_type?
+            # libcudf doesn't support precision, so result.dtype doesn't
+            # maintain output_type.precision
             result.dtype.precision = output_type.precision
         elif op in {
             "__eq__",
@@ -430,7 +436,11 @@ def _with_type_metadata(
         return self
 
 
-def _get_decimal_type(lhs_dtype, rhs_dtype, op):
+def _get_decimal_type(
+    lhs_dtype: DecimalDtype,
+    rhs_dtype: DecimalDtype,
+    op: str,
+) -> DecimalDtype:
     """
     Returns the resulting decimal type after calculating
     precision & scale when performing the binary operation
@@ -441,6 +451,7 @@ def _get_decimal_type(lhs_dtype, rhs_dtype, op):
 
     # This should at some point be hooked up to libcudf's
     # binary_operation_fixed_point_scale
+    # Note: libcudf decimal types don't have a concept of precision
 
     p1, p2 = lhs_dtype.precision, rhs_dtype.precision
     s1, s2 = lhs_dtype.scale, rhs_dtype.scale

python/cudf/cudf/tests/test_decimal.py

@@ -398,3 +398,13 @@ def test_decimal_overflow():
     s = cudf.Series([1, 2], dtype=cudf.Decimal128Dtype(precision=38, scale=0))
     result = s * Decimal("1.0")
     assert_eq(cudf.Decimal128Dtype(precision=38, scale=1), result.dtype)
+
+
+def test_decimal_binop_upcast_operands():
+    ser1 = cudf.Series([0.51, 1.51, 2.51]).astype(cudf.Decimal64Dtype(18, 2))
+    ser2 = cudf.Series([0.90, 0.96, 0.99]).astype(cudf.Decimal128Dtype(19, 2))
+    result = ser1 + ser2
+    expected = cudf.Series([1.41, 2.47, 3.50]).astype(
+        cudf.Decimal128Dtype(20, 2)
+    )
+    assert_eq(result, expected)