ENH: resolution inference for array_to_timedelta64

pandas/_libs/lib.pyx

@@ -106,7 +106,6 @@ from pandas._libs.tslibs.nattype cimport (
 )
 from pandas._libs.tslibs.offsets cimport is_offset_object
 from pandas._libs.tslibs.period cimport is_period_object
-from pandas._libs.tslibs.timedeltas cimport convert_to_timedelta64
 from pandas._libs.tslibs.timezones cimport tz_compare
 
 # constants that will be compared to potentially arbitrarily large
@@ -2670,11 +2669,6 @@ def maybe_convert_objects(ndarray[object] objects,
         elif is_timedelta(val):
             if convert_non_numeric:
                 seen.timedelta_ = True
-                try:
-                    convert_to_timedelta64(val, "ns")
-                except OutOfBoundsTimedelta:
-                    seen.object_ = True
-                    break
                 break
             else:
                 seen.object_ = True

pandas/_libs/tslibs/conversion.pxd

@@ -45,9 +45,6 @@ cdef int64_t get_datetime64_nanos(object val, NPY_DATETIMEUNIT reso) except? -1
 
 cpdef datetime localize_pydatetime(datetime dt, tzinfo tz)
 cdef int64_t cast_from_unit(object ts, str unit, NPY_DATETIMEUNIT out_reso=*) except? -1
-cdef (int64_t, int) precision_from_unit(
-    NPY_DATETIMEUNIT in_reso, NPY_DATETIMEUNIT out_reso=*
-)
 
 cdef maybe_localize_tso(_TSObject obj, tzinfo tz, NPY_DATETIMEUNIT reso)
 

pandas/_libs/tslibs/timedeltas.pxd

@@ -9,7 +9,6 @@ cpdef int64_t get_unit_for_round(freq, NPY_DATETIMEUNIT creso) except? -1
 cpdef int64_t delta_to_nanoseconds(
     delta, NPY_DATETIMEUNIT reso=*, bint round_ok=*
 ) except? -1
-cdef convert_to_timedelta64(object ts, str unit)
 cdef bint is_any_td_scalar(object obj)
 
 

pandas/_libs/tslibs/timedeltas.pyi

@@ -70,6 +70,7 @@ def array_to_timedelta64(
     values: npt.NDArray[np.object_],
     unit: str | None = ...,
     errors: str = ...,
+    creso: int = ...,
 ) -> np.ndarray: ...  # np.ndarray[m8ns]
 def parse_timedelta_unit(unit: str | None) -> UnitChoices: ...
 def delta_to_nanoseconds(

pandas/_libs/tslibs/timedeltas.pyx

@@ -41,7 +41,6 @@ from pandas._libs.missing cimport checknull_with_nat_and_na
 from pandas._libs.tslibs.base cimport ABCTimestamp
 from pandas._libs.tslibs.conversion cimport (
     cast_from_unit,
-    precision_from_unit,
 )
 from pandas._libs.tslibs.dtypes cimport (
     c_DEPR_UNITS,
@@ -289,68 +288,6 @@ cpdef int64_t delta_to_nanoseconds(
         ) from err
 
 
-@cython.overflowcheck(True)
-cdef object ensure_td64ns(object ts):
-    """
-    Overflow-safe implementation of td64.astype("m8[ns]")
-
-    Parameters
-    ----------
-    ts : np.timedelta64
-
-    Returns
-    -------
-    np.timedelta64[ns]
-    """
-    cdef:
-        NPY_DATETIMEUNIT td64_unit
-        int64_t td64_value, mult
-
-    td64_unit = get_datetime64_unit(ts)
-    if (
-        td64_unit != NPY_DATETIMEUNIT.NPY_FR_ns
-        and td64_unit != NPY_DATETIMEUNIT.NPY_FR_GENERIC
-    ):
-
-        td64_value = cnp.get_timedelta64_value(ts)
-
-        mult = precision_from_unit(td64_unit)[0]
-        try:
-            # NB: cython#1381 this cannot be *=
-            td64_value = td64_value * mult
-        except OverflowError as err:
-            raise OutOfBoundsTimedelta(ts) from err
-
-        return np.timedelta64(td64_value, "ns")
-
-    return ts
-
-
-cdef convert_to_timedelta64(object ts, str unit):
-    """
-    Convert an incoming object to a timedelta64 if possible.
-    Before calling, unit must be standardized to avoid repeated unit conversion
-
-    Handle these types of objects:
-        - timedelta/Timedelta
-
-    Return an timedelta64[ns] object
-    """
-    # Caller is responsible for checking unit not in ["Y", "y", "M"]
-    if isinstance(ts, _Timedelta):
-        # already in the proper format
-        if ts._creso != NPY_FR_ns:
-            ts = ts.as_unit("ns").asm8
-        else:
-            ts = np.timedelta64(ts._value, "ns")
-
-    elif PyDelta_Check(ts):
-        ts = np.timedelta64(delta_to_nanoseconds(ts), "ns")
-    elif not cnp.is_timedelta64_object(ts):
-        raise TypeError(f"Invalid type for timedelta scalar: {type(ts)}")
-    return ts.astype("timedelta64[ns]")
-
-
 cdef _numeric_to_td64ns(object item, str unit):
     # caller is responsible for checking
     #  assert unit not in ["Y", "y", "M"]
@@ -369,10 +306,34 @@ cdef _numeric_to_td64ns(object item, str unit):
     return ts
 
 
+# TODO: de-duplicate with DatetimeParseState
+cdef class ResoState:
+    cdef:
+        NPY_DATETIMEUNIT creso
+        bint creso_ever_changed
+
+    def __cinit__(self, NPY_DATETIMEUNIT creso):
+        self.creso = creso
+        self.creso_ever_changed = False
+
+    cdef bint update_creso(self, NPY_DATETIMEUNIT item_reso) noexcept:
+        # Return a bool indicating whether we bumped to a higher resolution
+        if self.creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC:
+            self.creso = item_reso
+        elif item_reso > self.creso:
+            self.creso = item_reso
+            self.creso_ever_changed = True
+            return True
+        return False
+
+
 @cython.boundscheck(False)
 @cython.wraparound(False)
 def array_to_timedelta64(
-    ndarray values, str unit=None, str errors="raise"
+    ndarray values,
+    str unit=None,
+    str errors="raise",
+    NPY_DATETIMEUNIT creso=NPY_DATETIMEUNIT.NPY_FR_GENERIC,
 ) -> ndarray:
     # values is object-dtype, may be 2D
     """
@@ -394,6 +355,10 @@ def array_to_timedelta64(
         cnp.broadcast mi = cnp.PyArray_MultiIterNew2(result, values)
         cnp.flatiter it
         str parsed_unit = parse_timedelta_unit(unit or "ns")
+        NPY_DATETIMEUNIT item_reso
+        ResoState state = ResoState(creso)
+        bint infer_reso = creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC
+        ndarray iresult = result.view("i8")
 
     if values.descr.type_num != cnp.NPY_OBJECT:
         # raise here otherwise we segfault below
@@ -421,18 +386,58 @@ def array_to_timedelta64(
                 ival = NPY_NAT
 
             elif cnp.is_timedelta64_object(item):
-                td64ns_obj = ensure_td64ns(item)
-                ival = cnp.get_timedelta64_value(td64ns_obj)
+                # TODO: de-duplicate this with Timedelta.__new__
+                ival = cnp.get_timedelta64_value(item)
+                dt64_reso = get_datetime64_unit(item)
+                if not (
+                    is_supported_unit(dt64_reso) or
+                    dt64_reso in [
+                        NPY_DATETIMEUNIT.NPY_FR_m,
+                        NPY_DATETIMEUNIT.NPY_FR_h,
+                        NPY_DATETIMEUNIT.NPY_FR_D,
+                        NPY_DATETIMEUNIT.NPY_FR_W,
+                        NPY_DATETIMEUNIT.NPY_FR_GENERIC
+                    ]
+                ):
+                    err = npy_unit_to_abbrev(dt64_reso)
+                    raise ValueError(
+                        f"Unit {err} is not supported. "
+                        "Only unambiguous timedelta values durations are supported. "
+                        "Allowed units are 'W', 'D', 'h', 'm', 's', 'ms', 'us', 'ns'")
+
+                item_reso = get_supported_reso(dt64_reso)
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+                if dt64_reso != NPY_DATETIMEUNIT.NPY_FR_GENERIC:
+                    try:
+                        ival = convert_reso(
+                            ival,
+                            dt64_reso,
+                            creso,
+                            round_ok=True,
+                        )
+                    except (OverflowError, OutOfBoundsDatetime) as err:
+                        raise OutOfBoundsTimedelta(item) from err
+                else:
+                    # e.g. NaT
+                    pass
 
             elif isinstance(item, _Timedelta):
-                if item._creso != NPY_FR_ns:
-                    ival = item.as_unit("ns")._value
-                else:
-                    ival = item._value
+                item_reso = item._creso
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+
+                ival = (<_Timedelta>item)._as_creso(creso)._value
 
             elif PyDelta_Check(item):
                 # i.e. isinstance(item, timedelta)
-                ival = delta_to_nanoseconds(item)
+                item_reso = NPY_DATETIMEUNIT.NPY_FR_us
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+                ival = delta_to_nanoseconds(item, reso=creso)
 
             elif isinstance(item, str):
                 if (
@@ -443,13 +448,27 @@ def array_to_timedelta64(
                 else:
                     ival = parse_timedelta_string(item)
 
+                item_reso = NPY_FR_ns
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+
             elif is_tick_object(item):
-                ival = item.nanos
+                item_reso = get_supported_reso(item._creso)
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+                ival = delta_to_nanoseconds(item, reso=creso)
 
             elif is_integer_object(item) or is_float_object(item):
                 td64ns_obj = _numeric_to_td64ns(item, parsed_unit)
                 ival = cnp.get_timedelta64_value(td64ns_obj)
 
+                item_reso = NPY_FR_ns
+                state.update_creso(item_reso)
+                if infer_reso:
+                    creso = state.creso
+
             else:
                 raise TypeError(f"Invalid type for timedelta scalar: {type(item)}")
 
@@ -467,7 +486,29 @@ def array_to_timedelta64(
 
         cnp.PyArray_MultiIter_NEXT(mi)
 
-    return result
+    if infer_reso:
+        if state.creso_ever_changed:
+            # We encountered mismatched resolutions, need to re-parse with
+            #  the correct one.
+            return array_to_timedelta64(
+                values,
+                unit=unit,
+                errors=errors,
+                creso=state.creso,
+            )
+        elif state.creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC:
+            # i.e. we never encountered anything non-NaT, default to "s". This
+            # ensures that insert and concat-like operations with NaT
+            # do not upcast units
+            result = iresult.view("m8[s]")
+        else:
+            # Otherwise we can use the single reso that we encountered and avoid
+            #  a second pass.
+            abbrev = npy_unit_to_abbrev(state.creso)
+            result = iresult.view(f"m8[{abbrev}]")
+
+    abbrev = npy_unit_to_abbrev(creso)
+    return result.view(f"m8[{abbrev}]")
 
 
 @cython.cpow(True)

pandas/core/arrays/timedeltas.py

@@ -1251,7 +1251,7 @@ def _objects_to_td64ns(
     values = np.asarray(data, dtype=np.object_)
 
     result = array_to_timedelta64(values, unit=unit, errors=errors)
-    return result.view("timedelta64[ns]")
+    return result
 
 
 def _validate_td64_dtype(dtype) -> DtypeObj:

pandas/tests/arithmetic/test_timedelta64.py

@@ -717,7 +717,7 @@ def test_tdi_add_overflow(self):
             )
 
         # These should not overflow!
-        exp = TimedeltaIndex([NaT])
+        exp = TimedeltaIndex([NaT], dtype="m8[ns]")
         result = pd.to_timedelta([NaT]) - Timedelta("1 days")
         tm.assert_index_equal(result, exp)
 
@@ -2216,7 +2216,7 @@ def test_float_series_rdiv_td64arr(self, box_with_array, names):
 
     def test_td64arr_all_nat_div_object_dtype_numeric(self, box_with_array):
         # GH#39750 make sure we infer the result as td64
-        tdi = TimedeltaIndex([NaT, NaT])
+        tdi = TimedeltaIndex([NaT, NaT], dtype="m8[ns]")
 
         left = tm.box_expected(tdi, box_with_array)
         right = np.array([2, 2.0], dtype=object)

pandas/tests/dtypes/test_inference.py

@@ -808,7 +808,7 @@ def test_maybe_convert_objects_datetime(self):
         tm.assert_numpy_array_equal(out, exp)
 
         arr = np.array([pd.NaT, np.timedelta64(1, "s")], dtype=object)
-        exp = np.array([np.timedelta64("NaT"), np.timedelta64(1, "s")], dtype="m8[ns]")
+        exp = np.array([np.timedelta64("NaT"), np.timedelta64(1, "s")], dtype="m8[s]")
         out = lib.maybe_convert_objects(arr, convert_non_numeric=True)
         tm.assert_numpy_array_equal(out, exp)
 
@@ -863,7 +863,7 @@ def test_maybe_convert_objects_datetime_overflow_safe(self, dtype):
         if dtype == "datetime64[ns]":
             expected = np.array(["2363-10-04"], dtype="M8[us]")
         else:
-            expected = arr
+            expected = arr.astype("m8[us]")
         tm.assert_numpy_array_equal(out, expected)
 
     def test_maybe_convert_objects_mixed_datetimes(self):

pandas/tests/extension/test_arrow.py

@@ -591,14 +591,6 @@ def test_reduce_frame(self, data, all_numeric_reductions, skipna, request):
             if data.dtype._is_numeric:
                 mark = pytest.mark.xfail(reason="skew not implemented")
                 request.applymarker(mark)
-        elif (
-            op_name in ["std", "sem"]
-            and pa.types.is_date64(data._pa_array.type)
-            and skipna
-        ):
-            # overflow
-            mark = pytest.mark.xfail(reason="Cannot cast")
-            request.applymarker(mark)
         return super().test_reduce_frame(data, all_numeric_reductions, skipna)
 
     @pytest.mark.parametrize("typ", ["int64", "uint64", "float64"])

pandas/tests/frame/methods/test_dtypes.py

@@ -103,7 +103,7 @@ def test_dtypes_timedeltas(self):
         )
         result = df.dtypes
         expected = Series(
-            [np.dtype("datetime64[ns]"), np.dtype("timedelta64[ns]")], index=list("AB")
+            [np.dtype("datetime64[ns]"), np.dtype("timedelta64[us]")], index=list("AB")
         )
         tm.assert_series_equal(result, expected)
 
@@ -112,7 +112,7 @@ def test_dtypes_timedeltas(self):
         expected = Series(
             [
                 np.dtype("datetime64[ns]"),
-                np.dtype("timedelta64[ns]"),
+                np.dtype("timedelta64[us]"),
                 np.dtype("datetime64[ns]"),
             ],
             index=list("ABC"),
@@ -125,7 +125,7 @@ def test_dtypes_timedeltas(self):
         expected = Series(
             [
                 np.dtype("datetime64[ns]"),
-                np.dtype("timedelta64[ns]"),
+                np.dtype("timedelta64[us]"),
                 np.dtype("datetime64[ns]"),
                 np.dtype("int64"),
             ],