PERF: avoid potentially-copying ravel (pandas-dev#46733)

Author: jbrockmendel

pandas/_libs/tslibs/conversion.pyx

@@ -2,6 +2,7 @@ import cython
 import numpy as np
 
 cimport numpy as cnp
+from cpython.object cimport PyObject
 from numpy cimport (
     int32_t,
     int64_t,
@@ -273,7 +274,8 @@ def ensure_timedelta64ns(arr: ndarray, copy: bool = True):
 
 @cython.boundscheck(False)
 @cython.wraparound(False)
-def datetime_to_datetime64(ndarray[object] values):
+def datetime_to_datetime64(ndarray values):
+    # ndarray[object], but can't declare object without ndim
     """
     Convert ndarray of datetime-like objects to int64 array representing
     nanosecond timestamps.
@@ -288,20 +290,27 @@ def datetime_to_datetime64(ndarray[object] values):
     inferred_tz : tzinfo or None
     """
     cdef:
-        Py_ssize_t i, n = len(values)
+        Py_ssize_t i, n = values.size
         object val
-        int64_t[:] iresult
+        int64_t ival
+        ndarray iresult  # int64_t, but can't declare that without specifying ndim
         npy_datetimestruct dts
         _TSObject _ts
         bint found_naive = False
         tzinfo inferred_tz = None
 
-    result = np.empty(n, dtype='M8[ns]')
+        cnp.broadcast mi
+
+    result = np.empty((<object>values).shape, dtype='M8[ns]')
     iresult = result.view('i8')
+
+    mi = cnp.PyArray_MultiIterNew2(iresult, values)
     for i in range(n):
-        val = values[i]
+        # Analogous to: val = values[i]
+        val = <object>(<PyObject**>cnp.PyArray_MultiIter_DATA(mi, 1))[0]
+
         if checknull_with_nat(val):
-            iresult[i] = NPY_NAT
+            ival = NPY_NAT
         elif PyDateTime_Check(val):
             if val.tzinfo is not None:
                 if found_naive:
@@ -314,18 +323,23 @@ def datetime_to_datetime64(ndarray[object] values):
                     inferred_tz = val.tzinfo
 
                 _ts = convert_datetime_to_tsobject(val, None)
-                iresult[i] = _ts.value
+                ival = _ts.value
                 check_dts_bounds(&_ts.dts)
             else:
                 found_naive = True
                 if inferred_tz is not None:
                     raise ValueError('Cannot mix tz-aware with '
                                      'tz-naive values')
-                iresult[i] = pydatetime_to_dt64(val, &dts)
+                ival = pydatetime_to_dt64(val, &dts)
                 check_dts_bounds(&dts)
         else:
             raise TypeError(f'Unrecognized value type: {type(val)}')
 
+        # Analogous to: iresult[i] = ival
+        (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = ival
+
+        cnp.PyArray_MultiIter_NEXT(mi)
+
     return result, inferred_tz
 
 

pandas/core/arrays/datetimes.py

@@ -2247,10 +2247,9 @@ def objects_to_datetime64ns(
         result = result.reshape(data.shape, order=order)
     except ValueError as err:
         try:
-            values, tz_parsed = conversion.datetime_to_datetime64(data.ravel("K"))
+            values, tz_parsed = conversion.datetime_to_datetime64(data)
             # If tzaware, these values represent unix timestamps, so we
             #  return them as i8 to distinguish from wall times
-            values = values.reshape(data.shape, order=order)
             return values.view("i8"), tz_parsed
         except (ValueError, TypeError):
             raise err

pandas/core/arrays/timedeltas.py

@@ -429,14 +429,15 @@ def _formatter(self, boxed: bool = False):
 
         return get_format_timedelta64(self, box=True)
 
-    @dtl.ravel_compat
     def _format_native_types(
         self, *, na_rep="NaT", date_format=None, **kwargs
     ) -> npt.NDArray[np.object_]:
         from pandas.io.formats.format import get_format_timedelta64
 
         formatter = get_format_timedelta64(self._ndarray, na_rep)
-        return np.array([formatter(x) for x in self._ndarray])
+        # equiv: np.array([formatter(x) for x in self._ndarray])
+        #  but independent of dimension
+        return np.frompyfunc(formatter, 1, 1)(self._ndarray)
 
     # ----------------------------------------------------------------
     # Arithmetic Methods