PERF: Vectorized Timedelta property access (#18092) (#18225)

WillAyd · jreback · commit cd80f08fba8e · 2017-11-12T13:02:08.000-05:00
diff --git a/asv_bench/benchmarks/timedelta.py b/asv_bench/benchmarks/timedelta.py
@@ -40,3 +40,46 @@ def setup(self):
 
     def test_add_td_ts(self):
         self.td + self.ts
+
+
+class TimedeltaProperties(object):
+    goal_time = 0.2
+
+    def setup(self):
+        self.td = Timedelta(days=365, minutes=35, seconds=25, milliseconds=35)
+
+    def time_timedelta_days(self):
+        self.td.days
+
+    def time_timedelta_seconds(self):
+        self.td.seconds
+
+    def time_timedelta_microseconds(self):
+        self.td.microseconds
+
+    def time_timedelta_nanoseconds(self):
+        self.td.nanoseconds
+
+
+class DatetimeAccessor(object):
+    goal_time = 0.2
+
+    def setup(self):
+        self.N = 100000
+        self.series = pd.Series(
+            pd.timedelta_range('1 days', periods=self.N, freq='h')
+        )
+    def time_dt_accessor(self):
+        self.series.dt
+
+    def time_timedelta_dt_accessor_days(self):
+        self.series.dt.days
+
+    def time_timedelta_dt_accessor_seconds(self):
+        self.series.dt.seconds
+
+    def time_timedelta_dt_accessor_microseconds(self):
+        self.series.dt.microseconds
+
+    def time_timedelta_dt_accessor_nanoseconds(self):
+        self.series.dt.nanoseconds
diff --git a/doc/source/whatsnew/v0.22.0.txt b/doc/source/whatsnew/v0.22.0.txt
@@ -72,6 +72,7 @@ Performance Improvements
 - Indexers on ``Series`` or ``DataFrame`` no longer create a reference cycle (:issue:`17956`)
 - Added a keyword argument, ``cache``, to :func:`to_datetime` that improved the performance of converting duplicate datetime arguments (:issue:`11665`)
 - :class`DateOffset` arithmetic performance is improved (:issue:`18218`)
+- Converting a ``Series`` of ``Timedelta`` objects to days, seconds, etc... sped up through vectorization of underlying methods (:issue:`18092`)
 -
 
 .. _whatsnew_0220.docs:
diff --git a/pandas/_libs/src/datetime/np_datetime.c b/pandas/_libs/src/datetime/np_datetime.c
@@ -562,6 +562,17 @@ void pandas_datetime_to_datetimestruct(npy_datetime val, PANDAS_DATETIMEUNIT fr,
     convert_datetime_to_datetimestruct(&meta, val, result);
 }
 
+void pandas_timedelta_to_timedeltastruct(npy_timedelta val,
+                                         PANDAS_DATETIMEUNIT fr,
+                                          pandas_timedeltastruct *result) {
+  pandas_datetime_metadata meta;
+
+  meta.base = fr;
+  meta.num - 1;
+
+  convert_timedelta_to_timedeltastruct(&meta, val, result);
+}
+
 PANDAS_DATETIMEUNIT get_datetime64_unit(PyObject *obj) {
     return (PANDAS_DATETIMEUNIT)((PyDatetimeScalarObject *)obj)->obmeta.base;
 }
@@ -980,3 +991,107 @@ int convert_datetime_to_datetimestruct(pandas_datetime_metadata *meta,
 
     return 0;
 }
+
+/*
+ * Converts a timedelta from a timedeltastruct to a timedelta based
+ * on some metadata. The timedelta is assumed to be valid.
+ *
+ * Returns 0 on success, -1 on failure.
+ */
+int convert_timedelta_to_timedeltastruct(pandas_timedelta_metadata *meta,
+                                         npy_timedelta td,
+                                         pandas_timedeltastruct *out) {
+    npy_int64 perday;
+    npy_int64 frac;
+    npy_int64 sfrac;
+    npy_int64 ifrac;
+    int sign;
+    npy_int64 DAY_NS = 86400000000000LL;
+
+    /* Initialize the output to all zeros */
+    memset(out, 0, sizeof(pandas_timedeltastruct));
+
+    switch (meta->base) {
+        case PANDAS_FR_ns:
+
+        // put frac in seconds
+        if (td < 0 && td % (1000LL * 1000LL * 1000LL) != 0)
+          frac = td / (1000LL * 1000LL * 1000LL) - 1;
+        else
+            frac = td / (1000LL * 1000LL * 1000LL);
+
+          if (frac < 0) {
+            sign = -1;
+
+            // even fraction
+            if ((-frac % 86400LL) != 0) {
+              out->days = -frac / 86400LL + 1;
+              frac += 86400LL * out->days;
+            } else {
+              frac = -frac;
+            }
+          } else {
+            sign = 1;
+            out->days = 0;
+          }
+
+          if (frac >= 86400) {
+            out->days += frac / 86400LL;
+            frac -= out->days * 86400LL;
+          }
+
+          if (frac >= 3600) {
+            out->hrs = frac / 3600LL;
+            frac -= out->hrs * 3600LL;
+          } else {
+            out->hrs = 0;
+          }
+
+          if (frac >= 60) {
+            out->min = frac / 60LL;
+            frac -= out->min * 60LL;
+          } else {
+            out->min = 0;
+          }
+
+          if (frac >= 0) {
+            out->sec = frac;
+            frac -= out->sec;
+          } else {
+            out->sec = 0;
+          }
+
+          sfrac = (out->hrs * 3600LL + out->min * 60LL
+                   + out->sec) * (1000LL * 1000LL * 1000LL);
+
+          if (sign < 0)
+            out->days = -out->days;
+
+          ifrac = td - (out->days * DAY_NS + sfrac);
+
+          if (ifrac != 0) {
+            out->ms = ifrac / (1000LL * 1000LL);
+            ifrac -= out->ms * 1000LL * 1000LL;
+            out->us = ifrac / 1000LL;
+            ifrac -= out->us * 1000LL;
+            out->ns = ifrac;
+          } else {
+            out->ms = 0;
+            out->us = 0;
+            out->ns = 0;
+          }
+
+          out->seconds = out->hrs * 3600 + out->min * 60 + out->sec;
+          out->microseconds = out->ms * 1000 + out->us;
+          out->nanoseconds = out->ns;
+          break;
+
+        default:
+            PyErr_SetString(PyExc_RuntimeError,
+                            "NumPy datetime metadata is corrupted with invalid "
+                            "base unit");
+            return -1;
+    }
+
+    return 0;
+}
diff --git a/pandas/_libs/src/datetime/np_datetime.h b/pandas/_libs/src/datetime/np_datetime.h
@@ -49,11 +49,18 @@ typedef struct {
         npy_int32 month, day, hour, min, sec, us, ps, as;
 } pandas_datetimestruct;
 
+typedef struct {
+        npy_int64 days;
+        npy_int32 hrs, min, sec, ms, us, ns, seconds, microseconds, nanoseconds;
+} pandas_timedeltastruct;
+
 typedef struct {
     PANDAS_DATETIMEUNIT base;
     int num;
 } pandas_datetime_metadata;
 
+typedef pandas_datetime_metadata pandas_timedelta_metadata;
+
 extern const pandas_datetimestruct _NS_MIN_DTS;
 extern const pandas_datetimestruct _NS_MAX_DTS;
 
@@ -71,6 +78,10 @@ npy_datetime pandas_datetimestruct_to_datetime(PANDAS_DATETIMEUNIT fr,
 void pandas_datetime_to_datetimestruct(npy_datetime val, PANDAS_DATETIMEUNIT fr,
                                        pandas_datetimestruct *result);
 
+void pandas_timedelta_to_timedeltastruct(npy_timedelta val,
+                                         PANDAS_DATETIMEUNIT fr,
+                                         pandas_timedeltastruct *result);
+
 int dayofweek(int y, int m, int d);
 
 extern const int days_per_month_table[2][12];
@@ -131,6 +142,11 @@ convert_datetime_to_datetimestruct(pandas_datetime_metadata *meta,
                                    npy_datetime dt,
                                    pandas_datetimestruct *out);
 
+int
+convert_timedelta_to_timedeltastruct(pandas_timedelta_metadata *meta,
+                                     npy_timedelta td,
+                                     pandas_timedeltastruct *out);
+
 
 PANDAS_DATETIMEUNIT get_datetime64_unit(PyObject *obj);
 
diff --git a/pandas/_libs/tslibs/fields.pyx b/pandas/_libs/tslibs/fields.pyx
@@ -17,7 +17,8 @@ from numpy cimport ndarray, int64_t, int32_t, int8_t
 np.import_array()
 
 
-from np_datetime cimport pandas_datetimestruct, dt64_to_dtstruct
+from np_datetime cimport (pandas_datetimestruct, pandas_timedeltastruct,
+                          dt64_to_dtstruct, td64_to_tdstruct)
 
 from datetime cimport (
     days_per_month_table,
@@ -545,6 +546,123 @@ def get_date_field(ndarray[int64_t] dtindex, object field):
     raise ValueError("Field %s not supported" % field)
 
 
+@cython.wraparound(False)
+@cython.boundscheck(False)
+def get_timedelta_field(ndarray[int64_t] tdindex, object field):
+    """
+    Given a int64-based timedelta index, extract the days, hrs, sec.,
+    field and return an array of these values.
+    """
+    cdef:
+        Py_ssize_t i, count = 0
+        ndarray[int32_t] out
+        pandas_timedeltastruct tds
+
+    count = len(tdindex)
+    out = np.empty(count, dtype='i4')
+
+    if field == 'days':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.days
+        return out
+
+    elif field == 'h':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.hrs
+        return out
+
+    elif field == 's':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.sec
+        return out
+
+    elif field == 'seconds':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.seconds
+        return out
+
+    elif field == 'ms':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.ms
+        return out
+
+    elif field == 'microseconds':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.microseconds
+        return out
+
+    elif field == 'us':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.us
+        return out
+
+    elif field == 'ns':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.ns
+        return out
+
+    elif field == 'nanoseconds':
+        with nogil:
+            for i in range(count):
+                if tdindex[i] == NPY_NAT:
+                    out[i] = -1
+                    continue
+
+                td64_to_tdstruct(tdindex[i], &tds)
+                out[i] = tds.nanoseconds
+        return out
+
+    raise ValueError("Field %s not supported" % field)
+
+
 cdef inline int days_in_month(pandas_datetimestruct dts) nogil:
     return days_per_month_table[is_leapyear(dts.year)][dts.month - 1]
 
diff --git a/pandas/_libs/tslibs/np_datetime.pxd b/pandas/_libs/tslibs/np_datetime.pxd
@@ -30,6 +30,10 @@ cdef extern from "../src/datetime/np_datetime.h":
         int64_t year
         int32_t month, day, hour, min, sec, us, ps, as
 
+    ctypedef struct pandas_timedeltastruct:
+        int64_t days
+        int32_t hrs, min, sec, ms, us, ns, seconds, microseconds, nanoseconds
+
     ctypedef enum PANDAS_DATETIMEUNIT:
         PANDAS_FR_Y
         PANDAS_FR_M
@@ -54,6 +58,7 @@ cdef check_dts_bounds(pandas_datetimestruct *dts)
 
 cdef int64_t dtstruct_to_dt64(pandas_datetimestruct* dts) nogil
 cdef void dt64_to_dtstruct(int64_t dt64, pandas_datetimestruct* out) nogil
+cdef void td64_to_tdstruct(int64_t td64, pandas_timedeltastruct* out) nogil
 
 cdef int64_t pydatetime_to_dt64(datetime val, pandas_datetimestruct *dts)
 cdef int64_t pydate_to_dt64(date val, pandas_datetimestruct *dts)
diff --git a/pandas/_libs/tslibs/np_datetime.pyx b/pandas/_libs/tslibs/np_datetime.pyx
diff --git a/pandas/_libs/tslibs/timedeltas.pyx b/pandas/_libs/tslibs/timedeltas.pyx
diff --git a/pandas/core/indexes/timedeltas.py b/pandas/core/indexes/timedeltas.py

Original file line number	Diff line number	Diff line change
`@@ -72,6 +72,7 @@ Performance Improvements`
`72`	`72`	- Indexers on ``Series`` or ``DataFrame`` no longer create a reference cycle (:issue:`17956`)
`73`	`73`	- Added a keyword argument, ``cache``, to :func:`to_datetime` that improved the performance of converting duplicate datetime arguments (:issue:`11665`)
`74`	`74`	- :class`DateOffset` arithmetic performance is improved (:issue:`18218`)
	`75`	+- Converting a ``Series`` of ``Timedelta`` objects to days, seconds, etc... sped up through vectorization of underlying methods (:issue:`18092`)
`75`	`76`	`-`
`76`	`77`
`77`	`78`	`.. _whatsnew_0220.docs:`