PERF: using murmur hash for float64 khash-tables (#36729)

Author: realead

asv_bench/benchmarks/hash_functions.py

@@ -0,0 +1,164 @@
+import numpy as np
+
+import pandas as pd
+
+
+class IsinAlmostFullWithRandomInt:
+    params = [
+        [np.float64, np.int64, np.uint64, np.object],
+        range(10, 21),
+    ]
+    param_names = ["dtype", "exponent"]
+
+    def setup(self, dtype, exponent):
+        M = 3 * 2 ** (exponent - 2)
+        # 0.77-the maximal share of occupied buckets
+        np.random.seed(42)
+        self.s = pd.Series(np.random.randint(0, M, M)).astype(dtype)
+        self.values = np.random.randint(0, M, M).astype(dtype)
+        self.values_outside = self.values + M
+
+    def time_isin(self, dtype, exponent):
+        self.s.isin(self.values)
+
+    def time_isin_outside(self, dtype, exponent):
+        self.s.isin(self.values_outside)
+
+
+class IsinWithRandomFloat:
+    params = [
+        [np.float64, np.object],
+        [
+            1_300,
+            2_000,
+            7_000,
+            8_000,
+            70_000,
+            80_000,
+            750_000,
+            900_000,
+        ],
+    ]
+    param_names = ["dtype", "M"]
+
+    def setup(self, dtype, M):
+        np.random.seed(42)
+        self.values = np.random.rand(M)
+        self.s = pd.Series(self.values).astype(dtype)
+        np.random.shuffle(self.values)
+        self.values_outside = self.values + 0.1
+
+    def time_isin(self, dtype, M):
+        self.s.isin(self.values)
+
+    def time_isin_outside(self, dtype, M):
+        self.s.isin(self.values_outside)
+
+
+class IsinWithArangeSorted:
+    params = [
+        [np.float64, np.int64, np.uint64, np.object],
+        [
+            1_000,
+            2_000,
+            8_000,
+            100_000,
+            1_000_000,
+        ],
+    ]
+    param_names = ["dtype", "M"]
+
+    def setup(self, dtype, M):
+        self.s = pd.Series(np.arange(M)).astype(dtype)
+        self.values = np.arange(M).astype(dtype)
+
+    def time_isin(self, dtype, M):
+        self.s.isin(self.values)
+
+
+class IsinWithArange:
+    params = [
+        [np.float64, np.int64, np.uint64, np.object],
+        [
+            1_000,
+            2_000,
+            8_000,
+        ],
+        [-2, 0, 2],
+    ]
+    param_names = ["dtype", "M", "offset_factor"]
+
+    def setup(self, dtype, M, offset_factor):
+        offset = int(M * offset_factor)
+        np.random.seed(42)
+        tmp = pd.Series(np.random.randint(offset, M + offset, 10 ** 6))
+        self.s = tmp.astype(dtype)
+        self.values = np.arange(M).astype(dtype)
+
+    def time_isin(self, dtype, M, offset_factor):
+        self.s.isin(self.values)
+
+
+class Float64GroupIndex:
+    # GH28303
+    def setup(self):
+        self.df = pd.date_range(
+            start="1/1/2018", end="1/2/2018", periods=1e6
+        ).to_frame()
+        self.group_index = np.round(self.df.index.astype(int) / 1e9)
+
+    def time_groupby(self):
+        self.df.groupby(self.group_index).last()
+
+
+class UniqueAndFactorizeArange:
+    params = range(4, 16)
+    param_names = ["exponent"]
+
+    def setup(self, exponent):
+        a = np.arange(10 ** 4, dtype="float64")
+        self.a2 = (a + 10 ** exponent).repeat(100)
+
+    def time_factorize(self, exponent):
+        pd.factorize(self.a2)
+
+    def time_unique(self, exponent):
+        pd.unique(self.a2)
+
+
+class NumericSeriesIndexing:
+
+    params = [
+        (pd.Int64Index, pd.UInt64Index, pd.Float64Index),
+        (10 ** 4, 10 ** 5, 5 * 10 ** 5, 10 ** 6, 5 * 10 ** 6),
+    ]
+    param_names = ["index_dtype", "N"]
+
+    def setup(self, index, N):
+        vals = np.array(list(range(55)) + [54] + list(range(55, N - 1)))
+        indices = index(vals)
+        self.data = pd.Series(np.arange(N), index=indices)
+
+    def time_loc_slice(self, index, N):
+        # trigger building of mapping
+        self.data.loc[:800]
+
+
+class NumericSeriesIndexingShuffled:
+
+    params = [
+        (pd.Int64Index, pd.UInt64Index, pd.Float64Index),
+        (10 ** 4, 10 ** 5, 5 * 10 ** 5, 10 ** 6, 5 * 10 ** 6),
+    ]
+    param_names = ["index_dtype", "N"]
+
+    def setup(self, index, N):
+        vals = np.array(list(range(55)) + [54] + list(range(55, N - 1)))
+        np.random.seed(42)
+        np.random.shuffle(vals)
+        indices = index(vals)
+        self.data = pd.Series(np.arange(N), index=indices)
+
+    def time_loc_slice(self, index, N):
+        # trigger building of mapping
+        self.data.loc[:800]

doc/source/whatsnew/v1.2.0.rst

@@ -424,6 +424,7 @@ Performance improvements
 - Reduced peak memory usage in :meth:`DataFrame.to_pickle` when using ``protocol=5`` in python 3.8+ (:issue:`34244`)
 - faster ``dir`` calls when many index labels, e.g. ``dir(ser)`` (:issue:`37450`)
 - Performance improvement in :class:`ExpandingGroupby` (:issue:`37064`)
+- Performance improvement in :meth:`pd.DataFrame.groupby` for ``float`` ``dtype`` (:issue:`28303`), changes of the underlying hash-function can lead to changes in float based indexes sort ordering for ties (e.g. :meth:`pd.Index.value_counts`)
 
 .. ---------------------------------------------------------------------------
 

pandas/_libs/src/klib/khash.h

@@ -143,10 +143,86 @@ typedef khint_t khiter_t;
 #define __ac_set_isboth_false(flag, i) __ac_set_isempty_false(flag, i)
 #define __ac_set_isdel_true(flag, i) ((void)0)
 
+
+// specializations of https://github.com/aappleby/smhasher/blob/master/src/MurmurHash2.cpp
+khint32_t PANDAS_INLINE murmur2_32to32(khint32_t k){
+    const khint32_t SEED = 0xc70f6907UL;
+    // 'm' and 'r' are mixing constants generated offline.
+    // They're not really 'magic', they just happen to work well.
+    const khint32_t M_32 = 0x5bd1e995;
+    const int R_32 = 24;
+
+    // Initialize the hash to a 'random' value
+    khint32_t h = SEED ^ 4;
+
+    //handle 4 bytes:
+    k *= M_32;
+    k ^= k >> R_32;
+    k *= M_32;
+
+    h *= M_32;
+    h ^= k;
+
+    // Do a few final mixes of the hash to ensure the "last few
+    // bytes" are well-incorporated. (Really needed here?)
+    h ^= h >> 13;
+    h *= M_32;
+    h ^= h >> 15;
+    return h;
+}
+
+// it is possible to have a special x64-version, which would need less operations, but
+// using 32bit version always has also some benifits:
+//    - one code for 32bit and 64bit builds
+//    - the same case for 32bit and 64bit builds
+//    - no performance difference could be measured compared to a possible x64-version
+
+khint32_t PANDAS_INLINE murmur2_32_32to32(khint32_t k1, khint32_t k2){
+    const khint32_t SEED = 0xc70f6907UL;
+    // 'm' and 'r' are mixing constants generated offline.
+    // They're not really 'magic', they just happen to work well.
+    const khint32_t M_32 = 0x5bd1e995;
+    const int R_32 = 24;
+
+    // Initialize the hash to a 'random' value
+    khint32_t h = SEED ^ 4;
+
+    //handle first 4 bytes:
+    k1 *= M_32;
+    k1 ^= k1 >> R_32;
+    k1 *= M_32;
+
+    h *= M_32;
+    h ^= k1;
+
+    //handle second 4 bytes:
+    k2 *= M_32;
+    k2 ^= k2 >> R_32;
+    k2 *= M_32;
+
+    h *= M_32;
+    h ^= k2;
+
+    // Do a few final mixes of the hash to ensure the "last few
+    // bytes" are well-incorporated.
+    h ^= h >> 13;
+    h *= M_32;
+    h ^= h >> 15;
+    return h;
+}
+
+khint32_t PANDAS_INLINE murmur2_64to32(khint64_t k){
+    khint32_t k1 = (khint32_t)k;
+    khint32_t k2 = (khint32_t)(k >> 32);
+
+    return murmur2_32_32to32(k1, k2);
+}
+
+
 #ifdef KHASH_LINEAR
 #define __ac_inc(k, m) 1
 #else
-#define __ac_inc(k, m) (((k)>>3 ^ (k)<<3) | 1) & (m)
+#define __ac_inc(k, m) (murmur2_32to32(k) | 1) & (m)
 #endif
 
 #define __ac_fsize(m) ((m) < 32? 1 : (m)>>5)

pandas/_libs/src/klib/khash_python.h

@@ -13,25 +13,31 @@
 // is 64 bits the truncation causes collission issues.  Given all that, we use our own
 // simple hash, viewing the double bytes as an int64 and using khash's default
 // hash for 64 bit integers.
-// GH 13436
+// GH 13436 showed that _Py_HashDouble doesn't work well with khash
+// GH 28303 showed, that the simple xoring-version isn't good enough
+// See GH 36729 for evaluation of the currently used murmur2-hash version
+
 khint64_t PANDAS_INLINE asint64(double key) {
-  khint64_t val;
-  memcpy(&val, &key, sizeof(double));
-  return val;
+    khint64_t val;
+    memcpy(&val, &key, sizeof(double));
+    return val;
 }
 
-// correct for all inputs but not -0.0 and NaNs
-#define kh_float64_hash_func_0_NAN(key) (khint32_t)((asint64(key))>>33^(asint64(key))^(asint64(key))<<11)
-
-// correct for all inputs but not NaNs
-#define kh_float64_hash_func_NAN(key) ((key) == 0.0 ?                       \
-                                        kh_float64_hash_func_0_NAN(0.0) : \
-                                        kh_float64_hash_func_0_NAN(key))
+#define ZERO_HASH 0
+#define NAN_HASH  0
 
-// correct for all
-#define kh_float64_hash_func(key) ((key) != (key) ?                       \
-                                   kh_float64_hash_func_NAN(Py_NAN) :     \
-                                   kh_float64_hash_func_NAN(key))
+khint32_t PANDAS_INLINE kh_float64_hash_func(double val){
+    // 0.0 and -0.0 should have the same hash:
+    if (val == 0.0){
+        return ZERO_HASH;
+    }
+    // all nans should have the same hash:
+    if ( val!=val ){
+        return NAN_HASH;
+    }
+    khint64_t as_int = asint64(val);
+    return murmur2_64to32(as_int);
+}
 
 #define kh_float64_hash_equal(a, b) ((a) == (b) || ((b) != (b) && (a) != (a)))
 

pandas/core/base.py

@@ -982,9 +982,9 @@ def value_counts(
         >>> index = pd.Index([3, 1, 2, 3, 4, np.nan])
         >>> index.value_counts()
         3.0    2
-        1.0    1
         2.0    1
         4.0    1
+        1.0    1
         dtype: int64
 
         With `normalize` set to `True`, returns the relative frequency by
@@ -993,9 +993,9 @@ def value_counts(
         >>> s = pd.Series([3, 1, 2, 3, 4, np.nan])
         >>> s.value_counts(normalize=True)
         3.0    0.4
-        1.0    0.2
         2.0    0.2
         4.0    0.2
+        1.0    0.2
         dtype: float64
 
         **bins**
@@ -1017,10 +1017,10 @@ def value_counts(
 
         >>> s.value_counts(dropna=False)
         3.0    2
-        1.0    1
         2.0    1
-        4.0    1
         NaN    1
+        4.0    1
+        1.0    1
         dtype: int64
         """
         result = value_counts(

pandas/tests/base/test_value_counts.py

@@ -232,18 +232,14 @@ def test_value_counts_datetime64(index_or_series):
 
     # with NaT
     s = df["dt"].copy()
-    s = klass(list(s.values) + [pd.NaT])
+    s = klass(list(s.values) + [pd.NaT] * 4)
 
     result = s.value_counts()
     assert result.index.dtype == "datetime64[ns]"
     tm.assert_series_equal(result, expected_s)
 
     result = s.value_counts(dropna=False)
-    # GH 35922. NaN-like now sorts to the beginning of duplicate counts
-    idx = pd.to_datetime(
-        ["2010-01-01 00:00:00", "2008-09-09 00:00:00", pd.NaT, "2009-01-01 00:00:00"]
-    )
-    expected_s = Series([3, 2, 1, 1], index=idx)
+    expected_s = pd.concat([Series([4], index=DatetimeIndex([pd.NaT])), expected_s])
     tm.assert_series_equal(result, expected_s)
 
     unique = s.unique()