add bench for google's cityhash function (64bit, http://code.google.com/p/cityhash/)

convenient graph of output can be found at http://www.enlightenment.org/~discomfitor/hash_bench.png


SVN revision: 58610
This commit is contained in:
Mike Blumenkrantz 2011-04-12 23:12:52 +00:00
parent 09c490369a
commit d422c0a88e
6 changed files with 495 additions and 4 deletions

View File

@ -270,6 +270,7 @@ AM_CONDITIONAL(EINA_ENABLE_BENCHMARK_E17, test "x${enable_benchmark_e17}" = "xye
### Checks for programs
AC_PROG_CC
AC_PROG_CXX
# pkg-config
PKG_PROG_PKG_CONFIG
@ -590,6 +591,10 @@ fi
EFL_CHECK_BENCHMARK([enable_benchmark="yes"], [enable_benchmark="no"])
EINA_BENCH_MODULE([glib], [${enable_benchmark}], [glib-2.0], [enable_benchmark_glib="yes"], [enable_benchmark_glib="no"])
if test -n "$CXX" && test "x$enable_benchmark" = "xyes" ; then
AC_DEFINE_UNQUOTED([CITYHASH_BENCH], [1], [enable bench tests for google hash method])
fi
AM_CONDITIONAL([CITYHASH_BENCH], [test -n "$CXX" && test "x$enable_benchmark" = "xyes"])
AC_SUBST(requirement_eina)

View File

@ -90,7 +90,7 @@ bench_PROGRAMS = eina_bench
eina_bench_SOURCES = \
eina_bench.c \
eina_bench_sort.c \
eina_bench_hash.c \
eina_bench_hash.c \
eina_bench_stringshare.c \
eina_bench_convert.c \
eina_bench_mempool.c \
@ -108,8 +108,19 @@ evas_object_list.c \
evas_stringshare.c \
eina_bench_quad.c
if CITYHASH_BENCH
noinst_LTLIBRARIES = libcity.la
libcity_la_SOURCES = city.cc
nodist_EXTRA_eina_bench_SOURCES = dummy.cc
eina_bench_LDADD = @GLIB_LIBS@ $(top_builddir)/src/lib/libeina.la libcity.la
else
eina_bench_LDADD = @GLIB_LIBS@ $(top_builddir)/src/lib/libeina.la
endif
endif
EXTRA_DIST = eina_bench.h \

View File

@ -0,0 +1,307 @@
// Copyright (c) 2011 Google, Inc.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
//
// CityHash Version 1, by Geoff Pike and Jyrki Alakuijala
//
// This file provides CityHash64() and related functions.
//
// It's probably possible to create even faster hash functions by
// writing a program that systematically explores some of the space of
// possible hash functions, by using SIMD instructions, or by
// compromising on hash quality.
#include "city.h"
#include <algorithm>
using namespace std;
#define UNALIGNED_LOAD64(p) (*(const uint64*)(p))
#define UNALIGNED_LOAD32(p) (*(const uint32*)(p))
#if !defined(LIKELY)
#if defined(__GNUC__)
#define LIKELY(x) (__builtin_expect(!!(x), 1))
#else
#define LIKELY(x) (x)
#endif
#endif
// Some primes between 2^63 and 2^64 for various uses.
static const uint64 k0 = 0xc3a5c85c97cb3127ULL;
static const uint64 k1 = 0xb492b66fbe98f273ULL;
static const uint64 k2 = 0x9ae16a3b2f90404fULL;
static const uint64 k3 = 0xc949d7c7509e6557ULL;
// Bitwise right rotate. Normally this will compile to a single
// instruction, especially if the shift is a manifest constant.
static uint64 Rotate(uint64 val, int shift) {
// Avoid shifting by 64: doing so yields an undefined result.
return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
}
// Equivalent to Rotate(), but requires the second arg to be non-zero.
// On x86-64, and probably others, it's possible for this to compile
// to a single instruction if both args are already in registers.
static uint64 RotateByAtLeast1(uint64 val, int shift) {
return (val >> shift) | (val << (64 - shift));
}
static uint64 ShiftMix(uint64 val) {
return val ^ (val >> 47);
}
static uint64 HashLen16(uint64 u, uint64 v) {
return Hash128to64(uint128(u, v));
}
static uint64 HashLen0to16(const char *s, size_t len) {
if (len > 8) {
uint64 a = UNALIGNED_LOAD64(s);
uint64 b = UNALIGNED_LOAD64(s + len - 8);
return HashLen16(a, RotateByAtLeast1(b + len, len)) ^ b;
}
if (len >= 4) {
uint64 a = UNALIGNED_LOAD32(s);
return HashLen16(len + (a << 3), UNALIGNED_LOAD32(s + len - 4));
}
if (len > 0) {
uint8 a = s[0];
uint8 b = s[len >> 1];
uint8 c = s[len - 1];
uint32 y = static_cast<uint32>(a) + (static_cast<uint32>(b) << 8);
uint32 z = len + (static_cast<uint32>(c) << 2);
return ShiftMix(y * k2 ^ z * k3) * k2;
}
return k2;
}
// This probably works well for 16-byte strings as well, but it may be overkill
// in that case.
static uint64 HashLen17to32(const char *s, size_t len) {
uint64 a = UNALIGNED_LOAD64(s) * k1;
uint64 b = UNALIGNED_LOAD64(s + 8);
uint64 c = UNALIGNED_LOAD64(s + len - 8) * k2;
uint64 d = UNALIGNED_LOAD64(s + len - 16) * k0;
return HashLen16(Rotate(a - b, 43) + Rotate(c, 30) + d,
a + Rotate(b ^ k3, 20) - c + len);
}
// Return a 16-byte hash for 48 bytes. Quick and dirty.
// Callers do best to use "random-looking" values for a and b.
static pair<uint64, uint64> WeakHashLen32WithSeeds(
uint64 w, uint64 x, uint64 y, uint64 z, uint64 a, uint64 b) {
a += w;
b = Rotate(b + a + z, 21);
uint64 c = a;
a += x;
a += y;
b += Rotate(a, 44);
return make_pair(a + z, b + c);
}
// Return a 16-byte hash for s[0] ... s[31], a, and b. Quick and dirty.
static pair<uint64, uint64> WeakHashLen32WithSeeds(
const char* s, uint64 a, uint64 b) {
return WeakHashLen32WithSeeds(UNALIGNED_LOAD64(s),
UNALIGNED_LOAD64(s + 8),
UNALIGNED_LOAD64(s + 16),
UNALIGNED_LOAD64(s + 24),
a,
b);
}
// Return an 8-byte hash for 33 to 64 bytes.
static uint64 HashLen33to64(const char *s, size_t len) {
uint64 z = UNALIGNED_LOAD64(s + 24);
uint64 a = UNALIGNED_LOAD64(s) + (len + UNALIGNED_LOAD64(s + len - 16)) * k0;
uint64 b = Rotate(a + z, 52);
uint64 c = Rotate(a, 37);
a += UNALIGNED_LOAD64(s + 8);
c += Rotate(a, 7);
a += UNALIGNED_LOAD64(s + 16);
uint64 vf = a + z;
uint64 vs = b + Rotate(a, 31) + c;
a = UNALIGNED_LOAD64(s + 16) + UNALIGNED_LOAD64(s + len - 32);
z = UNALIGNED_LOAD64(s + len - 8);
b = Rotate(a + z, 52);
c = Rotate(a, 37);
a += UNALIGNED_LOAD64(s + len - 24);
c += Rotate(a, 7);
a += UNALIGNED_LOAD64(s + len - 16);
uint64 wf = a + z;
uint64 ws = b + Rotate(a, 31) + c;
uint64 r = ShiftMix((vf + ws) * k2 + (wf + vs) * k0);
return ShiftMix(r * k0 + vs) * k2;
}
uint64 CityHash64(const char *s, size_t len) {
if (len <= 32) {
if (len <= 16) {
return HashLen0to16(s, len);
} else {
return HashLen17to32(s, len);
}
} else if (len <= 64) {
return HashLen33to64(s, len);
}
// For strings over 64 bytes we hash the end first, and then as we
// loop we keep 56 bytes of state: v, w, x, y, and z.
uint64 x = UNALIGNED_LOAD64(s);
uint64 y = UNALIGNED_LOAD64(s + len - 16) ^ k1;
uint64 z = UNALIGNED_LOAD64(s + len - 56) ^ k0;
pair<uint64, uint64> v = WeakHashLen32WithSeeds(s + len - 64, len, y);
pair<uint64, uint64> w = WeakHashLen32WithSeeds(s + len - 32, len * k1, k0);
z += ShiftMix(v.second) * k1;
x = Rotate(z + x, 39) * k1;
y = Rotate(y, 33) * k1;
// Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
len = (len - 1) & ~static_cast<size_t>(63);
do {
x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
x ^= w.second;
y ^= v.first;
z = Rotate(z ^ w.first, 33);
v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
std::swap(z, x);
s += 64;
len -= 64;
} while (len != 0);
return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
HashLen16(v.second, w.second) + x);
}
uint64 CityHash64WithSeed(const char *s, size_t len, uint64 seed) {
return CityHash64WithSeeds(s, len, k2, seed);
}
uint64 CityHash64WithSeeds(const char *s, size_t len,
uint64 seed0, uint64 seed1) {
return HashLen16(CityHash64(s, len) - seed0, seed1);
}
// A subroutine for CityHash128(). Returns a decent 128-bit hash for strings
// of any length representable in ssize_t. Based on City and Murmur.
static uint128 CityMurmur(const char *s, size_t len, uint128 seed) {
uint64 a = Uint128Low64(seed);
uint64 b = Uint128High64(seed);
uint64 c = 0;
uint64 d = 0;
ssize_t l = len - 16;
if (l <= 0) { // len <= 16
c = b * k1 + HashLen0to16(s, len);
d = Rotate(a + (len >= 8 ? UNALIGNED_LOAD64(s) : c), 32);
} else { // len > 16
c = HashLen16(UNALIGNED_LOAD64(s + len - 8) + k1, a);
d = HashLen16(b + len, c + UNALIGNED_LOAD64(s + len - 16));
a += d;
do {
a ^= ShiftMix(UNALIGNED_LOAD64(s) * k1) * k1;
a *= k1;
b ^= a;
c ^= ShiftMix(UNALIGNED_LOAD64(s + 8) * k1) * k1;
c *= k1;
d ^= c;
s += 16;
l -= 16;
} while (l > 0);
}
a = HashLen16(a, c);
b = HashLen16(d, b);
return uint128(a ^ b, HashLen16(b, a));
}
uint128 CityHash128WithSeed(const char *s, size_t len, uint128 seed) {
if (len < 128) {
return CityMurmur(s, len, seed);
}
// We expect len >= 128 to be the common case. Keep 56 bytes of state:
// v, w, x, y, and z.
pair<uint64, uint64> v, w;
uint64 x = Uint128Low64(seed);
uint64 y = Uint128High64(seed);
uint64 z = len * k1;
v.first = Rotate(y ^ k1, 49) * k1 + UNALIGNED_LOAD64(s);
v.second = Rotate(v.first, 42) * k1 + UNALIGNED_LOAD64(s + 8);
w.first = Rotate(y + z, 35) * k1 + x;
w.second = Rotate(x + UNALIGNED_LOAD64(s + 88), 53) * k1;
// This is the same inner loop as CityHash64(), manually unrolled.
do {
x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
x ^= w.second;
y ^= v.first;
z = Rotate(z ^ w.first, 33);
v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
std::swap(z, x);
s += 64;
x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
x ^= w.second;
y ^= v.first;
z = Rotate(z ^ w.first, 33);
v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
std::swap(z, x);
s += 64;
len -= 128;
} while (LIKELY(len >= 128));
y += Rotate(w.first, 37) * k0 + z;
x += Rotate(v.first + z, 49) * k0;
// If 0 < len < 128, hash up to 4 chunks of 32 bytes each from the end of s.
for (size_t tail_done = 0; tail_done < len; ) {
tail_done += 32;
y = Rotate(y - x, 42) * k0 + v.second;
w.first += UNALIGNED_LOAD64(s + len - tail_done + 16);
x = Rotate(x, 49) * k0 + w.first;
w.first += v.first;
v = WeakHashLen32WithSeeds(s + len - tail_done, v.first, v.second);
}
// At this point our 48 bytes of state should contain more than
// enough information for a strong 128-bit hash. We use two
// different 48-byte-to-8-byte hashes to get a 16-byte final result.
x = HashLen16(x, v.first);
y = HashLen16(y, w.first);
return uint128(HashLen16(x + v.second, w.second) + y,
HashLen16(x + w.second, y + v.second));
}
uint128 CityHash128(const char *s, size_t len) {
if (len >= 16) {
return CityHash128WithSeed(s + 16,
len - 16,
uint128(UNALIGNED_LOAD64(s) ^ k3,
UNALIGNED_LOAD64(s + 8)));
} else if (len >= 8) {
return CityHash128WithSeed(NULL,
0,
uint128(UNALIGNED_LOAD64(s) ^ (len * k0),
UNALIGNED_LOAD64(s + len - 8) ^ k1));
} else {
return CityHash128WithSeed(s, len, uint128(k0, k1));
}
}

View File

@ -0,0 +1,96 @@
// Copyright (c) 2011 Google, Inc.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
//
// CityHash Version 1, by Geoff Pike and Jyrki Alakuijala
//
// This file provides a few functions for hashing strings. On x86-64
// hardware in 2011, CityHash64() is faster than other high-quality
// hash functions, such as Murmur. This is largely due to higher
// instruction-level parallelism. CityHash64() and CityHash128() also perform
// well on hash-quality tests.
//
// CityHash128() is optimized for relatively long strings and returns
// a 128-bit hash. For strings more than about 2000 bytes it can be
// faster than CityHash64().
//
// Functions in the CityHash family are not suitable for cryptography.
//
// WARNING: This code has not been tested on big-endian platforms!
// It is known to work well on little-endian platforms that have a small penalty
// for unaligned reads, such as current Intel and AMD moderate-to-high-end CPUs.
//
// By the way, for some hash functions, given strings a and b, the hash
// of a+b is easily derived from the hashes of a and b. This property
// doesn't hold for any hash functions in this file.
#ifndef CITY_HASH_H_
#define CITY_HASH_H_
#include <stdlib.h> // for size_t.
#include <stdint.h>
#include <utility>
typedef uint8_t uint8;
typedef uint32_t uint32;
typedef uint64_t uint64;
typedef std::pair<uint64, uint64> uint128;
#ifdef __cplusplus
extern "C" {
#endif
inline uint64 Uint128Low64(const uint128& x) { return x.first; }
inline uint64 Uint128High64(const uint128& x) { return x.second; }
// Hash function for a byte array.
uint64 CityHash64(const char *buf, size_t len);
// Hash function for a byte array. For convenience, a 64-bit seed is also
// hashed into the result.
uint64 CityHash64WithSeed(const char *buf, size_t len, uint64 seed);
// Hash function for a byte array. For convenience, two seeds are also
// hashed into the result.
uint64 CityHash64WithSeeds(const char *buf, size_t len,
uint64 seed0, uint64 seed1);
// Hash function for a byte array.
uint128 CityHash128(const char *s, size_t len);
// Hash function for a byte array. For convenience, a 128-bit seed is also
// hashed into the result.
uint128 CityHash128WithSeed(const char *s, size_t len, uint128 seed);
// Hash 128 input bits down to 64 bits of output.
// This is intended to be a reasonably good hash function.
inline uint64 Hash128to64(const uint128& x) {
// Murmur-inspired hashing.
const uint64 kMul = 0x9ddfea08eb382d69ULL;
uint64 a = (Uint128Low64(x) ^ Uint128High64(x)) * kMul;
a ^= (a >> 47);
uint64 b = (Uint128High64(x) ^ a) * kMul;
b ^= (b >> 47);
b *= kMul;
return b;
}
#ifdef __cplusplus
}
#endif
#endif // CITY_HASH_H_

View File

@ -35,14 +35,14 @@ struct _Eina_Benchmark_Case
};
static const Eina_Benchmark_Case etc[] = {
/* { "Hash", eina_bench_hash }, */
{ "Hash", eina_bench_hash },
/* { "Array vs List vs Inlist", eina_bench_array }, */
/* { "Stringshare", eina_bench_stringshare }, */
/* { "Convert", eina_bench_convert }, */
/* { "Sort", eina_bench_sort }, */
/* { "Mempool", eina_bench_mempool }, */
/* { "Rectangle_Pool", eina_bench_rectangle_pool }, */
{ "Render Loop", eina_bench_quadtree },
// { "Render Loop", eina_bench_quadtree },
{ NULL, NULL }
};

View File

@ -38,6 +38,28 @@
#include "eina_rbtree.h"
#include "eina_convert.h"
#ifdef CITYHASH_BENCH
// Hash function for a byte array.
uint64_t CityHash64(const char *buf, size_t len);
static unsigned int
_eina_string_key_length(const char *key)
{
if (!key)
return 0;
return (int)strlen(key) + 1;
}
static int
_eina_string_key_cmp(const char *key1, __UNUSED__ int key1_length,
const char *key2, __UNUSED__ int key2_length)
{
return strcmp(key1, key2);
}
#endif
typedef struct _Eina_Bench_Rbtree Eina_Bench_Rbtree;
struct _Eina_Bench_Rbtree
{
@ -122,6 +144,51 @@ eina_bench_lookup_rbtree(int request)
eina_rbtree_delete(root, EINA_RBTREE_FREE_CB(_eina_bench_rbtree_free), NULL);
}
#ifdef CITYHASH_BENCH
static void
eina_bench_lookup_cityhash(int request)
{
Eina_Hash *hash = NULL;
int *tmp_val;
unsigned int i;
unsigned int j;
hash = eina_hash_new(EINA_KEY_LENGTH(_eina_string_key_length),
EINA_KEY_CMP(_eina_string_key_cmp),
EINA_KEY_HASH(CityHash64),
free,
8);
for (i = 0; i < (unsigned int)request; ++i)
{
char tmp_key[10];
tmp_val = malloc(sizeof (int));
if (!tmp_val)
continue;
eina_convert_itoa(i, tmp_key);
*tmp_val = i;
eina_hash_add(hash, tmp_key, tmp_val);
}
srand(time(NULL));
for (j = 0; j < 200; ++j)
for (i = 0; i < (unsigned int)request; ++i)
{
char tmp_key[10];
eina_convert_itoa(rand() % request, tmp_key);
tmp_val = eina_hash_find(hash, tmp_key);
}
eina_hash_free(hash);
}
#endif
static void
eina_bench_lookup_superfast(int request)
{
@ -311,7 +378,7 @@ eina_bench_lookup_evas(int request)
unsigned int i;
unsigned int j;
array = eina_array_new(1000);
array = eina_array_new(10000);
for (i = 0; i < (unsigned int)request; ++i)
{
@ -409,6 +476,11 @@ void eina_bench_hash(Eina_Benchmark *bench)
eina_benchmark_register(bench, "djb2-lookup-inline",
EINA_BENCHMARK(
eina_bench_lookup_djb2_inline), 10, 10000, 10);
#ifdef CITYHASH_BENCH
eina_benchmark_register(bench, "cityhash",
EINA_BENCHMARK(
eina_bench_lookup_cityhash), 10, 10000, 10);
#endif
eina_benchmark_register(bench, "rbtree",
EINA_BENCHMARK(
eina_bench_lookup_rbtree), 10, 10000, 10);