[cfe-commits] [patch] change libc++'s string hash function to cityhash64
Howard Hinnant
hhinnant at apple.com
Fri Dec 9 06:43:58 PST 2011
Thanks! And thanks for updating CREDITS.TXT as well. I will review this patch, hopefully some time today.
Question regarding the llvm Developer's Policy:
http://llvm.org/docs/DeveloperPolicy.html#clp
I'm not a lawyer, nor even very knowledgeable about copyright issues. But my understanding is that CityHash is (opensource) copyrighted by Google. But I'm not seeing such a credit in CREDITS.TXT, nor do I know if such credit would be acceptable to Google.
Are you a representative of Google, and if so can you address these concerns?
Thanks,
Howard
On Dec 8, 2011, at 8:16 PM, Craig Silverstein wrote:
> Below is a patch that changes libc++'s hash<string> to use cityhash64
> (http://code.google.com/p/cityhash/) for machines where size_t is 64
> bits. I did not change the code where size_t is 32 bits; it will
> still use murmur2 for that. This is because cityhash64 needs a fast
> 64bit x 64bit multiply, which 32-bit systems are unlikely to have.
>
> I wrote code to test the change, based on Howard's code posted
> earlier. It's below for reference. Here is the speed improvement for
> using cityhash, on my linux desktop machine:
> model name : Intel(R) Core(TM)2 CPU 6600 @ 2.40GHz
> All statistics below are averaged over 3 runs.
>
> RUN TIME (compiled with -O2):
> old (murmur2): 0:09.20
> new (cityhash64): 0:08.60 // 6.5% faster
>
> This is corroborated by lower numbers for the collision rate, as
> emitted by the test program (smaller numbers are better):
>
> murmur2:
> score = 1.3934
> score = 1.42452
> score = 1.35706
> score = 1.47355
> score = 1.59226
> score = 1.4267
> score = 1.494
> score = 1.56537
> score = 1.63725
> score = 1.71011
>
> cityhash64:
> score = 1.38081
> score = 1.3898
> score = 1.2847
> score = 1.38114
> score = 1.47815
> score = 1.28579
> score = 1.33372
> score = 1.38128
> score = 1.42893
> score = 1.47695
>
> The downside: cityhash64 is a non-trivial amount of code (it has
> special-case code for different input sizes, for speed purposes).
> Here is the comparison of compile time and resulting binary size. The
> compile command I used was:
> clang++ -std=c++0x -stdlib=libc++ -isystem`pwd`/../include -L`pwd`/../lib -Wl,-rpath `pwd`/../lib /var/tmp/libcxx_test.cc -o /var/tmp/libcxx_test
>
> COMPILE TIME (no flags):
> old (murmur2): 0:08.82
> new (cityhash64): 0:08.95 // 1.5% slower
>
> SIZE (no flags):
> murmur2: 78121
> cityhash64: 83082 // 6.4% bigger
>
> SIZE (-g):
> murmur2: 377745
> cityhash64: 387554 // 2.6% bigger
>
> SIZE (-g -O2):
> murmur2: 326673
> cityhash64: 342554 // 4.9% bigger
>
> SIZE (-O2):
> murmur2: 31841
> cityhash64: 36082 // 13.3% bigger
>
> My viewpoint (for what it's worth), is that the speed increase is
> worth these costs. If the maintainers disagree, there are other
> points along the speed/size tradeoff, however, including removing the
> special-case code for small strings, or moving from cityhash64 to
> murmur64.
>
> Here is the test code:
> ---
> #include <unordered_set>
> #include <iostream>
> #include <fstream>
> #include <string>
> #include <vector>
>
> // Computes the average number of comparisions per lookup.
> // A perfect hash will return 1.
> template <class C>
> float
> grade(const C& c)
> {
> using namespace std;
> if (c.size() <= 1)
> return 100;
> float score = 0;
> size_t bc = c.bucket_count();
> for (size_t i = 0; i != bc; ++i)
> {
> size_t bs = c.bucket_size(i);
> score += bs * (bs+1) / 2;
> }
> return score / c.size();
> }
>
> int main()
> {
> using namespace std;
> typedef string T;
> vector<T> words;
> filebuf fb;
> fb.open("/usr/share/dict/words",ios::in);
> for (istream wordstream(&fb); wordstream; )
> {
> string word;
> wordstream >> word;
> words.push_back(word);
> }
> for (int repeat = 0; repeat < 200; ++repeat)
> {
> unordered_set<T> table;
> table.max_load_factor(1);
> // /usr/dict/words has 98569 words in it
> for (int i = 0; i < 10; ++i)
> {
> for (int j = 0; j < 9850; ++j)
> {
> table.insert(words[j*10+i]);
> }
> if (repeat == 0) cout << "score = " << grade(table) << '\n';
> }
> }
> }
> ---
>
> craig
> Index: include/memory
> ===================================================================
> --- include/memory (revision 146148)
> +++ include/memory (working copy)
> @@ -2719,18 +2719,22 @@
>
> template <class _Tp> struct hash;
>
> +// We use murmur2 when size_t is 32 bits, and cityhash64 when size_t
> +// is 64 bits. This is because cityhash64 uses 64bit x 64bit
> +// multiplication, which can be very slow on 32-bit systems.
> template <class _Size, size_t = sizeof(_Size)*__CHAR_BIT__>
> -struct __murmur2;
> +struct __murmur2_or_cityhash;
>
> template <class _Size>
> -struct __murmur2<_Size, 32>
> +struct __murmur2_or_cityhash<_Size, 32>
> {
> _Size operator()(const void* __key, _Size __len);
> };
>
> +// murmur2
> template <class _Size>
> _Size
> -__murmur2<_Size, 32>::operator()(const void* __key, _Size __len)
> +__murmur2_or_cityhash<_Size, 32>::operator()(const void* __key, _Size __len)
> {
> const _Size __m = 0x5bd1e995;
> const _Size __r = 24;
> @@ -2762,50 +2766,166 @@
> }
>
> template <class _Size>
> -struct __murmur2<_Size, 64>
> +struct __murmur2_or_cityhash<_Size, 64>
> {
> _Size operator()(const void* __key, _Size __len);
> +
> + private:
> + // Some primes between 2^63 and 2^64.
> + static const _Size __k0 = 0xc3a5c85c97cb3127ULL;
> + static const _Size __k1 = 0xb492b66fbe98f273ULL;
> + static const _Size __k2 = 0x9ae16a3b2f90404fULL;
> + static const _Size __k3 = 0xc949d7c7509e6557ULL;
> +
> + static _Size __rotate(_Size __val, int __shift) {
> + return __shift == 0 ? __val : ((__val >> __shift) | (__val << (64 - __shift)));
> + }
> +
> + static _Size __rotate_by_at_least_1(_Size __val, int __shift) {
> + return (__val >> __shift) | (__val << (64 - __shift));
> + }
> +
> + static _Size __shift_mix(_Size __val) {
> + return __val ^ (__val >> 47);
> + }
> +
> + static _Size __hash_len_16(_Size __u, _Size __v) {
> + const _Size __mul = 0x9ddfea08eb382d69ULL;
> + _Size __a = (__u ^ __v) * __mul;
> + __a ^= (__a >> 47);
> + _Size __b = (__v ^ __a) * __mul;
> + __b ^= (__b >> 47);
> + __b *= __mul;
> + return __b;
> + }
> +
> + static _Size __hash_len_0_to_16(const char* __s, _Size __len) {
> + if (__len > 8) {
> + const _Size __a = *(const _Size*)__s;
> + const _Size __b = *(const _Size*)(__s + __len - 8);
> + return __hash_len_16(__a, __rotate_by_at_least_1(__b + __len, __len)) ^ __b;
> + }
> + if (__len >= 4) {
> + const uint32_t __a = *(const uint32_t*)(__s);
> + const uint32_t __b = *(const uint32_t*)(__s + __len - 4);
> + return __hash_len_16(__len + (__a << 3), __b);
> + }
> + if (__len > 0) {
> + const unsigned char __a = __s[0];
> + const unsigned char __b = __s[__len >> 1];
> + const unsigned char __c = __s[__len - 1];
> + const uint32_t __y = static_cast<uint32_t>(__a) +
> + (static_cast<uint32_t>(__b) << 8);
> + const uint32_t __z = __len + (static_cast<uint32_t>(__c) << 2);
> + return __shift_mix(__y * __k2 ^ __z * __k3) * __k2;
> + }
> + return __k2;
> + }
> +
> + static _Size __hash_len_17_to_32(const char *__s, _Size __len) {
> + const _Size __a = *(const _Size*)(__s) * __k1;
> + const _Size __b = *(const _Size*)(__s + 8);
> + const _Size __c = *(const _Size*)(__s + __len - 8) * __k2;
> + const _Size __d = *(const _Size*)(__s + __len - 16) * __k0;
> + return __hash_len_16(__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<_Size, _Size> __weak_hash_len_32_with_seeds(
> + _Size __w, _Size __x, _Size __y, _Size __z, _Size __a, _Size __b) {
> + __a += __w;
> + __b = __rotate(__b + __a + __z, 21);
> + const _Size __c = __a;
> + __a += __x;
> + __a += __y;
> + __b += __rotate(__a, 44);
> + return pair<_Size, _Size>(__a + __z, __b + __c);
> + }
> +
> + // Return a 16-byte hash for s[0] ... s[31], a, and b. Quick and dirty.
> + static pair<_Size, _Size> __weak_hash_len_32_with_seeds(
> + const char* __s, _Size __a, _Size __b) {
> + return __weak_hash_len_32_with_seeds(*(const _Size*)(__s),
> + *(const _Size*)(__s + 8),
> + *(const _Size*)(__s + 16),
> + *(const _Size*)(__s + 24),
> + __a,
> + __b);
> + }
> +
> + // Return an 8-byte hash for 33 to 64 bytes.
> + static _Size __hash_len_33_to_64(const char *__s, size_t __len) {
> + _Size __z = *(const _Size*)(__s + 24);
> + _Size __a = *(const _Size*)(__s) +
> + (__len + *(const _Size*)(__s + __len - 16)) * __k0;
> + _Size __b = __rotate(__a + __z, 52);
> + _Size __c = __rotate(__a, 37);
> + __a += *(const _Size*)(__s + 8);
> + __c += __rotate(__a, 7);
> + __a += *(const _Size*)(__s + 16);
> + _Size __vf = __a + __z;
> + _Size __vs = __b + __rotate(__a, 31) + __c;
> + __a = *(const _Size*)(__s + 16) + *(const _Size*)(__s + __len - 32);
> + __z += *(const _Size*)(__s + __len - 8);
> + __b = __rotate(__a + __z, 52);
> + __c = __rotate(__a, 37);
> + __a += *(const _Size*)(__s + __len - 24);
> + __c += __rotate(__a, 7);
> + __a += *(const _Size*)(__s + __len - 16);
> + _Size __wf = __a + __z;
> + _Size __ws = __b + __rotate(__a, 31) + __c;
> + _Size __r = __shift_mix((__vf + __ws) * __k2 + (__wf + __vs) * __k0);
> + return __shift_mix(__r * __k0 + __vs) * __k2;
> + }
> };
>
> +// cityhash64
> template <class _Size>
> _Size
> -__murmur2<_Size, 64>::operator()(const void* __key, _Size __len)
> +__murmur2_or_cityhash<_Size, 64>::operator()(const void* __key, _Size __len)
> {
> - const _Size __m = 0xc6a4a7935bd1e995ull;
> - const _Size __r = 47;
> - _Size __h = __len * __m;
> - const unsigned char* __data = static_cast<const unsigned char*>(__key);
> - for (; __len >= 8; __data += 8, __len -= 8)
> - {
> - _Size __k = *(const _Size*)__data;
> - __k *= __m;
> - __k ^= __k >> __r;
> - __k *= __m;
> - __h ^= __k;
> - __h *= __m;
> + const char* __s = static_cast<const char*>(__key);
> + if (__len <= 32) {
> + if (__len <= 16) {
> + return __hash_len_0_to_16(__s, __len);
> + } else {
> + return __hash_len_17_to_32(__s, __len);
> }
> - switch (__len)
> - {
> - case 7:
> - __h ^= __data[6] << 48;
> - case 6:
> - __h ^= __data[5] << 40;
> - case 5:
> - __h ^= __data[4] << 32;
> - case 4:
> - __h ^= __data[3] << 24;
> - case 3:
> - __h ^= __data[2] << 16;
> - case 2:
> - __h ^= __data[1] << 8;
> - case 1:
> - __h ^= __data[0];
> - __h *= __m;
> - }
> - __h ^= __h >> __r;
> - __h *= __m;
> - __h ^= __h >> __r;
> - return __h;
> + } else if (__len <= 64) {
> + return __hash_len_33_to_64(__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.
> + _Size __x = *(const _Size*)(__s + __len - 40);
> + _Size __y = *(const _Size*)(__s + __len - 16) +
> + *(const _Size*)(__s + __len - 56);
> + _Size __z = __hash_len_16(*(const _Size*)(__s + __len - 48) + __len,
> + *(const _Size*)(__s + __len - 24));
> + pair<_Size, _Size> __v = __weak_hash_len_32_with_seeds(__s + __len - 64, __len, __z);
> + pair<_Size, _Size> __w = __weak_hash_len_32_with_seeds(__s + __len - 32, __y + __k1, __x);
> + __x = __x * __k1 + *(const _Size*)(__s);
> +
> + // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
> + __len = (__len - 1) & ~static_cast<_Size>(63);
> + do {
> + __x = __rotate(__x + __y + __v.first + *(const _Size*)(__s + 8), 37) * __k1;
> + __y = __rotate(__y + __v.second + *(const _Size*)(__s + 48), 42) * __k1;
> + __x ^= __w.second;
> + __y += __v.first + *(const _Size*)(__s + 40);
> + __z = __rotate(__z + __w.first, 33) * __k1;
> + __v = __weak_hash_len_32_with_seeds(__s, __v.second * __k1, __x + __w.first);
> + __w = __weak_hash_len_32_with_seeds(__s + 32, __z + __w.second,
> + __y + *(const _Size*)(__s + 16));
> + std::swap(__z, __x);
> + __s += 64;
> + __len -= 64;
> + } while (__len != 0);
> + return __hash_len_16(
> + __hash_len_16(__v.first, __w.first) + __shift_mix(__y) * __k1 + __z,
> + __hash_len_16(__v.second, __w.second) + __x);
> }
>
> template <class _Tp, size_t = sizeof(_Tp) / sizeof(size_t)>
> @@ -2863,7 +2983,7 @@
> };
> } __u;
> __u.__t = __v;
> - return __murmur2<size_t>()(&__u, sizeof(__u));
> + return __murmur2_or_cityhash<size_t>()(&__u, sizeof(__u));
> }
> };
>
> @@ -2885,7 +3005,7 @@
> };
> } __u;
> __u.__t = __v;
> - return __murmur2<size_t>()(&__u, sizeof(__u));
> + return __murmur2_or_cityhash<size_t>()(&__u, sizeof(__u));
> }
> };
>
> @@ -2908,7 +3028,7 @@
> };
> } __u;
> __u.__t = __v;
> - return __murmur2<size_t>()(&__u, sizeof(__u));
> + return __murmur2_or_cityhash<size_t>()(&__u, sizeof(__u));
> }
> };
>
> Index: include/string
> ===================================================================
> --- include/string (revision 146148)
> +++ include/string (working copy)
> @@ -3916,7 +3916,7 @@
> size_t _LIBCPP_INLINE_VISIBILITY __do_string_hash(_Ptr __p, _Ptr __e)
> {
> typedef typename iterator_traits<_Ptr>::value_type value_type;
> - return __murmur2<size_t>()(__p, (__e-__p)*sizeof(value_type));
> + return __murmur2_or_cityhash<size_t>()(__p, (__e-__p)*sizeof(value_type));
> }
>
> template<class _CharT, class _Traits, class _Allocator>
> Index: CREDITS.TXT
> ===================================================================
> --- CREDITS.TXT (revision 146148)
> +++ CREDITS.TXT (working copy)
> @@ -32,3 +32,7 @@
> N: Arvid Picciani
> E: aep at exys dot org
> D: Minor patches and musl port.
> +
> +N: Craig Silverstein
> +E: csilvers at google.com
> +D: Implemented Cityhash as the string hash function on 64-bit machines
> _______________________________________________
> cfe-commits mailing list
> cfe-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
More information about the cfe-commits
mailing list