I\'ve read in other posts that this seems to be the best way to combine hash-values. Could somebody please break this down and explain why this is the best way to do it?
It's not the best, surprisingly to me it's not even particularily good. The main problem is the bad distribution, which is not really the fault of boost::hash_combine
in itself, but in conjunction with a badly distributing hash like std::hash
which is most commonly implemented with the identity function.
Figure 2: The effect of a single bit change in one of two random 32 bit numbers on the result of boost::hash_combine
To demonstrate how bad things can become these are the collisions for points on a 32x32 grid when using hash_combine
as intended, and with std::hash
:
# hash x₀ y₀ x₁ y₁ ...
3449074105 6 30 8 15
3449074104 6 31 8 16
3449074107 6 28 8 17
3449074106 6 29 8 18
3449074109 6 26 8 19
3449074108 6 27 8 20
3449074111 6 24 8 21
3449074110 6 25 8 22
For a well distributed hash there should be none, statistically. Using bit-rotations instead of bit-shifts and xor instead of addition one could easily create a similar hash_combine that preserves entropy better. But really what you should do is use a good hash function in the first place, then after that a simple xor is sufficient to combine the seed and the hash.
#include <limits>
#include <cstdint>
template<typename T>
T xorshift(const T& n,int i){
return n^(n>>i);
}
uint32_t distribute(const uint32_t& n){
uint32_t p = 0x55555555ul; // pattern of alternating 0 and 1
uint32_t c = 3423571495ul; // random uneven integer constant;
return c*xorshift(p*xorshift(n,16),16);
}
uint64_t hash(const uint64_t& n){
uint64_t p = 0x5555555555555555; // pattern of alternating 0 and 1
uint64_t c = 17316035218449499591ull;// random uneven integer constant;
return c*xorshift(p*xorshift(n,32),32);
}
// if c++20 rotl is not available:
template <typename T,typename S>
typename std::enable_if<std::is_unsigned<T>::value,T>::type
constexpr rotl(const T n, const S i){
const T m = (std::numeric_limits<T>::digits-1);
const T c = i&m;
return (n<<c)|(n>>((T(0)-c)&m)); // this is usually recognized by the compiler to mean rotation, also c++20 now gives us rotl directly
}
template <class T>
inline size_t hash_combine(std::size_t& seed, const T& v)
{
return rotl(seed,std::numeric_limits<size_t>::digits/3) ^ distribute(std::hash<T>(v));
}
The seed is rotated once before combining it to make the order in which the hash was computed relevant.
The hash_combine
from boost
needs two operations less, and more importantly no multiplications, in fact it's about 5x faster, but at about 2 cyles per hash on my machine the proposed solution is still very fast and pays off quickly when used for a hash table. There are 118 collisions on a 1024x1024 grid (vs. 982017 for boosts
hash_combine
+ std::hash
), about as many as expected for a well distributed hash function and that is all we can ask for.
Now even when used in conjunction with a good hash function boost::hash_combine
is not ideal. If all entropy is in the seed at some point some of it will get lost. There are 2948667289 distinct results of boost::hash_combine(x,0)
, but there should be 4294967296 .
In conclusion, they tried to create a hash function that does both, combining and cascading, and fast, but ended up with something that does both just good enough to not be recognised as bad immediately.
ROTL For VS studio (you can deduce ROTR easily). (This is actually in reply to @WolfgangBrehm.)
Reason: the standard trick to induce the compiler to emit ror and/or rol instructions gives an error in VS: error C4146: unary minus operator applied to unsigned type, result still unsigned.
So... I solved the compiler error by replacing (-c) by (T(0) -c) but that is not going to be optimized.
Adding (MS specific) specialisations solves that as inspection of the emitted optimised assembly will show.
#include <intrin.h> // and some more includes, see above...
template <typename T> // default template is not good for optimisation
typename std::enable_if<std::is_unsigned<T>::value, T>::type
constexpr rotl(const T n, const int i)
{
constexpr T m = (std::numeric_limits<T>::digits - 1);
const T c = i & m;
//return (n << c) | (n >> (-c) & m);
return (n << c) | (n >> (T(0) - c) & m);
}
template<>
inline uint32_t rotl(const uint32_t n, const int i)
{
constexpr int m = (std::numeric_limits<uint32_t>::digits - 1);
const int c = i & m;
return _rotl(n, c);
}
template<>
inline uchar rotl(const uchar n, const int i)
{
constexpr uchar m = (std::numeric_limits<uchar>::digits - 1);
const uchar c = i & m;
return _rotl8(n, c);
}
template<>
inline ushort rotl(const ushort n, const int i)
{
constexpr uchar m = (std::numeric_limits<ushort>::digits - 1);
const uchar c = i & m;
return _rotl16(n, c);
}
template<>
inline uint64_t rotl(const uint64_t n, const int i)
{
constexpr int m = (std::numeric_limits<uint64_t>::digits - 1);
const int c = i & m;
return _rotl64(n, c);
}
It being the "best" is argumentative.
It being "good", or even "very good", at least superficially, is easy.
seed ^= hasher(v) + 0x9e3779b9 + (seed<<6) + (seed>>2);
We'll presume seed
is a previous result of hasher
or this algorithm.
^=
means that the bits on the left and bits on the right all change the bits of the result.
hasher(v)
is presumed to be a decent hash on v
. But the rest is defence in case it isn't a decent hash.
0x9e3779b9
is a 32 bit value (it could be extended to 64 bit if size_t
was 64 bit arguably) that contains half 0s and half 1s. It is basically a random series of 0s and 1s done by approximating particular irrational constant as a base-2 fixed point value. This helps ensure that if the hasher returns bad values, we still get a smear of 1s and 0s in our output.
(seed<<6) + (seed>>2)
is a bit shuffle of the incoming seed.
Imagine the 0x
constant was missing. Imagine the hasher returns the constant 0x01000
for almost every v
passed in. Now, each bit of the seed is spread out over the next iteration of the hash, during which it is again spread out.
The seed ^= (seed<<6) + (seed>>2)
0x00001000
becomes 0x00041400
after one iteration. Then 0x00859500
. As you repeat the operation, any set bits are "smeared out" over the output bits. Eventually the right and left bits collide, and carry moves the set bit from "even locations" to "odd locations".
The bits dependent on the value of an input seed grows relatively fast and in complex ways as the combine operation recurses on the seed operation. Adding causes carries, which smear things even more. The 0x
constant adds a bunch of pseudo-random bits that make boring hash values occupy more than a few bits of the hash space after being combined.
It is asymmetric thanks to addition (combining the hashes of "dog"
and "god"
gives different results), it handles boring hash values (mapping characters to their ascii value, which only involves twiddling a handful of bits). And, it is reasonably fast.
Slower hash combines that are cryptographically strong can be better in other situations. I, naively, would presume that making the shifts be a combination of even and odd shifts might be a good idea (but maybe addition, which moves even bits from odd bits, makes that less of a problem: after 3 iterations, incoming lone seed bits will collide and add and cause a carry).
The downside to this kind of analysis is that it only takes one mistake to make a hash function really bad. Pointing out all the good things doesn't help that much. So another thing that makes it good now is that it is reasonably famous and in an open-source repository, and I haven't heard anyone point out why it is bad.