Use word-wise hash instead of FNV-1a for arrays of word-aligned scalars

Also special case int-aligned scalars on 64-bit machines.
On a 64-bit machine hashing an array of pointers is now ~5.95x faster
with LDC2 and ~8.54x faster with DMD, and hashing an array of ints is
~3.34x faster with LDC2 and ~8.12x faster with DMD.
This commit is contained in:
Nathan Sashihara 2019-02-12 09:33:48 -05:00
parent 0bef2ef76d
commit a786bdbec5

View File

@ -18,7 +18,7 @@ import tanya.meta.trait;
import tanya.range.primitive;
version (unittest) import tanya.test.stub;
private struct FNV
private struct Hasher
{
static if (size_t.sizeof == 4)
{
@ -50,11 +50,33 @@ private struct FNV
}
else static if (isScalarType!T || isPointer!T)
{
(() @trusted => add((cast(const ubyte*) &key)[0 .. T.sizeof]))();
// Treat as an array of words
static if (T.sizeof % size_t.sizeof == 0
&& T.alignof >= size_t.alignof)
alias CastT = size_t;
// (64-bit or 128-bit) Treat as an array of ints
else static if (T.sizeof % uint.sizeof == 0
&& T.alignof >= uint.alignof)
alias CastT = uint;
// Treat as an array of bytes
else
alias CastT = ubyte;
add((() @trusted => (cast(const CastT*) &key)[0 .. T.sizeof / CastT.sizeof])());
}
else static if (isArray!T && isScalarType!(ElementType!T))
{
add(cast(const ubyte[]) key);
// Treat as an array of words
static if (ElementType!T.sizeof % size_t.sizeof == 0
&& ElementType!T.alignof >= size_t.alignof)
alias CastT = size_t;
// (64-bit or 128-bit) Treat as an array of ints
else static if (ElementType!T.sizeof % uint.sizeof == 0
&& ElementType!T.alignof >= uint.alignof)
alias CastT = uint;
// Treat as an array of bytes
else
alias CastT = ubyte;
add(cast(const CastT[]) key);
}
else static if (is(T == typeof(null)))
{
@ -73,13 +95,166 @@ private struct FNV
}
}
void add(const ubyte[] key) @nogc nothrow pure @safe
void add(scope const ubyte[] key) @nogc nothrow pure @safe
{
// FNV-1a
foreach (c; key)
{
this.hash = (this.hash ^ c) * prime;
}
}
void add(scope const size_t[] key) @nogc nothrow pure @safe
{
static if (size_t.sizeof == 4)
{
// Partial MurmurHash3_x86_32 (no finalization)
enum uint c1 = 0xcc9e2d51;
enum uint c2 = 0x1b873593;
alias h1 = hash;
foreach (x; key)
{
auto k1 = x * c1;
k1 = (k1 << 15) | (k1 >> (32 - 15));
k1 *= c2;
h1 ^= k1;
h1 = (h1 << 13) | (h1 >> (32 - 13));
h1 = h1 * 5 + 0xe6546b64;
}
}
else static if (size_t.sizeof == 8)
{
// Partial 64-bit MurmurHash64A (no finalization)
alias h = hash;
enum ulong m = 0xc6a4a7935bd1e995UL;
foreach (x; key)
{
auto k = x * m;
k ^= k >>> 47;
k *= m;
h ^= k;
h *= m;
}
}
else static if (size_t.sizeof == 16)
{
// Partial MurmurHash3_x64_128 (no finalization)
// treating each size_t as a pair of ulong.
ulong h1 = cast(ulong) hash;
ulong h2 = cast(ulong) (hash >> 64);
enum ulong c1 = 0x87c37b91114253d5UL;
enum ulong c2 = 0x4cf5ad432745937fUL;
foreach (x; key)
{
auto k1 = cast(ulong) x;
auto k2 = cast(ulong) (x >> 64);
k1 *= c1; k1 = (k1 << 32) | (k1 >> (64 - 31)); k1 *= c2; h1 ^= k1;
h1 = (h1 << 27) | (h1 >> (64 - 27)); h1 += h2; h1 = h1*5+0x52dce729;
k2 *= c2; k2 = (k2 << 33) | (k2 >> (64 - 33)); k2 *= c1; h2 ^= k2;
h2 = (h2 << 31) | (h2 >> (64 - 31)); h2 += h1; h2 = h2*5+0x38495ab5;
}
hash = cast(size_t) h1 + ((cast(size_t) h2) << 64);
}
else
{
static assert(0, "Hash length must be either 32, 64, or 128 bits.");
}
}
static if (size_t.sizeof != uint.sizeof)
void add(scope const uint[] key) @nogc nothrow pure @trusted
{
static if (size_t.sizeof == 8)
{
// Partial 32-bit MurmurHash64B (no finalization)
enum uint m = 0x5bd1e995;
enum r = 24;
uint h1 = cast(uint) hash;
uint h2 = cast(uint) (hash >> 32);
const(uint)* data = key.ptr;
auto len = key.length;
for (; len >= 2; data += 2, len -= 2)
{
uint k1 = data[0];
k1 *= m; k1 ^= k1 >> r; k1 *= m;
h1 *= m; h1 ^= k1;
uint k2 = data[1];
k2 *= m; k2 ^= k2 >> r; k2 *= m;
h2 *= m; h2 ^= k2;
}
if (len)
{
uint k1 = data[0];
k1 *= m; k1 ^= k1 >> r; k1 *= m;
h1 *= m; h1 ^= k1;
}
hash = cast(ulong) h1 + ((cast(ulong) h2) << 32);
}
else static if (size_t.sizeof == 16)
{
// Partial MurmurHash3_x86_128 (no finalization)
enum uint c1 = 0x239b961b;
enum uint c2 = 0xab0e9789;
enum uint c3 = 0x38b34ae5;
enum uint c4 = 0xa1e38b93;
uint h1 = cast(uint) hash;
uint h2 = cast(uint) (hash >> 32);
uint h3 = cast(uint) (hash >> 64);
uint h4 = cast(uint) (hash >> 96);
const(uint)* data = key.ptr;
auto len = key.length;
for (; len >= 4; data += 4, len -= 4)
{
uint k1 = data[0];
uint k2 = data[1];
uint k3 = data[2];
uint k4 = data[3];
h1 = (h1 << 19) | (h1 >> (32 - 19)); h1 += h2; h1 = h1*5+0x561ccd1b;
k2 *= c2; k2 = (k2 << 16) | (k2 >> (32 - 16)); k2 *= c3; h2 ^= k2;
h2 = (h2 << 17) | (h2 >> (32 - 17)); h2 += h3; h2 = h2*5+0x0bcaa747;
k3 *= c3; k3 = (k3 << 17) | (k3 >> (32 - 17)); k3 *= c4; h3 ^= k3;
h3 = (h3 << 15) | (h3 >> (32 - 15)); h3 += h4; h3 = h3*5+0x96cd1c35;
k4 *= c4; k4 = (k4 << 18) | (k4 >> (32 - 18)); k4 *= c1; h4 ^= k4;
h4 = (h4 << 13) | (h4 >> (32 - 13)); h4 += h1; h4 = h4*5+0x32ac3b17;
}
uint k1, k2, k3;
switch (len) // 0, 1, 2, 3
{
case 3:
k3 = data[2];
k3 *= c3; k3 = (k3 << 17) | (k3 >> (32 - 17)); k3 *= c4; h3 ^= k3;
goto case;
case 2:
k2 = data[1];
k2 *= c2; k2 = (k2 << 16) | (k2 >> (32 - 16)); k2 *= c3; h2 ^= k2;
goto case;
case 1:
k1 = data[0];
k1 *= c1; k1 = (k1 << 15) | (k1 >> (32 - 15)); k1 *= c2; h1 ^= k1;
break;
}
hash = cast(size_t) h1 +
((cast(size_t) h2) << 32) +
((cast(size_t) h3) << 64) +
((cast(size_t) h4) << 96);
}
else
{
static assert(0, "Hash length must be either 32, 64, or 128 bits.");
}
}
}
/**
@ -98,7 +273,8 @@ private struct FNV
* For pointers and for scalar types implicitly convertible to `size_t` this
* is an identity operation (i.e. the value is cast to `size_t` and returned
* unaltered). Integer types wider than `size_t` are XOR folded down to
* `size_t`. Other scalar types use the FNV-1a (Fowler-Noll-Vo) hash function.
* `size_t`. Other scalar types use an architecture-dependent hash function
* based on their width and alignment.
* If the type provides a `toHash`-function, only `toHash()` is called and its
* result is returned.
*
@ -134,9 +310,9 @@ size_t hash(T)(auto ref T key)
}
else
{
FNV fnv;
fnv(key);
return fnv.hash;
Hasher hasher;
hasher(key);
return hasher.hash;
}
}
@ -199,12 +375,12 @@ version (unittest)
static if (size_t.sizeof == 4) @nogc nothrow pure @safe unittest
{
assert(hash(HashRange()) == 0x6222e842U);
assert(hash(ToHashRange()) == 1268118805U);
assert(hash(ToHashRange()) == 3371162643U);
}
static if (size_t.sizeof == 8) @nogc nothrow pure @safe unittest
{
assert(hash(HashRange()) == 0x08985907b541d342UL);
assert(hash(ToHashRange()) == 12161962213042174405UL);
assert(hash(ToHashRange()) == 2072958611659694473);
}
static if (size_t.sizeof == 4) @nogc nothrow pure @system unittest