Inline xxHash64

This commit is contained in:
Oleksandr Melnyk 2022-06-11 17:21:10 +03:00
parent d7cf848606
commit 5d6d07a006
3 changed files with 315 additions and 212 deletions

View File

@ -0,0 +1,247 @@
/*
* This is the auto generated code.
* All function calls are inlined in XXH64
* Please don't try to analyze it.
*/
using System.Runtime.CompilerServices;
namespace Standart.Hash.xxHash;
public partial class xxHash64
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong __inline__XXH64(byte* input, int len, ulong seed)
{
ulong h64;
if (len >= 32)
{
byte* end = input + len;
byte* limit = end - 31;
ulong v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
ulong v2 = seed + XXH_PRIME64_2;
ulong v3 = seed + 0;
ulong v4 = seed - XXH_PRIME64_1;
do
{
// XXH64_round
v1 += *((ulong*) input) * XXH_PRIME64_2;
v1 = (v1 << 31) | (v1 >> (64 - 31));
v1 *= XXH_PRIME64_1;
input += 8;
// XXH64_round
v2 += *((ulong*) input) * XXH_PRIME64_2;
v2 = (v2 << 31) | (v2 >> (64 - 31));
v2 *= XXH_PRIME64_1;
input += 8;
// XXH64_round
v3 += *((ulong*) input) * XXH_PRIME64_2;
v3 = (v3 << 31) | (v3 >> (64 - 31));
v3 *= XXH_PRIME64_1;
input += 8;
// XXH64_round
v4 += *((ulong*) input) * XXH_PRIME64_2;
v4 = (v4 << 31) | (v4 >> (64 - 31));
v4 *= XXH_PRIME64_1;
input += 8;
} while (input < limit);
h64 = ((v1 << 1) | (v1 >> (64 - 1))) +
((v2 << 7) | (v2 >> (64 - 7))) +
((v3 << 12) | (v3 >> (64 - 12))) +
((v4 << 18) | (v4 >> (64 - 18)));
// XXH64_mergeRound
v1 *= XXH_PRIME64_2;
v1 = (v1 << 31) | (v1 >> (64 - 31));
v1 *= XXH_PRIME64_1;
h64 ^= v1;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v2 *= XXH_PRIME64_2;
v2 = (v2 << 31) | (v2 >> (64 - 31));
v2 *= XXH_PRIME64_1;
h64 ^= v2;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v3 *= XXH_PRIME64_2;
v3 = (v3 << 31) | (v3 >> (64 - 31));
v3 *= XXH_PRIME64_1;
h64 ^= v3;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v4 *= XXH_PRIME64_2;
v4 = (v4 << 31) | (v4 >> (64 - 31));
v4 *= XXH_PRIME64_1;
h64 ^= v4;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
}
else
{
h64 = seed + XXH_PRIME64_5;
}
h64 += (ulong) len;
// XXH64_finalize
len &= 31;
while (len >= 8) {
ulong k1 = XXH64_round(0, *(ulong*)input);
input += 8;
h64 ^= k1;
h64 = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
len -= 8;
}
if (len >= 4) {
h64 ^= *(uint*)input * XXH_PRIME64_1;
input += 4;
h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
len -= 4;
}
while (len > 0) {
h64 ^= (*input++) * XXH_PRIME64_5;
h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
--len;
}
// XXH64_avalanche
h64 ^= h64 >> 33;
h64 *= XXH_PRIME64_2;
h64 ^= h64 >> 29;
h64 *= XXH_PRIME64_3;
h64 ^= h64 >> 32;
return h64;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe void __inline__XXH64_stream_process(byte[] input, int len, ref ulong v1, ref ulong v2, ref ulong v3,
ref ulong v4)
{
fixed (byte* pData = &input[0])
{
byte* ptr = pData;
byte* limit = ptr + len;
do
{
// XXH64_round
v1 += *((ulong*) ptr) * XXH_PRIME64_2;
v1 = (v1 << 31) | (v1 >> (64 - 31));
v1 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v2 += *((ulong*) ptr) * XXH_PRIME64_2;
v2 = (v2 << 31) | (v2 >> (64 - 31));
v2 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v3 += *((ulong*) ptr) * XXH_PRIME64_2;
v3 = (v3 << 31) | (v3 >> (64 - 31));
v3 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v4 += *((ulong*) ptr) * XXH_PRIME64_2;
v4 = (v4 << 31) | (v4 >> (64 - 31));
v4 *= XXH_PRIME64_1;
ptr += 8;
} while (ptr < limit);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong __inline__XXH64_stream_finalize(byte[] input, int len, ref ulong v1, ref ulong v2, ref ulong v3,
ref ulong v4, long length, ulong seed)
{
fixed (byte* pData = &input[0])
{
byte* ptr = pData;
byte* end = pData + len;
ulong h64;
if (length >= 32)
{
h64 = ((v1 << 1) | (v1 >> (64 - 1))) +
((v2 << 7) | (v2 >> (64 - 7))) +
((v3 << 12) | (v3 >> (64 - 12))) +
((v4 << 18) | (v4 >> (64 - 18)));
// XXH64_mergeRound
v1 *= XXH_PRIME64_2;
v1 = (v1 << 31) | (v1 >> (64 - 31));
v1 *= XXH_PRIME64_1;
h64 ^= v1;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v2 *= XXH_PRIME64_2;
v2 = (v2 << 31) | (v2 >> (64 - 31));
v2 *= XXH_PRIME64_1;
h64 ^= v2;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v3 *= XXH_PRIME64_2;
v3 = (v3 << 31) | (v3 >> (64 - 31));
v3 *= XXH_PRIME64_1;
h64 ^= v3;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v4 *= XXH_PRIME64_2;
v4 = (v4 << 31) | (v4 >> (64 - 31));
v4 *= XXH_PRIME64_1;
h64 ^= v4;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
}
else
{
h64 = seed + XXH_PRIME64_5;
}
h64 += (ulong) length;
// XXH64_finalize
len &= 31;
while (len >= 8) {
ulong k1 = XXH64_round(0, *(ulong*)ptr);
ptr += 8;
h64 ^= k1;
h64 = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
len -= 8;
}
if (len >= 4) {
h64 ^= *(uint*)ptr * XXH_PRIME64_1;
ptr += 4;
h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
len -= 4;
}
while (len > 0) {
h64 ^= (*ptr++) * XXH_PRIME64_5;
h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
--len;
}
// XXH64_avalanche
h64 ^= h64 >> 33;
h64 *= XXH_PRIME64_2;
h64 ^= h64 >> 29;
h64 *= XXH_PRIME64_3;
h64 ^= h64 >> 32;
return h64;
}
}
}

View File

@ -7,15 +7,14 @@ namespace Standart.Hash.xxHash
public static partial class xxHash64 public static partial class xxHash64
{ {
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong XXH64_internal(byte* input, int len, ulong seed) private static unsafe ulong XXH64(byte* input, int len, ulong seed)
{ {
byte* end = input + len;
ulong h64; ulong h64;
if (len >= 32) if (len >= 32)
{ {
byte* limit = end - 32; byte* end = input + len;
byte* limit = end - 31;
ulong v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2; ulong v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
ulong v2 = seed + XXH_PRIME64_2; ulong v2 = seed + XXH_PRIME64_2;
@ -24,64 +23,21 @@ namespace Standart.Hash.xxHash
do do
{ {
// XXH64_round v1 = XXH64_round(v1, *(ulong*) input); input += 8;
v1 += *((ulong*)input) * XXH_PRIME64_2; v2 = XXH64_round(v2, *(ulong*) input); input += 8;
v1 = XXH_rotl64(v1, 31); v3 = XXH64_round(v3, *(ulong*) input); input += 8;
v1 *= XXH_PRIME64_1; v4 = XXH64_round(v4, *(ulong*) input); input += 8;
input += 8; } while (input < limit);
// XXH64_round
v2 += *((ulong*)input) * XXH_PRIME64_2;
v2 = XXH_rotl64(v2, 31);
v2 *= XXH_PRIME64_1;
input += 8;
// XXH64_round
v3 += *((ulong*)input) * XXH_PRIME64_2;
v3 = XXH_rotl64(v3, 31);
v3 *= XXH_PRIME64_1;
input += 8;
// XXH64_round
v4 += *((ulong*)input) * XXH_PRIME64_2;
v4 = XXH_rotl64(v4, 31);
v4 *= XXH_PRIME64_1;
input += 8;
} while (input <= limit);
h64 = XXH_rotl64(v1, 1) + h64 = XXH_rotl64(v1, 1) +
XXH_rotl64(v2, 7) + XXH_rotl64(v2, 7) +
XXH_rotl64(v3, 12) + XXH_rotl64(v3, 12) +
XXH_rotl64(v4, 18); XXH_rotl64(v4, 18);
// XXH64_mergeRound h64 = XXH64_mergeRound(h64, v1);
v1 *= XXH_PRIME64_2; h64 = XXH64_mergeRound(h64, v2);
v1 = XXH_rotl64(v1, 31); h64 = XXH64_mergeRound(h64, v3);
v1 *= XXH_PRIME64_1; h64 = XXH64_mergeRound(h64, v4);
h64 ^= v1;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v2 *= XXH_PRIME64_2;
v2 = XXH_rotl64(v2, 31);
v2 *= XXH_PRIME64_1;
h64 ^= v2;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v3 *= XXH_PRIME64_2;
v3 = XXH_rotl64(v3, 31);
v3 *= XXH_PRIME64_1;
h64 ^= v3;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v4 *= XXH_PRIME64_2;
v4 = XXH_rotl64(v4, 31);
v4 *= XXH_PRIME64_1;
h64 ^= v4;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
} }
else else
{ {
@ -90,164 +46,61 @@ namespace Standart.Hash.xxHash
h64 += (ulong)len; h64 += (ulong)len;
// XXH64_finalize return XXH64_finalize(h64, input, len);
while (input <= end - 8)
{
ulong t1 = *((ulong*)input) * XXH_PRIME64_2;
t1 = XXH_rotl64(t1, 31);
t1 *= XXH_PRIME64_1;
h64 ^= t1;
h64 = XXH_rotl64(h64, 27) * XXH_PRIME64_1 + XXH_PRIME64_4;
input += 8;
}
if (input <= end - 4)
{
h64 ^= *((uint*)input) * XXH_PRIME64_1;
h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
input += 4;
}
while (input < end)
{
h64 ^= *((byte*)input) * XXH_PRIME64_5;
h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
input += 1;
}
// XXH64_avalanche
h64 ^= h64 >> 33;
h64 *= XXH_PRIME64_2;
h64 ^= h64 >> 29;
h64 *= XXH_PRIME64_3;
h64 ^= h64 >> 32;
return h64;
} }
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe void __XXH64_stream_align(byte[] input, int len, ref ulong v1, ref ulong v2, ref ulong v3, ref ulong v4) private static ulong XXH64_round(ulong acc, ulong input)
{ {
fixed (byte* pData = &input[0]) acc += input * XXH_PRIME64_2;
{ acc = XXH_rotl64(acc, 31);
byte* ptr = pData; acc *= XXH_PRIME64_1;
byte* limit = ptr + len; return acc;
do
{
// XXH64_round
v1 += *((ulong*)ptr) * XXH_PRIME64_2;
v1 = XXH_rotl64(v1, 31);
v1 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v2 += *((ulong*)ptr) * XXH_PRIME64_2;
v2 = XXH_rotl64(v2, 31);
v2 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v3 += *((ulong*)ptr) * XXH_PRIME64_2;
v3 = XXH_rotl64(v3, 31);
v3 *= XXH_PRIME64_1;
ptr += 8;
// XXH64_round
v4 += *((ulong*)ptr) * XXH_PRIME64_2;
v4 = XXH_rotl64(v4, 31);
v4 *= XXH_PRIME64_1;
ptr += 8;
} while (ptr < limit);
}
} }
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong __XXH64_stream_finalize(byte[] input, int len, ref ulong v1, ref ulong v2, ref ulong v3, ref ulong v4, long length, ulong seed) private static ulong XXH64_mergeRound(ulong acc, ulong val)
{ {
fixed (byte* pData = &input[0]) val = XXH64_round(0, val);
{ acc ^= val;
byte* ptr = pData; acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
byte* end = pData + len; return acc;
ulong h64;
if (length >= 32)
{
h64 = XXH_rotl64(v1, 1) +
XXH_rotl64(v2, 7) +
XXH_rotl64(v3, 12) +
XXH_rotl64(v4, 18);
// XXH64_mergeRound
v1 *= XXH_PRIME64_2;
v1 = XXH_rotl64(v1, 31);
v1 *= XXH_PRIME64_1;
h64 ^= v1;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v2 *= XXH_PRIME64_2;
v2 = XXH_rotl64(v2, 31);
v2 *= XXH_PRIME64_1;
h64 ^= v2;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v3 *= XXH_PRIME64_2;
v3 = XXH_rotl64(v3, 31);
v3 *= XXH_PRIME64_1;
h64 ^= v3;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
// XXH64_mergeRound
v4 *= XXH_PRIME64_2;
v4 = XXH_rotl64(v4, 31);
v4 *= XXH_PRIME64_1;
h64 ^= v4;
h64 = h64 * XXH_PRIME64_1 + XXH_PRIME64_4;
}
else
{
h64 = seed + XXH_PRIME64_5;
} }
h64 += (ulong)length; [MethodImpl(MethodImplOptions.AggressiveInlining)]
private static ulong XXH64_avalanche(ulong hash)
// XXH64_finalize
while (ptr <= end - 8)
{ {
ulong t1 = *((ulong*)ptr) * XXH_PRIME64_2; hash ^= hash >> 33;
t1 = XXH_rotl64(t1, 31); hash *= XXH_PRIME64_2;
t1 *= XXH_PRIME64_1; hash ^= hash >> 29;
h64 ^= t1; hash *= XXH_PRIME64_3;
h64 = XXH_rotl64(h64, 27) * XXH_PRIME64_1 + XXH_PRIME64_4; hash ^= hash >> 32;
return hash;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong XXH64_finalize(ulong hash, byte* ptr, int len)
{
len &= 31;
while (len >= 8) {
ulong k1 = XXH64_round(0, *(ulong*)ptr);
ptr += 8; ptr += 8;
hash ^= k1;
hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
len -= 8;
} }
if (len >= 4) {
if (ptr <= end - 4) hash ^= *(uint*)ptr * XXH_PRIME64_1;
{
h64 ^= *((uint*)ptr) * XXH_PRIME64_1;
h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
ptr += 4; ptr += 4;
hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
len -= 4;
} }
while (len > 0) {
while (ptr < end) hash ^= (*ptr++) * XXH_PRIME64_5;
{ hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
h64 ^= *((byte*)ptr) * XXH_PRIME64_5; --len;
h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
ptr += 1;
}
// XXH64_avalanche
h64 ^= h64 >> 33;
h64 *= XXH_PRIME64_2;
h64 ^= h64 >> 29;
h64 *= XXH_PRIME64_3;
h64 ^= h64 >> 32;
return h64;
} }
return XXH64_avalanche(hash);
} }
} }
} }

View File

@ -119,7 +119,7 @@ public static partial class xxHash64
int l = offset - r; // length int l = offset - r; // length
// Process the next chunk // Process the next chunk
__XXH64_stream_align(buffer, l, ref v1, ref v2, ref v3, ref v4); __inline__XXH64_stream_process(buffer, l, ref v1, ref v2, ref v3, ref v4);
// Put remaining bytes to buffer // Put remaining bytes to buffer
Utils.BlockCopy(buffer, l, buffer, 0, r); Utils.BlockCopy(buffer, l, buffer, 0, r);
@ -127,7 +127,7 @@ public static partial class xxHash64
} }
// Process the final chunk // Process the final chunk
ulong h64 = __XXH64_stream_finalize(buffer, offset, ref v1, ref v2, ref v3, ref v4, length, seed); ulong h64 = __inline__XXH64_stream_finalize(buffer, offset, ref v1, ref v2, ref v3, ref v4, length, seed);
return h64; return h64;
} }
@ -215,7 +215,7 @@ public static partial class xxHash64
int l = offset - r; // length int l = offset - r; // length
// Process the next chunk // Process the next chunk
__XXH64_stream_align(buffer, l, ref v1, ref v2, ref v3, ref v4); __inline__XXH64_stream_process(buffer, l, ref v1, ref v2, ref v3, ref v4);
// Put remaining bytes to buffer // Put remaining bytes to buffer
Utils.BlockCopy(buffer, l, buffer, 0, r); Utils.BlockCopy(buffer, l, buffer, 0, r);
@ -223,7 +223,7 @@ public static partial class xxHash64
} }
// Process the final chunk // Process the final chunk
ulong h64 = __XXH64_stream_finalize(buffer, offset, ref v1, ref v2, ref v3, ref v4, length, seed); ulong h64 = __inline__XXH64_stream_finalize(buffer, offset, ref v1, ref v2, ref v3, ref v4, length, seed);
return h64; return h64;
} }
@ -256,6 +256,9 @@ public static partial class xxHash64
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining)]
private static unsafe ulong UnsafeComputeHash(byte* ptr, int length, ulong seed) private static unsafe ulong UnsafeComputeHash(byte* ptr, int length, ulong seed)
{ {
return XXH64_internal(ptr, length, seed); // Use inlined version
// return XXH64(ptr, length, seed);
return __inline__XXH64(ptr, length, seed);
} }
} }