youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

L4 is now on new algo, using new eviction arrays, using union of hlsf and cache line head.

Browse Source
This commit is contained in:
Youmu Liu
2026-04-23 17:39:38 +08:00
parent 5252be9436
commit 1a6ce24a27
5 changed files with 117 additions and 97 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Common/HighPerformance/Memory/L4.IntrinsicKernel.cs
+41 -29
View File
@@ -25,6 +25,7 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using StgSharp.Collections;
using System;
using System.Collections.Generic;
using System.Linq;
@@ -44,15 +45,12 @@ namespace StgSharp.HighPerformance.Memory
static IIntrinsicKernel Instance { get; } = Create();
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void AgeAll(
EvictionArray* eviction
);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public int ScanEvict(
byte* vPtr,
Vector128<byte> three
short* pPtr,
short* mPtr,
int count,
L4 owner
);
private static IIntrinsicKernel Create()
@@ -68,35 +66,49 @@ namespace StgSharp.HighPerformance.Memory
}
private struct SseKernel : IIntrinsicKernel
private readonly struct SseKernel : IIntrinsicKernel
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void AgeAll(
EvictionArray* eviction
public readonly int ScanEvict(
short* pPtr,
short* mPtr,
int count,
L4 owner
)
{
Vector128<byte> one = Vector128.Create((byte)0x01);
Vector128<byte> max = Vector128.Create((byte)0x03);
byte* p = (byte*)eviction;
for (int i = 0; i < CacheLineCount; i += 16)
int remain = count;
ConcurrentFlexibleArray<IL4Predict> predictors = owner._predictors;
predictors.RequestSpanOperation();
Span<IL4Predict> pSpan = predictors.AsSpan();
try
{
Vector128<byte> v = Sse2.LoadVector128(p + i);
v = Sse2.AddSaturate(v, one);
v = Sse2.Min(v, max);
Sse2.Store(p + i, v);
}
}
for (int i = 0; i < CacheLineCount; i += 8)
{
Vector128<short> p = Sse2.LoadVector128(pPtr + i);
Vector128<short> m = Sse2.LoadVector128(mPtr + i);
Vector128<short> great = Sse2.CompareGreaterThan(p, m);
Vector128<short> le = Sse42.Xor(great, Vector128.Create<short>(-1));
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public int ScanEvict(
byte* vPtr,
Vector128<byte> three
)
{
Vector128<byte> v = Sse2.LoadVector128(vPtr);
Vector128<byte> mask = Sse2.CompareEqual(v, three);
return Sse2.MoveMask(mask);
int mask = Sse42.MoveMask(le.AsByte());
while ((mask != 0) && (remain > 0))
{
int bit = BitOperations.TrailingZeroCount(mask);
int lane = bit >> 1;
if (owner.TryEvictAt(i + lane, pSpan)) {
remain--;
}
mask &= mask - 1;
mask &= mask - 1;
}
}
}
finally
{
predictors.ReleaseSpanOperation();
}
return remain;
}
}