Misaki/Misaki.HighPerformance
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor SPMD lane abstraction and add gather support

Browse Source 
- Rename ISPMD interfaces to ISPMDLane for clarity
- Add gather and mask load methods to ISPMDLane, implement for ScalarLane and WideLane
- Add GetUnsafePtr() for direct pointer access
- Update MathV and vector types to use new interface and gather methods
- Update SPMD job interfaces and implementations to ISPMDLane
- Improve hash codes, range checks, and safety checks in vector types
- Update codegen templates for new interface/methods
- Refactor SPMD jobs to use gather methods for efficient vectorized access
This commit is contained in:
Misaki
2026-04-25 11:50:51 +09:00
parent cfd01eb9b6
commit 9f7507ba71
18 changed files with 772 additions and 300 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
Misaki.HighPerformance.Test/Benchmark/GGXMipGenerationBenchmark.cs
View File

@@ -20,8 +20,8 @@ internal unsafe struct MipLevel
}
internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
where TFloat : unmanaged, ISPMD<TFloat, float>
where TInt : unmanaged, ISPMD<TInt, int>
where TFloat : unmanaged, ISPMDLane<TFloat, float>
where TInt : unmanaged, ISPMDLane<TInt, int>
{
public const uint SAMPLE_COUNT = 1024u;
@@ -104,34 +104,12 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
var py = (uv.y * (h - 1.0f)).Cast<TInt, int>();
// Clamp
px = TInt.Clamp(px, TInt.Zero, TInt.Create(w - 1));
py = TInt.Clamp(py, TInt.Zero, TInt.Create(h - 1));
px = TInt.Clamp(px, TInt.Zero, w - 1);
py = TInt.Clamp(py, TInt.Zero, h - 1);
// Assuming float RGB array format
var idx = (py * w + px) * 3;
var laneWidth = TFloat.LaneWidth;
var rBuffer = stackalloc float[laneWidth];
var gBuffer = stackalloc float[laneWidth];
var bBuffer = stackalloc float[laneWidth];
// Gather operation: extract scalar indices, perform random memory reads, and construct SoA buffers
for (var i = 0; i < laneWidth; i++)
{
var scalarIdx = idx[i];
rBuffer[i] = img[scalarIdx];
gBuffer[i] = img[scalarIdx + 1];
bBuffer[i] = img[scalarIdx + 2];
}
// Load the gathered contiguous arrays back into TLane types
var rLane = TFloat.Load(rBuffer);
var gLane = TFloat.Load(gBuffer);
var bLane = TFloat.Load(bBuffer);
return MathV.Create<TFloat, float>(rLane, gLane, bLane);
return MathV.GatherVector3<TFloat, float>(img, idx.GetUnsafePtr(), 1);
}
public void Execute(int loopIndex, ref readonly JobExecutionContext ctx)
@@ -179,9 +157,10 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
TFloat.Create(V.z)
);
var vPrefilteredColorX = TFloat.Zero;
var vPrefilteredColorY = TFloat.Zero;
var vPrefilteredColorZ = TFloat.Zero;
//var vPrefilteredColorX = TFloat.Zero;
//var vPrefilteredColorY = TFloat.Zero;
//var vPrefilteredColorZ = TFloat.Zero;
var vPrefilteredColor = Vector3<TFloat, float>.Zero;
var vTotalWeight = TFloat.Zero;
// 3. Monte Carlo Integration Loop
@@ -219,10 +198,7 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
var fireflyWeight = TFloat.One / (TFloat.One + luma);
var finalWeight = NdotL * fireflyWeight;
vPrefilteredColorX += sampleColor.x * finalWeight;
vPrefilteredColorY += sampleColor.y * finalWeight;
vPrefilteredColorZ += sampleColor.z * finalWeight;
vPrefilteredColor += sampleColor * finalWeight;
vTotalWeight += finalWeight;
}
@@ -231,9 +207,9 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
for (var i = 0; i < TFloat.LaneWidth; i++)
{
prefilteredColor.x += vPrefilteredColorX[i];
prefilteredColor.y += vPrefilteredColorY[i];
prefilteredColor.z += vPrefilteredColorZ[i];
prefilteredColor.x += vPrefilteredColor.x[i];
prefilteredColor.y += vPrefilteredColor.y[i];
prefilteredColor.z += vPrefilteredColor.z[i];
totalWeight += vTotalWeight[i];
}