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

Refactor GGXMipGenerationJobSPMD for SPMD support

Browse Source 
Replaced struct with generic SPMD version for SIMD, added type aliases (commented), optimized RadicalInverse_VdC, and adjusted SampleEquirectangularMap for better performance and code separation.
This commit is contained in:
Misaki
2026-04-29 01:35:12 +09:00
parent 0acaf00767
commit b4535eff00
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
Misaki.HighPerformance.Test/Benchmark/GGXMipGenerationBenchmark.cs
View File

@@ -10,6 +10,9 @@ using System.Runtime.CompilerServices;
using System.Runtime.InteropServices; using System.Runtime.InteropServices;
using static Misaki.HighPerformance.Mathematics.math; using static Misaki.HighPerformance.Mathematics.math;
//using TFloat = Misaki.HighPerformance.Mathematics.SPMD.WideLane<float>;
//using TInt = Misaki.HighPerformance.Mathematics.SPMD.WideLane<int>;
namespace Misaki.HighPerformance.Test.Benchmark; namespace Misaki.HighPerformance.Test.Benchmark;
internal unsafe struct MipLevel internal unsafe struct MipLevel
@@ -21,6 +24,7 @@ internal unsafe struct MipLevel
public float roughness; public float roughness;
} }
//internal unsafe struct GGXMipGenerationJobSPMD : IJobParallelFor
internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
where TFloat : unmanaged, ISPMDLane<TFloat, float> where TFloat : unmanaged, ISPMDLane<TFloat, float>
where TInt : unmanaged, ISPMDLane<TInt, int> where TInt : unmanaged, ISPMDLane<TInt, int>
@@ -32,7 +36,7 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
public float* radicalInverse_VdCLut; public float* radicalInverse_VdCLut;
public int numMipLevels; public int numMipLevels;
[MethodImpl(MethodImplOptions.AggressiveInlining)] [MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public static float RadicalInverse_VdC(uint bits) public static float RadicalInverse_VdC(uint bits)
{ {
bits = (bits << 16) | (bits >> 16); bits = (bits << 16) | (bits >> 16);
@@ -96,6 +100,7 @@ internal unsafe struct GGXMipGenerationJobSPMD<TFloat, TInt> : IJobParallelFor
} }
// Samples the source HDR image using bilinear interpolation (simplified to nearest neighbor for brevity here) // Samples the source HDR image using bilinear interpolation (simplified to nearest neighbor for brevity here)
// Do not inline this function to avoid register pressure and code bloat in the main sampling loop, as this is a relatively heavy operation and we want to keep it separate for better instruction cache usage.
[MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.AggressiveOptimization)] [MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.AggressiveOptimization)]
private static Vector3<TFloat, float> SampleEquirectangularMap(float* img, int w, int h, Vector3<TFloat, float> dir) private static Vector3<TFloat, float> SampleEquirectangularMap(float* img, int w, int h, Vector3<TFloat, float> dir)
{ {