Started converting the methods to parallel processing enabled kernels

SDFMapCreator.sln.DotSettings.user

@@ -1,4 +1,6 @@
 <wpf:ResourceDictionary xml:space="preserve" xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml" xmlns:s="clr-namespace:System;assembly=mscorlib" xmlns:ss="urn:shemas-jetbrains-com:settings-storage-xaml" xmlns:wpf="http://schemas.microsoft.com/winfx/2006/xaml/presentation">
+	<s:String x:Key="/Default/CodeInspection/ExcludedFiles/FilesAndFoldersToSkip2/=7020124F_002D9FFC_002D4AC3_002D8F3D_002DAAB8E0240759_002Ff_003AAccelerator_002EAllocations_002Ecs_002Fl_003AC_0021_003FUsers_003Fairon_003FAppData_003FRoaming_003FJetBrains_003FRider2024_002E3_003Fresharper_002Dhost_003FSourcesCache_003Fa3e5a0b5353a4a59d3be7bb386db3c46069739eacca1fc6b7323dca1ee7fd_003FAccelerator_002EAllocations_002Ecs/@EntryIndexedValue">ForceIncluded</s:String>
+	<s:String x:Key="/Default/CodeInspection/ExcludedFiles/FilesAndFoldersToSkip2/=7020124F_002D9FFC_002D4AC3_002D8F3D_002DAAB8E0240759_002Ff_003AEntryPointDescription_002Ecs_002Fl_003AC_0021_003FUsers_003Fairon_003FAppData_003FRoaming_003FJetBrains_003FRider2024_002E3_003Fresharper_002Dhost_003FSourcesCache_003F32b5380d8ca1aa7219c1622702a3e927b2bb32c9a53b43e12bb5a4af9a2862d_003FEntryPointDescription_002Ecs/@EntryIndexedValue">ForceIncluded</s:String>
 	<s:String x:Key="/Default/CodeInspection/ExcludedFiles/FilesAndFoldersToSkip2/=7020124F_002D9FFC_002D4AC3_002D8F3D_002DAAB8E0240759_002Ff_003AImage_007BTPixel_007D_002Ecs_002Fl_003AC_0021_003FUsers_003Fairon_003FAppData_003FRoaming_003FJetBrains_003FRider2024_002E3_003Fresharper_002Dhost_003FSourcesCache_003F98f0ece83ba33754ab932bd7b5c712d12f3a59029f9f14067f553a3a318c8f_003FImage_007BTPixel_007D_002Ecs/@EntryIndexedValue">ForceIncluded</s:String>
 	<s:String x:Key="/Default/CodeInspection/ExcludedFiles/FilesAndFoldersToSkip2/=7020124F_002D9FFC_002D4AC3_002D8F3D_002DAAB8E0240759_002Ff_003APixelTypeInfo_002Ecs_002Fl_003AC_0021_003FUsers_003Fairon_003FAppData_003FRoaming_003FJetBrains_003FRider2024_002E3_003Fresharper_002Dhost_003FSourcesCache_003Fc3cfdca1fb93eb6df5e51a81da5df646adfab8b862fd1a07ee5d247b49c5179_003FPixelTypeInfo_002Ecs/@EntryIndexedValue">ForceIncluded</s:String>
 	<s:String x:Key="/Default/CodeInspection/ExcludedFiles/FilesAndFoldersToSkip2/=7020124F_002D9FFC_002D4AC3_002D8F3D_002DAAB8E0240759_002Ff_003AThrowHelper_002Ecs_002Fl_003AC_0021_003FUsers_003Fairon_003FAppData_003FRoaming_003FJetBrains_003FRider2024_002E3_003Fresharper_002Dhost_003FSourcesCache_003F2c8e7ca976f350cba9836d5565dac56b11e0b56656fa786460eb1395857a6fa_003FThrowHelper_002Ecs/@EntryIndexedValue">ForceIncluded</s:String>

SDFMapCreator/Program.cs

@@ -16,7 +16,7 @@ public class Program {
     static List<TransitionMaskData> TransitionMasks = new();
     static List<SDFData> SDFs = new();
     static List<Vector3[,]> Gradients = new();
-
+    static readonly SdfKernels kernels = new();
     static void ConsoleUpdateLine(string s) => Console.Write("\r" + s);
 
     public static void Main(string[] args) {
@@ -48,12 +48,9 @@ public class Program {
             Environment.Exit(1);
         }
 
-        for (int i = 1; i < Images.Count; i++)
+        for (int i = 1; i < Images.Count; i++) {
-        {
+            for (int x = 0; x < Images[i].Width; x++) {
-            for (int x = 0; x < Images[i].Width; x++)
+                for (int y = 0; y < Images[i].Height; y++) {
-            {
-                for (int y = 0; y < Images[i].Height; y++)
-                {
                     Images[i].Pixels[x, y].X = MathF.Min(Images[i - 1].Pixels[x, y].X + Images[i].Pixels[x, y].X, MAX);
                     Images[i].Pixels[x, y].Y = MathF.Min(Images[i - 1].Pixels[x, y].Y + Images[i].Pixels[x, y].X, MAX);
                     Images[i].Pixels[x, y].Z = MathF.Min(Images[i - 1].Pixels[x, y].Z + Images[i].Pixels[x, y].X, MAX);
@@ -116,7 +113,8 @@ public class Program {
             for (var x = 0; x < mask.Mask.GetLength(0); x++) {
                 for (var y = 0; y < mask.Mask.GetLength(1); y++) {
                     if (gradient[x, y].X == 0) continue;
-                    var pixel = new Vector3(Remap(gradient[x, y].X, MIN, MAX, 1.0f-currStep, 1.0f-(currStep + stepIncrement)));
+                    var pixel = new Vector3(Remap(gradient[x, y].X, MIN, MAX, 1.0f - currStep,
+                        1.0f - (currStep + stepIncrement)));
                     if (pixel.X > finalImage[x, y].X) finalImage[x, y] = pixel;
                 }
             }
@@ -129,26 +127,15 @@ public class Program {
     private static void EdgeDetect(MaskData maskData) {
         uint width = (uint)maskData.Image.Width;
         uint height = (uint)maskData.Image.Height;
-        int iterCount = 0;
         var sw = new Stopwatch();
         sw.Start();
         Console.WriteLine("Running edge detection...");
-        Parallel.For(0, width * height, parallelOptions, (i) =>
+        kernels.EdgeDetect(maskData.Mask, out var temp);
-        {
+        maskData.Mask = temp;
-            int x = (int)(i % width);
+        Parallel.For(0, width * height, parallelOptions, (i) => {
-            int y = (int)(i / width);
+            if(maskData.Mask[i % width, i / width].Y != 0) lock (maskData.Edges) maskData.Edges.Add(new(i % width, i / width));
-
-            if (!EdgeKernel(maskData.Mask, x, y, width, height)) return;
-            var color = maskData.Mask[x, y];
-            color.Y = MAX;
-            maskData.Mask[x, y] = color;
-            lock (maskData.Edges) maskData.Edges.Add(new(x, y));
-            iterCount++;
-            if (iterCount % (width * height / 100) == 0) {
-                ConsoleUpdateLine(
-                    $"Progress: {iterCount / (float)(width * height):P} | {iterCount / (sw.Elapsed.TotalSeconds):N0} pixels/s");
-            }
         });
+        
         sw.Stop();
         Console.WriteLine(
             $"\nEdge pixels: {maskData.Edges.Count} | {maskData.Edges.Count / (sw.ElapsedMilliseconds + 1)} pixels/s\n Time: {sw.Elapsed.TotalSeconds:F4}s");
@@ -247,55 +234,16 @@ public class Program {
     private static float EuclideanDistance(Vector2 a, Vector2 b) =>
         MathF.Sqrt(MathF.Pow(a.X - b.X, 2) + MathF.Pow(a.Y - b.Y, 2));
     
-
     static Vector3[,] GetABMask(Vector3[,] A, Vector3[,] B, uint resX, uint resY) {
-        var temp = new Vector3[resX, resY];
+        kernels.ABMask(A, B, out var temp);
-        Parallel.For(0, resX * resY, parallelOptions, (i) =>
-        {
-            uint x = (uint)(i % resX);
-            uint y = (uint)(i / resX);
-            var pixelA = A[x, y];
-            var pixelB = B[x, y];
-            float lumaA = (pixelA.X + pixelA.Y + pixelA.Z) / 3;
-            float lumaB = (pixelB.X + pixelB.Y + pixelB.Z) / 3;
-            float resultPixel = lumaB > lumaA ? MAX : MIN;
-            temp[x, y] = new(resultPixel, 0, 0);
-        });
         return temp;
     }
 
     static Vector3[,] SelfMask(Vector3[,] A, uint resX, uint resY) {
-        var temp = new Vector3[resX, resY];
+        kernels.SelfMask(A, out var temp);
-        Parallel.For(0, resX * resY, parallelOptions, (i) =>
-        {
-            uint x = (uint)(i % resX);
-            uint y = (uint)(i / resX);
-            var pixelA = A[x, y];
-            float lumaA = (pixelA.X + pixelA.Y + pixelA.Z) / 3;
-            float resultPixel = lumaA > 0.99 ? MAX : MIN;
-            temp[x, y] = new(resultPixel, 0, 0);
-        });
         return temp;
     }
 
-    static bool EdgeKernel(Vector3[,] mask, int x, int y, uint width, uint height) {
-        //if we are already empty, return false
-        if (mask[x, y].X == 0) return false;
-        //if we are on the edge of the image, return false
-        if (x == 0 || y == 0 || x == width - 1 || y == height - 1) return false;
-        //check the 3x3 kernel
-        for (int xi = x - 1; xi <= x + 1; xi++) {
-            for (int yi = y - 1; yi <= y + 1; yi++) {
-                if (xi < 0 || xi >= width || yi < 0 || yi >= height)
-                    continue; //skip out of bounds pixels
-                if (mask[xi, yi].X == 0)
-                    return true; //if we find a black pixel, return true
-            }
-        }
-        //if we didn't find any black pixels, return false
-        return false;
-    }
-
     static T Lerp<T>(T a, T b, float t)
         where T : INumber<T>, IMultiplyOperators<T, float, T>, IAdditionOperators<T, T, T>
         => a * (1 - t) + b * t;

SDFMapCreator/Records.cs

@@ -3,6 +3,10 @@ using System.Numerics;
 namespace SDFMapCreator;
 
 public record Image(Vector3[,] Pixels, int Width, int Height);
-public record MaskData(Vector3[,] Mask, Image Image, List<Vector2> Edges);
+public record MaskData(Vector3[,] Mask, Image Image, List<Vector2> Edges) {
+    public Vector3[,] Mask { get; set; } = Mask;
+    public List<Vector2> Edges { get; set; } = Edges;
+}
+
 public record SDFData(Vector3[,] SDF);
 public record TransitionMaskData(Vector3[,] Mask, Image ImageA, Image ImageB);

SDFMapCreator/SDFMapCreator.csproj

@@ -8,6 +8,7 @@
     </PropertyGroup>
 
     <ItemGroup>
+      <PackageReference Include="ILGPU" Version="1.5.2" />
       <PackageReference Include="SixLabors.ImageSharp" Version="3.1.7" />
       <PackageReference Include="System.Numerics.Vectors" Version="4.6.1" />
     </ItemGroup>

SDFMapCreator/SdfKernels.Kernels.cs

@@ -0,0 +1,52 @@
+using System.Numerics;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace SDFMapCreator;
+
+public partial class SdfKernels {
+    private static void SelfMaskKernel(Index2D index, ArrayView2D<Vector3, Stride2D.DenseX> input, ArrayView2D<Vector3, Stride2D.DenseX> mask) {
+        int x = index.X;
+        int y = index.Y;
+        Vector3 value = input[x, y];
+        float lumaA = (value.X + value.Y + value.Z)/3f;
+        float r = lumaA > 0.99f ? 1f : 0f;
+        mask[x,y] = new (r, 0f, 0f);
+    }
+
+    private static void ABMaskKernel(Index2D index,
+        ArrayView2D<Vector3, Stride2D.DenseX> A,
+        ArrayView2D<Vector3, Stride2D.DenseX> B,
+        ArrayView2D<Vector3, Stride2D.DenseX> mask
+    ) {
+        int x = index.X;
+        int y = index.Y;
+        Vector3 valueA = A[x, y];
+        Vector3 valueB = B[x, y];
+        float lumaA = (valueA.X + valueA.Y + valueA.Z) / 3f;
+        float lumaB = (valueB.X + valueB.Y + valueB.Z) / 3f;
+        float r = lumaA > 0.99f && lumaB > 0.99f ? 1f : 0f;
+        mask[x, y] = new (r, 0f, 0f);
+    }
+
+    private static void EdgeKernel(Index2D index,
+        ArrayView2D<Vector3, Stride2D.DenseX> mask,
+        uint width, uint height
+    ) { // early exit if not on mask
+       if (mask[index].X == 0f) return;
+       int x = index.X;
+       int y = index.Y;
+       //if we are on the edge of the image, return false
+       if (x == 0 || y == 0 || x == width - 1 || y == height - 1) return;
+       
+       //check the 3x3 kernel
+       for (int xi = x - 1; xi <= x + 1; xi++) {
+           for (int yi = y - 1; yi <= y + 1; yi++) {
+               if (xi < 0 || xi >= width || yi < 0 || yi >= height)
+                   continue; //skip out of bounds pixels
+               if (mask[xi, yi].X == 0f)
+                   mask[index].Y = 1f; //if we find a black pixel, return true
+           }
+       }
+    }
+}

SDFMapCreator/SdfKernels.cs

@@ -0,0 +1,115 @@
+using System;
+using System.Numerics;
+using ILGPU;
+using ILGPU.Runtime;
+using ILGPU.Runtime.CPU;
+using ILGPU.Runtime.Cuda;
+using ILGPU.Runtime.OpenCL;
+
+namespace SDFMapCreator;
+
+public partial class SdfKernels {
+    private Context gpuContext;
+    
+    public SdfKernels() {
+        // Initialize the GPU context
+        gpuContext = Context.Create(builder => builder
+            .OpenCL()
+            .Cuda()
+            .CPU()
+        );
+
+        Console.WriteLine("Reading available accelerators (CUDA only)...");
+        foreach (var device in gpuContext.GetCudaDevices()) {
+            using Accelerator accelerator = device.CreateAccelerator(gpuContext);
+            Console.WriteLine($"{GetInfoString(accelerator)}");
+        }
+        Console.WriteLine("Reading available accelerators (OpenCL only)...");
+        foreach (var device in gpuContext.GetCLDevices()) {
+            using Accelerator accelerator = device.CreateAccelerator(gpuContext);
+            Console.WriteLine($"{GetInfoString(accelerator)}");
+        }
+        Console.WriteLine("Reading available accelerators (CPU only)...");
+        foreach (var device in gpuContext.GetCPUDevices()) {
+            using Accelerator accelerator = device.CreateAccelerator(gpuContext);
+            Console.WriteLine($"{GetInfoString(accelerator)}");
+        }
+        
+    }
+
+    public void SelfMask(Vector3[,] input, out Vector3[,] mask) {
+        var dev = gpuContext.GetPreferredDevice(preferCPU:false);
+        int width = input.GetLength(0);
+        int height = input.GetLength(1);
+        mask = new Vector3[width, height];
+        using Accelerator accelerator = dev.CreateAccelerator(gpuContext);
+        
+        using var buffer = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        buffer.CopyFromCPU(input);
+        
+        using var maskBuffer = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        
+        var selfMaskKernel = accelerator.LoadAutoGroupedStreamKernel<Index2D, ArrayView2D<Vector3, Stride2D.DenseX>,
+            ArrayView2D<Vector3, Stride2D.DenseX>>(SelfMaskKernel);
+
+        selfMaskKernel(new (width, height), buffer.View, maskBuffer.View);
+        
+        accelerator.Synchronize();
+
+        mask = maskBuffer.GetAsArray2D();
+    }
+    
+    public void ABMask(Vector3[,] A, Vector3[,] B, out Vector3[,] mask) {
+        var dev = gpuContext.GetPreferredDevice(preferCPU:false);
+        int width = A.GetLength(0);
+        int height = A.GetLength(1);
+        mask = new Vector3[width, height];
+        using Accelerator accelerator = dev.CreateAccelerator(gpuContext);
+        
+        using var bufferA = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        bufferA.CopyFromCPU(A);
+        
+        using var bufferB = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        bufferB.CopyFromCPU(B);
+        
+        using var maskBuffer = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        
+        var abMaskKernel = accelerator.LoadAutoGroupedStreamKernel<Index2D,
+            ArrayView2D<Vector3, Stride2D.DenseX>,
+            ArrayView2D<Vector3, Stride2D.DenseX>,
+            ArrayView2D<Vector3, Stride2D.DenseX>>(ABMaskKernel);
+
+        abMaskKernel(new (width, height), bufferA.View, bufferB.View, maskBuffer.View);
+        
+        accelerator.Synchronize();
+
+        mask = maskBuffer.GetAsArray2D();
+    }
+    
+    public void EdgeDetect(Vector3[,] mask, out Vector3[,] edge) {
+        var dev = gpuContext.GetPreferredDevice(preferCPU:false);
+        int width = mask.GetLength(0);
+        int height = mask.GetLength(1);
+        edge = new Vector3[width, height];
+        using Accelerator accelerator = dev.CreateAccelerator(gpuContext);
+        
+        using var buffer = accelerator.Allocate2DDenseX<Vector3>(new (width, height));
+        buffer.CopyFromCPU(mask);
+        
+        var edgeKernel = accelerator.LoadAutoGroupedStreamKernel<Index2D,
+            ArrayView2D<Vector3, Stride2D.DenseX>, uint, uint>(EdgeKernel);
+
+        edgeKernel(new (width, height), buffer.View, (uint)width, (uint)height);
+        
+        accelerator.Synchronize();
+
+        edge = buffer.GetAsArray2D();
+    }
+    
+    private static string GetInfoString(Accelerator a)
+    {
+        StringWriter infoString = new StringWriter();
+        a.PrintInformation(infoString);
+        return infoString.ToString();
+    }
+}