diff --git a/PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs b/PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
index b17f7bb..5b9bf5c 100644
--- a/PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
+++ b/PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
@@ -68,18 +68,10 @@ internal static class PrimeTesterKernels
         results[index] = result;
     }
 
-    public static void HeuristicTrialDivisionKernel(Index1D index, ArrayView<ulong> divisors, ulong n, ArrayView<byte> resultFlag)
+    public static void HeuristicTrialDivisionKernel(Index1D index, ArrayView<ulong> divisors, ulong n, ArrayView<int> resultFlag)
     {
         ulong divisor = divisors[index];
-        if (divisor <= 1UL)
-        {
-            return;
-        }
-
-        if (n % divisor == 0UL)
-        {
-            resultFlag[0] = 1;
-        }
+        Atomic.Or(ref resultFlag[0], divisor <= 1UL ? 0 : (n % divisor == 0UL ? 1 : 0));
     }
 
     public static void SharesFactorKernel(Index1D index, ArrayView<ulong> numbers, ArrayView<byte> results)
diff --git a/PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs b/PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs
index 5aea0f8..72add2f 100644
--- a/PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs
+++ b/PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs
@@ -221,16 +221,16 @@ public sealed class HeuristicCombinedPrimeTester
                 int count = 0;
 
                 var inputView = gpu.Input.View;
-                var outputView = gpu.Output.View;
+                var flagView = gpu.HeuristicFlag.View;
 
                 bool ProcessBatch(int length)
                 {
                         inputView.CopyFromCPU(ref divisorArray[0], length);
-                        byte compositeFlag = 0;
-                        outputView.CopyFromCPU(ref compositeFlag, 1);
-                        kernel(length, inputView, n, outputView);
+                        int compositeFlag = 0;
+                        flagView.CopyFromCPU(ref compositeFlag, 1);
+                        kernel(length, inputView, n, flagView);
                         accelerator.Synchronize();
-                        outputView.CopyToCPU(ref compositeFlag, 1);
+                        flagView.CopyToCPU(ref compositeFlag, 1);
 
                         return compositeFlag != 0;
                 }
diff --git a/PerfectNumbers.Core/HeuristicPrimeTester.cs b/PerfectNumbers.Core/HeuristicPrimeTester.cs
index a2abb2f..5d91856 100644
--- a/PerfectNumbers.Core/HeuristicPrimeTester.cs
+++ b/PerfectNumbers.Core/HeuristicPrimeTester.cs
@@ -279,7 +279,7 @@ public sealed class HeuristicPrimeTester
         int count = 0;
 
         var inputView = gpu.Input.View;
-        var outputView = gpu.Output.View;
+        var flagView = gpu.HeuristicFlag.View;
 
         bool ProcessBatch(int length)
         {
@@ -289,11 +289,11 @@ public sealed class HeuristicPrimeTester
             }
 
             inputView.CopyFromCPU(ref divisorArray[0], length);
-            byte compositeFlag = 0;
-            outputView.CopyFromCPU(ref compositeFlag, 1);
-            kernel(length, inputView, n, outputView);
+            int compositeFlag = 0;
+            flagView.CopyFromCPU(ref compositeFlag, 1);
+            kernel(length, inputView, n, flagView);
             accelerator.Synchronize();
-            outputView.CopyToCPU(ref compositeFlag, 1);
+            flagView.CopyToCPU(ref compositeFlag, 1);
             return compositeFlag != 0;
         }
 
diff --git a/PerfectNumbers.Core/PrimeTester.cs b/PerfectNumbers.Core/PrimeTester.cs
index 3832125..f18ca71 100644
--- a/PerfectNumbers.Core/PrimeTester.cs
+++ b/PerfectNumbers.Core/PrimeTester.cs
@@ -272,7 +272,7 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
                         internal readonly Context AcceleratorContext;
                         public readonly Accelerator Accelerator;
                         public readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> Kernel;
-                        public readonly Action<Index1D, ArrayView<ulong>, ulong, ArrayView<byte>> HeuristicTrialDivisionKernel;
+                        public readonly Action<Index1D, ArrayView<ulong>, ulong, ArrayView<int>> HeuristicTrialDivisionKernel;
                         public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault;
                         public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
                         public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
@@ -285,6 +285,7 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
                         public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
                         public MemoryBuffer1D<ulong, Stride1D.Dense> Input;
                         public MemoryBuffer1D<byte, Stride1D.Dense> Output;
+                        public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag;
                         public int BufferCapacity;
 
                         internal PrimeTesterGpuContextLease(int minBufferCapacity)
@@ -292,7 +293,7 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
                                 AcceleratorContext = Context.CreateDefault();
                                 Accelerator = AcceleratorContext.GetPreferredDevice(false).CreateAccelerator(AcceleratorContext);
                                 Kernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
-                                HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ulong, ArrayView<byte>>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
+                                HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ulong, ArrayView<int>>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
 
                                 var primesDefault = DivisorGenerator.SmallPrimes;
                                 DevicePrimesDefault = Accelerator.Allocate1D<uint>(primesDefault.Length);
@@ -337,6 +338,7 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
                                 BufferCapacity = minBufferCapacity < 1 ? 1 : minBufferCapacity;
                                 Input = Accelerator.Allocate1D<ulong>(BufferCapacity);
                                 Output = Accelerator.Allocate1D<byte>(BufferCapacity);
+                                HeuristicFlag = Accelerator.Allocate1D<int>(BufferCapacity);
                                 _returned = false;
                         }
 
@@ -350,8 +352,10 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
 
                                 Input.Dispose();
                                 Output.Dispose();
+                                HeuristicFlag.Dispose();
                                 Input = Accelerator.Allocate1D<ulong>(required);
                                 Output = Accelerator.Allocate1D<byte>(required);
+                                HeuristicFlag = Accelerator.Allocate1D<int>(required);
                                 BufferCapacity = required;
                         }
 
@@ -375,6 +379,7 @@ public static bool IsPrimeCpu(ulong n, CancellationToken ct)
                         {
                                 Input.Dispose();
                                 Output.Dispose();
+                                HeuristicFlag.Dispose();
                                 DevicePrimesDefault.Dispose();
                                 DevicePrimesLastOne.Dispose();
                                 DevicePrimesLastSeven.Dispose();