add scheduler method; fixing buffer parallel errors

StgSharp.Mathematics/Numeric/MatrixParallel.cs

@@ -70,9 +70,11 @@ namespace StgSharp.Mathematics.Numeric
         {
             RemainThreadCount = _threads.Count;
             _AfterWork = (ThreadPriority)afterWork;
+            MatrixParallelThread leader = handle.GetLeader();
             foreach (MatrixParallelWrap w in handle.Wraps) {
                 w.SetScheduler();
             }
+            leader.MatrixParallelLeader();
         }
 
         internal static void Init()

StgSharp.Mathematics/Numeric/MatrixParallelHandle.cs

@@ -97,6 +97,13 @@ namespace StgSharp.Mathematics.Numeric
 
         private bool IsMultithread { get; init; }
 
+        public MatrixParallelThread DefineLeader()
+        {
+            MatrixParallelThread leader = _threads[1];
+            leader.Role = 2;
+            return leader;
+        }
+
         public void Dispose()
         {
             if (_disposed) {

StgSharp.Mathematics/Numeric/MatrixParallelThread.cs

@@ -37,7 +37,7 @@ using System.Threading.Tasks;
 
 namespace StgSharp.Mathematics.Numeric
 {
-    public class MatrixParallelThread : IDisposable
+    public partial class MatrixParallelThread : IDisposable
     {
 
         // Affinity binding info for this thread
@@ -58,6 +58,9 @@ namespace StgSharp.Mathematics.Numeric
             _resetEvent = new ManualResetEventSlim(false);
         }
 
+        // Global runtime switch: true => UMA/no-affinity; false => NUMA/affinity (auto-detected)
+        public static bool UseUmaAffinityPath { get; } = DetectUmaAffinityPath();
+
         // ~MatrixComputeThread()
         // {
         // Dispose(disposing: false);
@@ -135,6 +138,36 @@ namespace StgSharp.Mathematics.Numeric
             _resetEvent.Set();
         }
 
+        internal void MatrixParallelLeader()
+        {
+            Thread current = Thread.CurrentThread;
+            current.Priority = ThreadPriority.Highest;
+            Thread.Sleep(0);
+
+#if WINDOWS
+            if (!UseUmaAffinityPath)
+            {
+                // NUMA/affinity path: bind to this thread's reserved mask, then restore
+                nuint previousMask = SetThreadAffinityMask(GetCurrentThread(), _affinity.Mask);
+                bool affinityChanged = previousMask != 0; // nonzero => success
+                ExecuteTaskPackage(Tasks);
+                if (affinityChanged) {
+                    _ = SetThreadAffinityMask(GetCurrentThread(), previousMask);
+                }
+            } else
+            {
+                // UMA path: no affinity binding
+                ExecuteTaskPackage(Tasks);
+            }
+#else
+            // Non-Windows: no affinity control
+            ExecuteTaskPackage(Tasks);
+#endif
+
+            this.Role = 1;
+            current.Priority = ThreadPriority.Normal;
+        }
+
         protected virtual void Dispose(bool disposing)
         {
             if (!disposedValue)
@@ -152,6 +185,36 @@ namespace StgSharp.Mathematics.Numeric
             }
         }
 
+        private static bool DetectUmaAffinityPath()
+        {
+#if WINDOWS
+            try
+            {
+                // Quick heuristic: single CPU looks like UMA
+                if (Environment.ProcessorCount <= 1)
+                {
+                    return true;
+                }
+
+                // Query NUMA topology: highest node number 0 => single node => UMA
+                uint highest = 0;
+                if (GetNumaHighestNodeNumber(out highest)) {
+                    return highest == 0;
+                }
+
+                // If query fails, prefer UMA (safer: no affinity changes)
+                return true;
+            }
+            catch
+            {
+                return true;
+            }
+#else
+            // Non-Windows: treat as UMA (no affinity API available)
+            return true;
+#endif
+        }
+
         private unsafe void MatrixParallelWorker()
         {
             // Bind OS-thread affinity once when the worker starts running
@@ -312,6 +375,10 @@ namespace StgSharp.Mathematics.Numeric
         [DllImport("kernel32.dll")]
         private static extern nuint SetThreadAffinityMask(IntPtr hThread, nuint dwThreadAffinityMask);
 
+        [LibraryImport("kernel32.dll")]
+        [return: MarshalAs(UnmanagedType.Bool)]
+        private static partial bool GetNumaHighestNodeNumber(out uint HighestNodeNumber);
+
 #elif LINUX
 
 #endif

StgSharp.Mathematics/Numeric/MatrixParallelWrap.cs

@@ -69,8 +69,11 @@ namespace StgSharp.Mathematics.Numeric
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void SetAllWorker()
         {
-            for (int i = _begin; i < _begin + WrapTaskCapacity; i++) {
-                _threads[i].Restart();
+            for (int i = _begin; i < _begin + WrapTaskCapacity; i++)
+            {
+                if (_threads[i].Role != 3) {
+                    _threads[i].Restart();
+                }
             }
         }