From daf9076e54bfbe16fb2170a85e1e1a835e3b8149 Mon Sep 17 00:00:00 2001
From: Youmu Liu <105042599+Nitload-NSI@users.noreply.github.com>
Date: Tue, 28 Oct 2025 20:44:38 +0800
Subject: [PATCH] daily

---
 StgSharp.Common/HighPerformance/M64.cs        |  6 --
 StgSharp.Common/Mathematics/Linear/Matrix.cs  |  4 +-
 .../Numeric/MatrixCompute.Basic.cs            | 57 +++++++++++++++----
 3 files changed, 49 insertions(+), 18 deletions(-)

diff --git a/StgSharp.Common/HighPerformance/M64.cs b/StgSharp.Common/HighPerformance/M64.cs
index 5b6a1798..19412588 100644
--- a/StgSharp.Common/HighPerformance/M64.cs
+++ b/StgSharp.Common/HighPerformance/M64.cs
@@ -25,15 +25,9 @@
 //     
 // -----------------------------------------------------------------------
 // -----------------------------------------------------------------------
-using System;
-using System.Collections.Generic;
-using System.Drawing;
-using System.Linq;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
-using System.Text;
-using System.Threading.Tasks;
 
 namespace StgSharp.HighPerformance
 {
diff --git a/StgSharp.Common/Mathematics/Linear/Matrix.cs b/StgSharp.Common/Mathematics/Linear/Matrix.cs
index 53feeb04..62868022 100644
--- a/StgSharp.Common/Mathematics/Linear/Matrix.cs
+++ b/StgSharp.Common/Mathematics/Linear/Matrix.cs
@@ -167,7 +167,7 @@ namespace StgSharp.Mathematics.Numeric
         {
 
             private readonly MatrixKernel<T>* _source = enumeration._source;
-            private int _indexCount = 0;
+            private int _indexCount = -1;
             private readonly int _stride = enumeration.PrimStride;
             private readonly long _bound = (long)enumeration.PrimSize * enumeration.PrimStride;
             private long _index = (long)enumeration.PrimOffset * enumeration.PrimStride;
@@ -230,7 +230,7 @@ namespace StgSharp.Mathematics.Numeric
 
             private readonly int _end = segment._size;
 
-            private int _indexCount = 0;
+            private int _indexCount = -1;
             private readonly long _bound = (long)segment._size * segment._stride;
             private long _index = (long)segment._offset * segment._stride;
             private readonly MatrixSegment<T> _segment = segment;
diff --git a/StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs b/StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
index 29fd591e..f28bbada 100644
--- a/StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
+++ b/StgSharp.Mathematics/Numeric/MatrixCompute.Basic.cs
@@ -45,24 +45,61 @@ namespace StgSharp.Mathematics.Numeric
                 throw new IndexOutOfRangeException("Matrix dimensions must agree.");
             }
             Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
-            if (left.KernelColumnLength * left.KernelRowLength <= 1024)
+            long count = (long)left.KernelColumnLength * left.KernelRowLength;
+            if (count <= 1024)
             {
                 MatrixParallelPool pool = MatrixParallel.LeadParallel();
-                if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle)) { }
-                MatrixSegmentEnumeration<float>.Enumerator leftEnum = left.GetColumnEnumeration().GetEnumerator();
-                MatrixSegmentEnumeration<float>.Enumerator rightEnum = right.GetColumnEnumeration().GetEnumerator();
-                MatrixSegmentEnumeration<float>.Enumerator ansEnum = ans.GetColumnEnumeration().GetEnumerator();
-
-                for (int i = 0; i < left.KernelColumnLength; i++)
+                if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle))
                 {
-                    MatrixSegment<float>.Enumerator leftColumEnum = leftEnum.Current.GetEnumerator();
-                    MatrixSegment<float>.Enumerator rightColumEnum = rightEnum.Current.GetEnumerator();
-                    MatrixSegment<float>.Enumerator ansColumEnum = ansEnum.Current.GetEnumerator();
+                    /*
+                     * wait
+                     */
+                }
+                MatrixKernel<float>* leftPtr = left.Buffer;
+                MatrixKernel<float>* rightPtr = right.Buffer;
+                MatrixKernel<float>* ansPtr = ans.Buffer;
+                for (int i = 0; i < count; i++) {
+                    NativeIntrinsic.Intrinsic.f32_add(leftPtr + i, rightPtr + i, ansPtr + i);
                 }
                 return ans;
             }
             Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
                 (delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_add);
+            /*
+             * run parallel
+             */
+            return ans;
+        }
+
+        public static unsafe Matrix<float> Sub(Matrix<float> left, Matrix<float> right)
+        {
+            if (left.ColumnLength != right.ColumnLength || left.RowLength != right.RowLength) {
+                throw new IndexOutOfRangeException("Matrix dimensions must agree.");
+            }
+            Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
+            long count = (long)left.KernelColumnLength * left.KernelRowLength;
+            if (count <= 1024)
+            {
+                MatrixParallelPool pool = MatrixParallel.LeadParallel();
+                if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle))
+                {
+                    /*
+                     * wait
+                     */
+                }
+                MatrixKernel<float>* leftPtr = left.Buffer;
+                MatrixKernel<float>* rightPtr = right.Buffer;
+                MatrixKernel<float>* ansPtr = ans.Buffer;
+                for (int i = 0; i < count; i++) {
+                    NativeIntrinsic.Intrinsic.f32_sub(leftPtr + i, rightPtr + i, ansPtr + i);
+                }
+                return ans;
+            }
+            Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
+                (delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_sub);
+            /*
+             * run parallel
+             */
             return ans;
         }