From fb7ed7262917f7307c91d1f834a36f663347d1b5 Mon Sep 17 00:00:00 2001
From: James Jackson-South <james_south@hotmail.com>
Date: Sun, 26 Jul 2026 01:46:29 +1000
Subject: [PATCH] Use AVX-512 for byte tensor addition

---
 src/ImageSharp/Common/Helpers/TensorPrimitives.cs | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/ImageSharp/Common/Helpers/TensorPrimitives.cs b/src/ImageSharp/Common/Helpers/TensorPrimitives.cs
index 4bff7882f..4be4ab3aa 100644
--- a/src/ImageSharp/Common/Helpers/TensorPrimitives.cs
+++ b/src/ImageSharp/Common/Helpers/TensorPrimitives.cs
@@ -204,15 +204,15 @@ internal static partial class TensorPrimitives_
         ref T yRef = ref MemoryMarshal.GetReference(y);
         ref T destinationRef = ref MemoryMarshal.GetReference(destination);
         nuint length = (uint)x.Length;
+        nuint vector512Threshold = Unsafe.SizeOf<T>() == 1 ? (uint)Vector512<T>.Count : 512;
 
-        // AVX-512 setup only pays off for larger multi-byte inputs. Byte addition remains on AVX2 because direct
-        // PNG/WebP measurements show that its higher lane count does not recover the wider dispatch cost.
+        // Match the runtime's AVX-512 selection for byte-sized elements once one complete vector is available.
+        // Wider elements retain the measured crossover point where their 512-bit setup cost becomes worthwhile.
         // Each pipeline preloads its final inputs when a tail overlaps so same-start in-place operation remains correct.
         if (TOperator.Vectorizable
             && Vector512.IsHardwareAccelerated
             && Vector512<T>.IsSupported
-            && Unsafe.SizeOf<T>() > 1
-            && length >= 512)
+            && length >= vector512Threshold)
         {
             InvokeVectorized512<T, TOperator>(ref xRef, ref yRef, ref destinationRef, length);
             return;