Inline GpuUInt128 Mul64 helper

EvenPerfectBitScanner.Benchmarks/GpuUInt128Mul64Benchmarks.cs

@@ -33,7 +33,7 @@ public class GpuUInt128Mul64Benchmarks
     public GpuUInt128 HighProductMaterializedInLocal()
     {
         GpuUInt128 value = Input.Multiplicand;
-        value.Mul64(Input.Multiplier);
+        GpuUInt128.Mul64(ref value, Input.Multiplier.High, Input.Multiplier.Low);
         return value;
     }
 

EvenPerfectBitScanner.Benchmarks/GpuUInt128MulModBenchmarks.cs

@@ -1,3 +1,4 @@
+using System.Numerics;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
@@ -67,7 +68,6 @@ public class GpuUInt128MulModBenchmarks
     [ArgumentsSource(nameof(GetGeneralCases))]
     public GpuUInt128 InPlaceMulMod(GpuMulModCase input)
     {
-        // TODO: Switch the extension method to allocate per iteration, as it's faster. Keep the benchmarks.
         GpuUInt128 value = input.Left;
         value.MulMod(input.Right, input.Modulus);
         return value;
@@ -87,6 +87,16 @@ public class GpuUInt128MulModBenchmarks
         return MulModInline(input.Left, input.Right, input.Modulus);
     }
 
+    /// <summary>
+    /// BigInteger-based reference reduction kept for validation; remains several microseconds slower than the GPU-compatible implementations.
+    /// </summary>
+    [Benchmark]
+    [ArgumentsSource(nameof(GetGeneralCases))]
+    public GpuUInt128 BigIntegerReference(GpuMulModCase input)
+    {
+        return MulModBigInteger(input.Left, input.Right, input.Modulus);
+    }
+
     /// <summary>
     /// Legacy implementation that allocates intermediate limbs; delivered 10.7 ns on tiny operands, 12.7 ns on high-word-dominant
     /// pairs, and 16.3 ns on mixed magnitude, making it the faster choice overall.
@@ -112,10 +122,7 @@ public class GpuUInt128MulModBenchmarks
     [ArgumentsSource(nameof(GetByLimbCases))]
     public GpuUInt128 InPlaceReduction(GpuMulModCase input)
     {
-        // TODO: Switch the extension method back to legacy method, as it's faster. Keep the benchmarks.
-        GpuUInt128 value = input.Left;
-        value.MulModByLimb(input.Right, input.Modulus);
-        return value;
+        return MulModByLimbInPlace(input.Left, input.Right, input.Modulus);
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -158,6 +165,46 @@ public class GpuUInt128MulModBenchmarks
         return result;
     }
 
+    private static GpuUInt128 MulModBigInteger(GpuUInt128 left, GpuUInt128 right, GpuUInt128 modulus)
+    {
+        BigInteger leftBig = (UInt128)left;
+        BigInteger rightBig = (UInt128)right;
+        BigInteger modulusBig = (UInt128)modulus;
+        if (modulusBig.IsZero)
+        {
+            return GpuUInt128.Zero;
+        }
+
+        UInt128 reduced = (UInt128)((leftBig * rightBig) % modulusBig);
+        return new GpuUInt128((ulong)(reduced >> 64), (ulong)reduced);
+    }
+
+    private static GpuUInt128 MulModByLimbInPlace(GpuUInt128 left, GpuUInt128 right, GpuUInt128 modulus)
+    {
+        var (p3, p2, p1, p0) = MultiplyFullLegacy(left, right);
+
+        GpuUInt128 remainder = new(p3, p2);
+        while (remainder.CompareTo(modulus) >= 0)
+        {
+            remainder.Sub(modulus);
+        }
+
+        ulong limb = p1;
+        for (int i = 0; i < 2; i++)
+        {
+            remainder.ShiftLeft(64);
+            remainder.Low = limb;
+            while (remainder.CompareTo(modulus) >= 0)
+            {
+                remainder.Sub(modulus);
+            }
+
+            limb = p0;
+        }
+
+        return remainder;
+    }
+
     private static GpuUInt128 MulModByLimbLegacy(GpuUInt128 left, GpuUInt128 right, GpuUInt128 modulus)
     {
         var (p3, p2, p1, p0) = MultiplyFullLegacy(left, right);

EvenPerfectBitScanner.Benchmarks/MulMod64Benchmarks.cs

@@ -157,9 +157,8 @@ public class MulMod64Benchmarks
     }
 
     /// <summary>
-    /// Uses the GPU-friendly <see cref="ULongExtensions.MulMod64Gpu(ulong, ulong, ulong)"/> helper; dense inputs now
-    /// cost 24.4–61.9 ns, while NearFullRange and SparseOperands finish near 7.27 ns and ZeroOperands stay baseline-fast at
-    /// 3.59 ns.
+    /// Benchmarks the legacy GPU-oriented shim that previously powered <see cref="ULongExtensions.MulMod64Gpu(ulong, ulong, ulong)"/>;
+    /// dense inputs still land in the 24.4–61.9 ns range while sparse operands remain closer to 3.6–7.3 ns.
     /// </summary>
     /// <remarks>
     /// Observed means: CrossWordBlend 24.422 ns (6.86×), MixedBitPattern 61.850 ns, NearFullRange 7.267 ns,
@@ -168,14 +167,12 @@ public class MulMod64Benchmarks
     [Benchmark]
     public ulong GpuCompatibleBaseline()
     {
-        // TODO: Callers in production should migrate to ULongExtensions.MulMod64 where GPU parity
-        // is not required; that baseline stayed 6–17× faster for dense 64-bit operands.
-        return Input.Left.MulMod64Gpu(Input.Right, Input.Modulus);
+        return MulMod64GpuLegacy(Input.Left, Input.Right, Input.Modulus);
     }
 
     /// <summary>
-    /// Uses the deferred native-modulo helper <see cref="ULongExtensions.MulMod64GpuDeferred(ulong, ulong, ulong)"/>;
-    /// excels when operands are tiny (2.01 ns on ZeroOperands) yet remains the slowest choice on dense data at 146–295 ns.
+    /// Benchmarks the deferred native-modulo shim that now lives only in this suite; tiny operands still excel around 2 ns
+    /// while dense datasets take 146–295 ns.
     /// </summary>
     /// <remarks>
     /// Observed means: CrossWordBlend 146.559 ns (40.63×), MixedBitPattern 255.555 ns, NearFullRange 21.390 ns,
@@ -184,9 +181,21 @@ public class MulMod64Benchmarks
     [Benchmark]
     public ulong GpuCompatibleDeferred()
     {
-        // TODO: Retire the deferred GPU shim from runtime paths and keep it only for benchmarks;
-        // ULongExtensions.MulMod64 avoids the 6×–82× slowdown on real-world operand mixes.
-        return Input.Left.MulMod64GpuDeferred(Input.Right, Input.Modulus);
+        return MulMod64GpuDeferredLegacy(Input.Left, Input.Right, Input.Modulus);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static ulong MulMod64GpuLegacy(ulong a, ulong b, ulong modulus)
+    {
+        GpuUInt128 state = new(a % modulus);
+        return state.MulMod(b, modulus);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static ulong MulMod64GpuDeferredLegacy(ulong a, ulong b, ulong modulus)
+    {
+        GpuUInt128 state = new(a);
+        return state.MulModWithNativeModulo(b, modulus);
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]