Various optimizations and cleanups. "in" removed from GpuUInt128 to avoid defensive copies from IL.

EvenPerfectBitScanner.Benchmarks/GpuPow2ModBenchmarks.cs

@@ -50,7 +50,7 @@ public class GpuPow2ModBenchmarks
         return Pow2ModSingleBit(Input.Exponent, Input.Modulus);
     }
 
-    private static GpuUInt128 Pow2ModSingleBit(ulong exponent, in GpuUInt128 modulus)
+    private static GpuUInt128 Pow2ModSingleBit(ulong exponent, GpuUInt128 modulus)
     {
         if (modulus.IsZero || modulus == GpuUInt128.One)
         {

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModBenchmarks.cs

@@ -52,14 +52,31 @@ public class Pow2MontgomeryModBenchmarks
     /// Baseline right-to-left Montgomery ladder; measured 36.36 μs on the small sample set and 117.32 μs on the large one.
     /// </summary>
     [Benchmark(Baseline = true)]
-    public ulong Baseline()
+    public ulong BaselineCpu()
     {
         GetData(out var exponents, out var divisors, out _);
         ulong checksum = 0UL;
 
         for (int i = 0; i < SampleCount; i++)
         {
-            checksum ^= exponents[i].Pow2MontgomeryModWindowed(divisors[i], keepMontgomery: false);
+            checksum ^= exponents[i].Pow2MontgomeryModWindowedCpu(divisors[i], keepMontgomery: false);
+        }
+
+        return checksum;
+    }
+
+    /// <summary>
+    /// Baseline right-to-left Montgomery ladder; measured ??.?? μs on the small sample set and ??.?? μs on the large one.
+    /// </summary>
+    [Benchmark]
+    public ulong BaselineGpu()
+    {
+        GetData(out var exponents, out var divisors, out _);
+        ulong checksum = 0UL;
+
+        for (int i = 0; i < SampleCount; i++)
+        {
+            checksum ^= exponents[i].Pow2MontgomeryModWindowedGpu(divisors[i], keepMontgomery: false);
         }
 
         return checksum;
@@ -69,14 +86,31 @@ public class Pow2MontgomeryModBenchmarks
     /// Uses a known divisor cycle to fold the exponent first, dropping runtimes to 26.04 μs on small inputs (0.72× baseline) and 9.53 μs on large inputs (0.08× baseline).
     /// </summary>
     [Benchmark]
-    public ulong BaselineWithKnownCycle()
+    public ulong BaselineWithKnownCycleCpu()
     {
         GetData(out var exponents, out var divisors, out var cycles);
         ulong checksum = 0UL;
 
         for (int i = 0; i < SampleCount; i++)
         {
-            checksum ^= exponents[i].Pow2MontgomeryModWithCycle(cycles[i], divisors[i]);
+            checksum ^= exponents[i].Pow2MontgomeryModWithCycleCpu(cycles[i], divisors[i]);
+        }
+
+        return checksum;
+    }
+
+    /// <summary>
+    /// Uses a known divisor cycle to fold the exponent first, dropping runtimes to 26.04 μs on small inputs (0.72× baseline) and 9.53 μs on large inputs (0.08× baseline).
+    /// </summary>
+    [Benchmark]
+    public ulong BaselineWithKnownCycleGpu()
+    {
+        GetData(out var exponents, out var divisors, out var cycles);
+        ulong checksum = 0UL;
+
+        for (int i = 0; i < SampleCount; i++)
+        {
+            checksum ^= exponents[i].Pow2MontgomeryModWithCycleGpu(cycles[i], divisors[i]);
         }
 
         return checksum;

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModCycleComputationBenchmarks.cs

@@ -89,14 +89,32 @@ public class Pow2MontgomeryModCycleComputationBenchmarks
 	/// large set.
 	/// </summary>
 	// [Benchmark(Baseline = true)]
-	public ulong MontgomeryWithoutCycle()
+	public ulong MontgomeryWithoutCycleCpu()
 	{
 		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
 		ulong checksum = 0UL;
 
 		for (int i = 0; i < SampleCount; i++)
 		{
-			checksum ^= exponents[i].Pow2MontgomeryModWindowed(divisors[i], keepMontgomery: false);
+			checksum ^= exponents[i].Pow2MontgomeryModWindowedCpu(divisors[i], keepMontgomery: false);
+		}
+
+		return checksum;
+	}
+
+	/// <summary>
+	/// Baseline Montgomery reduction without precomputed cycle data; measured 35.41 μs on the small sample and 117.33 μs on the
+	/// large set.
+	/// </summary>
+	// [Benchmark(Baseline = true)]
+	public ulong MontgomeryWithoutCycleGpu()
+	{
+		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
+		ulong checksum = 0UL;
+
+		for (int i = 0; i < SampleCount; i++)
+		{
+			checksum ^= exponents[i].Pow2MontgomeryModWindowedGpu(divisors[i], keepMontgomery: false);
 		}
 
 		return checksum;
@@ -107,14 +125,32 @@ public class Pow2MontgomeryModCycleComputationBenchmarks
 	/// on the large one (0.08×).
 	/// </summary>
 	[Benchmark]
-	public ulong MontgomeryWithPrecomputedCycle()
+	public ulong MontgomeryWithPrecomputedCycleCpu()
 	{
 		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out ulong[] cycles);
 		ulong checksum = 0UL;
 
 		for (int i = 0; i < SampleCount; i++)
 		{
-			checksum ^= exponents[i].Pow2MontgomeryModWithCycle(cycles[i], divisors[i]);
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleCpu(cycles[i], divisors[i]);
+		}
+
+		return checksum;
+	}
+
+	/// <summary>
+	/// Montgomery reduction with a known cycle length (cycle lookup only); 25.81 μs on the small set (0.73× baseline) and 9.12 μs
+	/// on the large one (0.08×).
+	/// </summary>
+	[Benchmark]
+	public ulong MontgomeryWithPrecomputedCycleGpu()
+	{
+		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out ulong[] cycles);
+		ulong checksum = 0UL;
+
+		for (int i = 0; i < SampleCount; i++)
+		{
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleGpu(cycles[i], divisors[i]);
 		}
 
 		return checksum;
@@ -125,7 +161,7 @@ public class Pow2MontgomeryModCycleComputationBenchmarks
 	/// cycle discovery, but drops to 12.91 μs on the large set (0.11× baseline).
 	/// </summary>
 	// [Benchmark]
-	public ulong MontgomeryWithGpuCycleComputation()
+	public ulong MontgomeryWithGpuCycleComputationCpu()
 	{
 		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
 		ulong checksum = 0UL;
@@ -133,15 +169,33 @@ public class Pow2MontgomeryModCycleComputationBenchmarks
 		for (int i = 0; i < SampleCount; i++)
 		{
 			ulong cycle = MersenneDivisorCycles.CalculateCycleLengthGpu(divisors[i].Modulus);
-			checksum ^= exponents[i].Pow2MontgomeryModWithCycle(cycle, divisors[i]);
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleCpu(cycle, divisors[i]);
 		}
 
 		return checksum;
 	}
 
+	/// <summary>
+	/// Montgomery reduction with the cycle computed on the fly using the GPU helper; costs 54.09 ms on the small benchmark due to
+	/// cycle discovery, but drops to 12.91 μs on the large set (0.11× baseline).
+	/// </summary>
+	// [Benchmark]
+	public ulong MontgomeryWithGpuCycleComputationGpu()
+	{
+		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
+		ulong checksum = 0UL;
+
+		for (int i = 0; i < SampleCount; i++)
+		{
+			ulong cycle = MersenneDivisorCycles.CalculateCycleLengthGpu(divisors[i].Modulus);
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleGpu(cycle, divisors[i]);
+		}
+
+		return checksum;
+	}
 
 	[Benchmark]
-	public ulong MontgomeryWithHeuristicCycleComputation()
+	public ulong MontgomeryWithHeuristicCycleComputationCpu()
 	{
 		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
 		ulong checksum = 0UL;
@@ -149,7 +203,22 @@ public class Pow2MontgomeryModCycleComputationBenchmarks
 		for (int i = 0; i < SampleCount; i++)
 		{
 			ulong cycle = CalculateCycleLengthWithHeuristics(divisors[i].Modulus);
-			checksum ^= exponents[i].Pow2MontgomeryModWithCycle(cycle, divisors[i]);
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleCpu(cycle, divisors[i]);
+		}
+
+		return checksum;
+	}
+
+	[Benchmark]
+	public ulong MontgomeryWithHeuristicCycleComputationGpu()
+	{
+		GetData(out ulong[] exponents, out MontgomeryDivisorData[] divisors, out _);
+		ulong checksum = 0UL;
+
+		for (int i = 0; i < SampleCount; i++)
+		{
+			ulong cycle = CalculateCycleLengthWithHeuristics(divisors[i].Modulus);
+			checksum ^= exponents[i].Pow2MontgomeryModWithCycleGpu(cycle, divisors[i]);
 		}
 
 		return checksum;

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModUInt128Benchmarks.cs

@@ -176,7 +176,7 @@ public class Pow2MontgomeryModUInt128Benchmarks
         return (UInt128)result;
     }
 
-    private static bool ShouldUseSingleBit(in GpuUInt128 exponent)
+    private static bool ShouldUseSingleBit(GpuUInt128 exponent)
     {
         if (exponent.High == 0UL && exponent.Low <= Pow2WindowFallbackThreshold)
         {
@@ -257,7 +257,7 @@ public class Pow2MontgomeryModUInt128Benchmarks
         return result;
     }
 
-    private static bool IsBitSet(in GpuUInt128 value, int bitIndex)
+    private static bool IsBitSet(GpuUInt128 value, int bitIndex)
     {
         if (bitIndex >= 64)
         {
@@ -267,7 +267,7 @@ public class Pow2MontgomeryModUInt128Benchmarks
         return ((value.Low >> bitIndex) & 1UL) != 0UL;
     }
 
-    private static ulong ExtractWindowValue(in GpuUInt128 exponent, int windowStart, int windowBitCount)
+    private static ulong ExtractWindowValue(GpuUInt128 exponent, int windowStart, int windowBitCount)
     {
         if (windowStart != 0)
         {