Unnecessary memory zeroing commented out. TODOs added for the remainders.

PerfectNumbers.Core/Gpu/Kernels/IncrementalKernels.cs

@@ -68,7 +68,6 @@ internal static class IncrementalKernels
             return;
         }
 
-        // TODO: Is this expected that the value of kStart is modified?
         GpuUInt128 k = kStart + (GpuUInt128)idx;
         GpuUInt128 q = twoP;
         GpuUInt128.Mul64(ref q, k.High, k.Low);
@@ -125,13 +124,16 @@ internal static class IncrementalKernels
         ulong exponent,
         GpuUInt128 twoP,
         GpuUInt128 kStart,
-        byte lastIsSeven,
+        byte _,
         ulong divMul,
         ResidueAutomatonArgs ra,
         ArrayView<int> found,
         ArrayView1D<ulong, Stride1D.Dense> smallCycles)
 	{
+		// TODO: Modify this kernel to always assign the required found element to the result of the calculations and
+		// remove any clearing of the output buffer in the caller methods.
         ulong idx = (ulong)index.X;
+		// TODO: Replace % operator in GPU and their outgoing paths with logical operations, where possible. 
         ulong idxMod3 = idx % 3UL;
         ulong idxMod5 = idx % 5UL;
         ulong r10 = ra.Q0M10 + (ra.Step10 * idx).Mod10(); r10 -= (r10 >= 10UL) ? 10UL : 0UL;

PerfectNumbers.Core/Gpu/Kernels/Pow2ModKernels.cs

@@ -6,6 +6,7 @@ namespace PerfectNumbers.Core.Gpu;
 
 internal static class Pow2ModKernels
 {
+	/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
 	public static void Pow2ModKernelScan(
 		Index1D index,
 		ulong exponent,
@@ -22,10 +23,12 @@ internal static class Pow2ModKernels
 		ArrayView1D<ulong, Stride1D.Dense> smallPrimesPow2LastSeven)
 	{
 		ulong idx = (ulong)index.X;
+		// TODO: Replace % operator in GPU and their outgoing paths with logical operations, where possible. 
 		ulong idxMod3 = idx % 3UL;
 		ulong idxMod5 = idx % 5UL;
 		ulong r10 = ra.Q0M10 + (ra.Step10 * idx).Mod10(); r10 -= (r10 >= 10UL) ? 10UL : 0UL;
 		bool shouldCheck = r10 != 5UL;
+		// TODO: Replace these "if" statements with ?:
 		if (shouldCheck)
 		{
 			ulong r8 = (ra.Q0M8 + ((ra.Step8 * idx) & 7UL)) & 7UL;
@@ -61,7 +64,6 @@ internal static class Pow2ModKernels
 			return;
 		}
 
-        // TODO: kStart is modified after this. Is this expected?
 		kStart.Add(idx);
 		GpuUInt128 q = twoP;
 		GpuUInt128.Mul64(ref q, kStart.High, kStart.Low);
@@ -104,13 +106,15 @@ internal static class Pow2ModKernels
 		orders[index] = exponent;
 	}
 
+	/// This kernel doesn't always set the result of the found array. It sets the value only when it needs to. The callers
+	/// must clean the output buffer before the call to get a deterministic result.
 	public static void Pow2ModOrderKernelScan(
 		Index1D index,
 		ulong exponent,
 		GpuUInt128 twoP,
 		GpuUInt128 kStart,
-        byte lastIsSeven,
-        ulong _,
+		byte _,
+		ulong __,
 		ResidueAutomatonArgs ra,
 		ArrayView<int> found,
 		ArrayView1D<ulong, Stride1D.Dense> smallCycles)

PerfectNumbers.Core/Gpu/Kernels/PrimeOrderKernels.cs

@@ -10,6 +10,8 @@ namespace PerfectNumbers.Core.Gpu;
 
 internal static partial class PrimeOrderGpuHeuristics
 {
+	/// This kernel always sets the result and operates within the allowed / required bounds, never accessing elements outside of them.
+	/// It sets the value only when it needs to. The callers don't need to clear the output / input buffers.
     internal static void PartialFactorKernel(
         Index1D index,
         ArrayView1D<uint, Stride1D.Dense> primes,
@@ -111,6 +113,8 @@ internal static partial class PrimeOrderGpuHeuristics
         }
     }
 
+	/// This kernel doesn't always set the result of the first element. It sets the value only when it needs to. The callers
+	/// must clean the output buffer before the call to get a deterministic result.
     internal static void CalculateOrderKernel(
         Index1D index,
         ulong prime,
@@ -126,6 +130,8 @@ internal static partial class PrimeOrderGpuHeuristics
             return;
         }
 
+		// TODO: Modify this kernel to always set the output result so that callers never need to clear any buffers and remove
+		// any existing cleaning of the buffers after that.
 		ArrayView1D<ulong, Stride1D.Dense> phiFactors = buffers.PhiFactors;
         ArrayView1D<int, Stride1D.Dense> phiExponents = buffers.PhiExponents;
         ArrayView1D<ulong, Stride1D.Dense> workFactors = buffers.WorkFactors;
@@ -145,6 +151,9 @@ internal static partial class PrimeOrderGpuHeuristics
 
 		statusOut[0] = (byte)PrimeOrderKernelStatus.Fallback;
 
+		// TODO: Is this condition ever satisfied on EventPerfectBitScanner's execution paths? Comment it out with explanatory
+		// comment before it, if not. Otherwise add a comment explaining how it's used. If tests or benchmarks require this,
+		// modify such tests and benchmark to avoid that.
 		if (prime <= 3UL)
 		{
 			ulong orderValue = prime == 3UL ? 2UL : 1UL;
@@ -1153,6 +1162,8 @@ internal static partial class PrimeOrderGpuHeuristics
         return order;
     }
 
+	/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
+
 	internal static void Pow2ModKernel(Index1D index, ArrayView1D<ulong, Stride1D.Dense> exponents, MontgomeryDivisorData divisor, ArrayView1D<ulong, Stride1D.Dense> remainders)
 	{
 		ulong exponent = exponents[index];
@@ -1169,6 +1180,7 @@ internal static partial class PrimeOrderGpuHeuristics
         });
     }
 
+	/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
     internal static void Pow2ModKernelWide(Index1D index, ArrayView1D<GpuUInt128, Stride1D.Dense> exponents, GpuUInt128 modulus, ArrayView1D<GpuUInt128, Stride1D.Dense> remainders)
     {
         GpuUInt128 exponent = exponents[index];

PerfectNumbers.Core/Gpu/MersenneNumberDivisorGpuTester.cs

@@ -41,7 +41,8 @@ public sealed class MersenneNumberDivisorGpuTester
         var accelerator = gpu.Accelerator;
         var kernel = GetKernel(accelerator);
         var resultBuffer = _resultBuffers.GetOrAdd(accelerator, acc => acc.Allocate1D<byte>(1));
-        resultBuffer.MemSetToZero();
+		// There is no point in clearing this buffer. We always override item [0] and never use it beyond item [0]
+        // resultBuffer.MemSetToZero();
         kernel(1, exponent, divisor, resultBuffer.View);
         accelerator.Synchronize();
         Span<byte> result = stackalloc byte[1];

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -35,6 +35,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 			{
 				remaining = maxK - kStart + 1UL;
 				currentSize = remaining > (UInt128)batchSize ? batchSize : (int)remaining;
+				// This clear is required because the result may not always be set. Remove it after the kernels are modified to always set the result.
 				foundBuffer.MemSetToZero();
 				// Precompute residue automaton bases for this batch
                                 UInt128 q0 = twoP * kStart + 1UL;

PerfectNumbers.Core/Gpu/MersenneNumberResidueGpuTester.cs

@@ -144,7 +144,8 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
         {
             if (resources.Capacity >= capacity)
 			{
-                resources.OrderBuffer.MemSetToZero();
+				// The caller always overwrites the required elements. We don't need to worry about clearing the output buffer.
+                // resources.OrderBuffer.MemSetToZero();
                 return resources;
             }
 
@@ -152,7 +153,8 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
         }
 
         var created = new ResidueResources(accelerator, capacity);
-        created.OrderBuffer.MemSetToZero();
+		// The caller always overwrites the required elements. We don't need to worry about clearing the output buffer.
+        // created.OrderBuffer.MemSetToZero();
         return created;
     }
 

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -201,17 +201,19 @@ internal static partial class PrimeOrderGpuHeuristics
             var kernel = GetPartialFactorKernel(accelerator);
 			SmallPrimeDeviceCache cache = GetSmallPrimeDeviceCache(accelerator);
 
+			// TODO: We should create / reallocate these buffer only once or if the new required length is bigger than capacity. 
             var factorBuffer = accelerator.Allocate1D<ulong>(primeTargets.Length);
             var exponentBuffer = accelerator.Allocate1D<int>(exponentTargets.Length);
             var countBuffer = accelerator.Allocate1D<int>(1);
             var remainingBuffer = accelerator.Allocate1D<ulong>(1);
             var fullyFactoredBuffer = accelerator.Allocate1D<byte>(1);
 
-            factorBuffer.MemSetToZero();
-            exponentBuffer.MemSetToZero();
-            countBuffer.MemSetToZero();
-            remainingBuffer.MemSetToZero();
-            fullyFactoredBuffer.MemSetToZero();
+			// TODO: There is no need to clear these buffers because the kernel will always assign values within the required bounds.
+			// factorBuffer.MemSetToZero();
+            // exponentBuffer.MemSetToZero();
+            // countBuffer.MemSetToZero();
+            // remainingBuffer.MemSetToZero();
+            // fullyFactoredBuffer.MemSetToZero();
 
             kernel(
                 stream,
@@ -240,6 +242,7 @@ internal static partial class PrimeOrderGpuHeuristics
             fullyFactoredBuffer.View.CopyToCPU(ref fullyFactoredFlag, 1);
             fullyFactored = fullyFactoredFlag != 0;
 
+			// TODO: These buffers shouldn't be disposed but rather reused accross call with the assigned accelerator.
             factorBuffer.Dispose();
             exponentBuffer.Dispose();
             countBuffer.Dispose();
@@ -366,6 +369,9 @@ internal static partial class PrimeOrderGpuHeuristics
         var kernel = GetOrderKernel(accelerator);
 		SmallPrimeDeviceCache cache = GetSmallPrimeDeviceCache(accelerator);
 
+		// TODO: These buffers should be created reallocated once and assigned to an accelerator. Callers should use their own
+		// thread static cache to prevent other threads from using taking accelerators with pre-allocated buffers if they don't
+		// use them.
         var phiFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
         var phiExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
         var workFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
@@ -377,6 +383,7 @@ internal static partial class PrimeOrderGpuHeuristics
         var resultBuffer = accelerator.Allocate1D<ulong>(1);
         var statusBuffer = accelerator.Allocate1D<byte>(1);
 
+		// TODO: Remove the cleaning after the order kernel is modified to always set the result.
 		phiFactorBuffer.MemSetToZero();
         phiExponentBuffer.MemSetToZero();
         workFactorBuffer.MemSetToZero();
@@ -523,11 +530,16 @@ internal static partial class PrimeOrderGpuHeuristics
         Accelerator accelerator = lease.Accelerator;
         AcceleratorStream stream = lease.Stream;
         var kernel = GetPow2ModKernel(accelerator);
+		// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
+		// Modify the callers to use their own pool of buffers per accelerator, so that other threads don't use the accelerators
+		// with out preallocated buffers. Share the pool with Pow2ModWide kernel.
         var exponentBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
         var remainderBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
 
         exponentBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(exponents), exponents.Length);
-        remainderBuffer.MemSetToZero();
+
+		// Pow2Mod kernel always assigns the required output elements so we don't need to worry about clearing these.
+		// remainderBuffer.MemSetToZero();
 
 		try
 		{
@@ -579,6 +591,9 @@ internal static partial class PrimeOrderGpuHeuristics
             Accelerator accelerator = lease.Accelerator;
             AcceleratorStream stream = lease.Stream;
             var kernel = GetPow2ModWideKernel(accelerator);
+			// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
+			// Modify the callers to use their own pool of buffers per accelerator, so that other threads don't use the accelerators
+			// with out preallocated buffers. Share the pool with Pow2Mod kernel.
             var exponentBuffer = accelerator.Allocate1D<GpuUInt128>(length);
             var remainderBuffer = accelerator.Allocate1D<GpuUInt128>(length);
 
@@ -592,7 +607,8 @@ internal static partial class PrimeOrderGpuHeuristics
             }
 
             exponentBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(exponentSpan), length);
-            remainderBuffer.MemSetToZero();
+			// Pow2ModWide kernel always assigns the required output elements so we don't need to worry about clearing these.
+            // remainderBuffer.MemSetToZero();
 
             GpuUInt128 modulus = (GpuUInt128)prime;
             kernel(stream, length, exponentBuffer.View, modulus, remainderBuffer.View);