SPMD SIMD math library & lock-free job system integration

- Add new SPMD SIMD math project with scalar/vector lanes
- Integrate SPMD jobs and scheduling into job system
- Implement lock-free job dependency management
- Update math functions for .NET 10 and SIMD performance
- Add SPMD benchmarks, compress-store tests, and race tests
- Introduce generic Result<T> error handling utilities
- Solution/project file updates and code cleanup

Misaki.HighPerformance.Jobs/IJob.cs

@@ -1,4 +1,4 @@
-namespace Misaki.HighPerformance.Jobs;
+namespace Misaki.HighPerformance.Jobs;
 
 /// <summary>
 /// Represents a job that performs a single unit of work.
@@ -24,3 +24,24 @@ public interface IJobParallelFor
     /// <param name="threadIndex">The index of the thread executing the job, useful for thread-specific operations.</param>
     void Execute(int loopIndex, int threadIndex);
 }
+
+public static class IJobExtensions
+{
+    public static void Run<T>(this ref T job, int threadIndex)
+        where T : unmanaged, IJob
+    {
+        job.Execute(threadIndex);
+    }
+}
+
+public static class IJobParallelForExtensions
+{
+    public static void Run<T>(this ref T job, int totalIterations, int threadIndex)
+        where T : unmanaged, IJobParallelFor
+    {
+        for (var i = 0; i < totalIterations; i++)
+        {
+            job.Execute(i, threadIndex);
+        }
+    }
+}

Misaki.HighPerformance.Jobs/JobHandle.cs

@@ -1,18 +1,21 @@
-namespace Misaki.HighPerformance.Jobs;
+namespace Misaki.HighPerformance.Jobs;
 
 public readonly struct JobHandle : IEquatable<JobHandle>
 {
-    internal readonly int _id;
+    private readonly int _id;
-    internal readonly int _generation;
+    private readonly int _generation;
 
-    public static JobHandle Invalid => new(-1, -1);
+    public int ID => _id - 1;
+    public int Generation => _generation - 1;
+
+    public static JobHandle Invalid => default;
 
     public bool IsValid => this != Invalid;
 
     internal JobHandle(int id, int generation)
     {
-        _id = id;
+        _id = id + 1;
-        _generation = generation;
+        _generation = generation + 1;
     }
 
     public bool Equals(JobHandle other)

Misaki.HighPerformance.Jobs/JobScheduler.cs

@@ -87,7 +87,7 @@ public interface IJobScheduler
     /// <param name="threadIndex">The index of the thread that will execute the job. This is used to assign thread-specific data. Use -1 to allow any thread to execute the job.</param>
     /// <returns>A <see cref="JobHandle"/> that can be used to track the completion of the scheduled job.
     ///     Returns <see cref="JobHandle.Invalid"/> if the job data allocation fails.</returns>
-    public JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize, int threadIndex)
+    JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize, int threadIndex)
         where T : unmanaged, IJobParallelFor;
 
     /// <summary>
@@ -100,7 +100,7 @@ public interface IJobScheduler
     /// <param name="threadIndex">The index of the thread that will execute the job. This is used to assign thread-specific data. Use -1 to allow any thread to execute the job.</param>
     /// <returns>A <see cref="JobHandle"/> that can be used to track the completion of the scheduled job.
     ///     Returns <see cref="JobHandle.Invalid"/> if the job data allocation fails.</returns>
-    public JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize, JobHandle dependency)
+    JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize, JobHandle dependency)
         where T : unmanaged, IJobParallelFor;
 
     /// <summary>
@@ -113,7 +113,7 @@ public interface IJobScheduler
     /// <param name="threadIndex">The index of the thread that will execute the job. This is used to assign thread-specific data. Use -1 to allow any thread to execute the job.</param>
     /// <returns>A <see cref="JobHandle"/> that can be used to track the completion of the scheduled job.
     ///     Returns <see cref="JobHandle.Invalid"/> if the job data allocation fails.</returns>
-    public JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize)
+    JobHandle ScheduleParallel<T>(ref T job, int totalIteration, int batchSize)
         where T : unmanaged, IJobParallelFor;
 
     /// <summary>
@@ -122,7 +122,7 @@ public interface IJobScheduler
     /// <param name="dependencies">A collection of <see cref="JobHandle"/> instances representing the dependencies to combine.</param>
     /// <returns>A <see cref="JobHandle"/> that represents the combined dependencies. The returned handle can be used to ensure
     ///     that all specified dependencies are completed before proceeding.</returns>
-    public JobHandle CombineDependencies(params ReadOnlySpan<JobHandle> dependencies);
+    JobHandle CombineDependencies(params ReadOnlySpan<JobHandle> dependencies);
 
     /// <summary>
     /// Retrieves the current status of a job identified by the specified handle.
@@ -130,13 +130,13 @@ public interface IJobScheduler
     /// <param name="handle">The handle representing the job whose status is to be retrieved. The handle must be valid.</param>
     /// <returns>The current status of the job as a <see cref="JobState"/> value.
     ///     Returns <see cref="JobState.Invalid"/> if the handle is invalid or the job does not exist.</returns>
-    public JobState GetJobStatus(JobHandle handle);
+    JobState GetJobStatus(JobHandle handle);
 
     /// <summary>
     /// Blocks the calling thread until the specified job is completed.
     /// </summary>
     /// <param name="handle">The handle of the job to wait for.</param>
-    public void WaitComplete(JobHandle handle);
+    void WaitComplete(JobHandle handle);
 
     /// <summary>
     /// Blocks the calling thread until all specified job handles have completed.
@@ -147,7 +147,7 @@ public interface IJobScheduler
     /// concurrently from multiple threads.</remarks>
     /// <param name="handles">A collection of job handles to wait for. Each handle represents an asynchronous job whose completion is awaited.
     /// The collection must not be empty.</param>
-    public void WaitAll(params ReadOnlySpan<JobHandle> handles);
+    void WaitAll(params ReadOnlySpan<JobHandle> handles);
 
     /// <summary>
     /// Waits until any of the specified job handles has completed and returns the first completed handle.
@@ -158,12 +158,14 @@ public interface IJobScheduler
     /// <param name="handles">A read-only span containing the job handles to monitor for completion. Each handle represents a job whose
     /// completion status will be checked.</param>
     /// <returns>The first job handle from the provided collection that has completed.</returns>
-    public JobHandle WaitAny(params ReadOnlySpan<JobHandle> handles);
+    JobHandle WaitAny(params ReadOnlySpan<JobHandle> handles);
 }
 
 public unsafe partial class JobScheduler
 {
-    public static readonly TempJobAllocator* pTempAllocator;
+    public static int MainThreadIndex => -1;
+
+    public static TempJobAllocator* pTempAllocator;
 
     /// <summary>
     /// Gets the allocation handle for the temporary job allocator.
@@ -173,7 +175,7 @@ public unsafe partial class JobScheduler
     /// </remarks>
     public static AllocationHandle TempAllocatorHandle => pTempAllocator->Handle;
 
-    static JobScheduler()
+    public static void InitTempAllocator()
     {
         pTempAllocator = (TempJobAllocator*)MemoryUtility.Malloc((nuint)sizeof(TempJobAllocator));
         pTempAllocator->Init();
@@ -194,14 +196,18 @@ public unsafe partial class JobScheduler
 /// </summary>
 public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
 {
-    private const int _SLEEP_THRESHOLD = 100;
+    // Don't sleep indefinitely because that causes our 1ms job to become 15ms.
+    private const int _SLEEP_THRESHOLD = -1;
+
+    // Lock-Free constants: State mask (low 16 bits) and RC unit (1 << 16)
+    private const int _STATE_MASK = 0xFFFF;
+    private const int _RC_ONE = 0x10000;
 
     private FreeList _jobDataAllocator;
     private readonly ConcurrentSlotMap<JobInfo> _jobInfoPool;
     private readonly ConcurrentQueue<JobHandle> _jobQueue;
     private readonly WorkerThread[] _workerThreads;
 
-    private readonly Lock _lock;
     private readonly SemaphoreSlim _workSignal;
     private readonly CancellationTokenSource _cts;
 
@@ -221,7 +227,6 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
         _jobInfoPool = new();
         _jobQueue = new();
 
-        _lock = new();
         _workSignal = new(0);
         _cts = new();
 
@@ -246,10 +251,11 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
 
     private void EnqueueJobIfReady(JobHandle handle)
     {
-        ref var jobInfo = ref _jobInfoPool.GetElementReferenceAt(handle._id, handle._generation, out var exist);
+        ref var jobInfo = ref _jobInfoPool.GetElementReferenceAt(handle.ID, handle.Generation, out var exist);
 
         if (exist && Volatile.Read(ref jobInfo.dependencyCount) == 0)
         {
+            // Note: JobState.Created is 0, JobState.Scheduled is 1. We assume RC logic doesn't touch initial state (RC=0).
             if (Interlocked.CompareExchange(ref jobInfo.state, JobState.Scheduled, JobState.Created) != JobState.Created)
             {
                 return;
@@ -293,28 +299,74 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
                 continue;
             }
 
-            lock (_lock)
+            ref var depJobInfo = ref _jobInfoPool.GetElementReferenceAt(dependency.ID, dependency.Generation, out var exist);
+            if (!exist)
             {
-                ref var depJobInfo = ref _jobInfoPool.GetElementReferenceAt(dependency._id, dependency._generation, out var exist);
+                // Dependency does not exist (likely completed already)
-                if (!exist || Volatile.Read(ref Unsafe.As<JobState, int>(ref depJobInfo.state)) == (int)JobState.Completed)
+                continue;
-                {
-                    continue;
-                }
-
-                if (depJobInfo.dependentCount >= JobInfo.MAX_DEPENDENTS)
-                {
-                    // Too many dependents
-                    // TODO: Handle this case properly
-                    _jobDataAllocator.Free(jobInfo.pJobData);
-                    return JobHandle.Invalid;
-                }
-
-                depJobInfo.dependentsID[depJobInfo.dependentCount] = id;
-                depJobInfo.dependentsGeneration[depJobInfo.dependentCount] = generation;
-                depJobInfo.dependentCount++;
             }
 
-            Interlocked.Increment(ref infoInPool.dependencyCount);
+            // Lock-free registration: Try to acquire "Reader Lock" by incrementing RC in high bits.
+            // If state is already Completed, we skip (dependency met).
+            var registered = false;
+            var completed = false;
+            var spin = new SpinWait();
+
+            while (true)
+            {
+                var stateVal = Volatile.Read(ref Unsafe.As<JobState, int>(ref depJobInfo.state));
+                var state = (JobState)(stateVal & _STATE_MASK);
+
+                if (state == JobState.Completed)
+                {
+                    completed = true;
+                    break;
+                }
+
+                // Attempt to increment RC (Reader Count)
+                if (Interlocked.CompareExchange(ref Unsafe.As<JobState, int>(ref depJobInfo.state), stateVal + _RC_ONE, stateVal) == stateVal)
+                {
+                    // RC acquired. We are safe from "Remove" and state change.
+                    var count = Interlocked.Increment(ref depJobInfo.dependentCount);
+                    if (count <= JobInfo.MAX_DEPENDENTS)
+                    {
+                        // Safely write to the fixed buffer
+                        depJobInfo.dependentsID[count - 1] = id;
+                        depJobInfo.dependentsGeneration[count - 1] = generation;
+                        registered = true;
+                    }
+
+                    // Release RC
+                    Interlocked.Add(ref Unsafe.As<JobState, int>(ref depJobInfo.state), -_RC_ONE);
+
+                    if (!registered)
+                    {
+                        // Failed to register because MAX_DEPENDENTS reached.
+                        // Backtrack the counter increment.
+                        Interlocked.Decrement(ref depJobInfo.dependentCount);
+
+                        // Cleanup and fail
+                        _jobDataAllocator.Free(jobInfo.pJobData);
+                        return JobHandle.Invalid;
+                    }
+
+                    break;
+                }
+
+                spin.SpinOnce(-1);
+            }
+
+            if (!registered && !completed)
+            {
+                // Should not happen if logic is correct, unless loop logic changed
+                Interlocked.Increment(ref infoInPool.dependencyCount);
+            }
+            else if (registered)
+            {
+                // Successfully added dependency
+                Interlocked.Increment(ref infoInPool.dependencyCount);
+            }
+            // else: completed is true, registered is false -> Dependency is already done, so we don't increment our dependencyCount.
         }
 
         EnqueueJobIfReady(handle);
@@ -325,7 +377,20 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     internal bool HasWork()
     {
-        return !_jobQueue.IsEmpty || _workerThreads.Any(w => !w.LocalQueue.IsEmpty);
+        if (!_jobQueue.IsEmpty)
+        {
+            return true;
+        }
+
+        for (var i = 0; i < _workerThreads.Length; i++)
+        {
+            if (!_workerThreads[i].LocalQueue.IsEmpty)
+            {
+                return true;
+            }
+        }
+
+        return false;
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -360,7 +425,7 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
             return ref Unsafe.NullRef<JobInfo>();
         }
 
-        return ref _jobInfoPool.GetElementReferenceAt(handle._id, handle._generation, out exist);
+        return ref _jobInfoPool.GetElementReferenceAt(handle.ID, handle.Generation, out exist);
     }
 
     internal void MarkJobComplete(JobHandle handle)
@@ -370,37 +435,77 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
             return;
         }
 
-        ref var info = ref _jobInfoPool.GetElementReferenceAt(handle._id, handle._generation, out var exist);
+        ref var info = ref _jobInfoPool.GetElementReferenceAt(handle.ID, handle.Generation, out var exist);
         if (!exist)
         {
             return;
         }
 
-        if (Interlocked.CompareExchange(ref info.state, JobState.Completed, JobState.Running) != JobState.Running)
+        // Lock-free Completion:
+        // 1. Transition State to Completed (preserving or setting upper bits?).
+        //    Actually, we want to block new Readers. Setting state to Completed blocks new Readers.
+        // 2. Wait for existing Readers (RC == 0).
+        var spin = new SpinWait();
+        while (true)
         {
-            return;
+            var stateVal = Volatile.Read(ref Unsafe.As<JobState, int>(ref info.state));
+            var state = (JobState)(stateVal & _STATE_MASK);
+
+            if (state == JobState.Completed)
+            {
+                return; // Already completed (shouldn't happen for single-execution jobs)
+            }
+
+            if (state != JobState.Running)
+            {
+                // If in valid state (e.g. Scheduled?), we still assume we can complete it.
+                // Usually it should be Running.
+            }
+
+            // Construct new value: State=Completed, preserve RC (temporarily) or strictly replace only low bits?
+            // We set low bits to Completed. High bits (RC) remain.
+            var newState = (stateVal & ~_STATE_MASK) | (int)JobState.Completed;
+
+            if (Interlocked.CompareExchange(ref Unsafe.As<JobState, int>(ref info.state), newState, stateVal) == stateVal)
+            {
+                // Successfully set State to Completed. New readers will see Completed and back off.
+                // Now we must wait for existing readers to finish (RC to become 0).
+                while (true)
+                {
+                    var current = Volatile.Read(ref Unsafe.As<JobState, int>(ref info.state));
+                    if (((uint)current >> 16) == 0)
+                    {
+                        break; // RC is 0. Safe to proceed.
+                    }
+
+                    spin.SpinOnce(-1);
+                }
+                break;
+            }
+
+            spin.SpinOnce(-1);
         }
 
-        var dependentsToNotify = stackalloc JobHandle[JobInfo.MAX_DEPENDENTS];
+        // We now have exclusive access to dependentsID (no new readers, old readers finished).
-        var dependentCount = 0;
+        // Safely capture dependents.
+        var dependentCount = info.dependentCount;
+        dependentCount = Math.Min(dependentCount, JobInfo.MAX_DEPENDENTS); // Safety cap
 
-        lock (_lock)
+        // Use stackalloc to avoid allocation, but we'll copy to notify after freeing parent.
+        var dependentsToNotify = stackalloc JobHandle[dependentCount];
+        for (var i = 0; i < dependentCount; i++)
         {
-            dependentCount = info.dependentCount;
+            dependentsToNotify[i] = new JobHandle(info.dependentsID[i], info.dependentsGeneration[i]);
-            for (var i = 0; i < dependentCount; i++)
-            {
-                dependentsToNotify[i] = new JobHandle(info.dependentsID[i], info.dependentsGeneration[i]);
-            }
         }
 
         _jobDataAllocator.Free(info.pJobData);
-        _jobInfoPool.Remove(handle._id, handle._generation);
+        _jobInfoPool.Remove(handle.ID, handle.Generation);
 
         for (var i = 0; i < dependentCount; i++)
         {
             var depHandle = dependentsToNotify[i];
 
-            ref var depJobInfo = ref _jobInfoPool.GetElementReferenceAt(depHandle._id, depHandle._generation, out var depExist);
+            ref var depJobInfo = ref _jobInfoPool.GetElementReferenceAt(depHandle.ID, depHandle.Generation, out var depExist);
             if (depExist && Interlocked.Decrement(ref depJobInfo.dependencyCount) == 0)
             {
                 EnqueueJobIfReady(depHandle);
@@ -437,8 +542,8 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
     }
 
     public JobHandle Schedule<T>(ref T job, int threadIndex)
-    where T : unmanaged, IJob
+        where T : unmanaged, IJob
-    => Schedule(ref job, threadIndex, JobHandle.Invalid);
+        => Schedule(ref job, threadIndex, JobHandle.Invalid);
 
     public JobHandle Schedule<T>(ref T job, JobHandle dependency)
         where T : unmanaged, IJob
@@ -519,13 +624,14 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
             return JobState.Invalid;
         }
 
-        ref var jobInfo = ref _jobInfoPool.GetElementReferenceAt(handle._id, handle._generation, out var exist);
+        ref var jobInfo = ref _jobInfoPool.GetElementReferenceAt(handle.ID, handle.Generation, out var exist);
         if (!exist)
         {
             return JobState.Completed; // We assume completed if not found. Invalid state is reserved for error.
         }
 
-        return (JobState)Volatile.Read(ref Unsafe.As<JobState, int>(ref jobInfo.state));
+        // Mask out the Reader Count (upper 16 bits) to return the actual State
+        return (JobState)(Volatile.Read(ref Unsafe.As<JobState, int>(ref jobInfo.state)) & _STATE_MASK);
     }
 
     public void WaitComplete(JobHandle handle)
@@ -536,9 +642,10 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
         }
 
         var spin = new SpinWait();
-        while (_jobInfoPool.TryGetElement(handle._id, handle._generation, out var jobInfo))
+        while (_jobInfoPool.TryGetElement(handle.ID, handle.Generation, out var jobInfo))
         {
-            if (jobInfo.state == JobState.Completed)
+            // Mask out RC
+            if ((jobInfo.state & (JobState)_STATE_MASK) == JobState.Completed)
             {
                 return;
             }
@@ -549,7 +656,6 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
 
     public void WaitAll(params ReadOnlySpan<JobHandle> handles)
     {
-        var sleepThreshold = _SLEEP_THRESHOLD * handles.Length;
         var spin = new SpinWait();
 
         while (true)
@@ -557,7 +663,7 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
             var completedCount = 0;
             foreach (var handle in handles)
             {
-                if (!_jobInfoPool.Contains(handle._id, handle._generation))
+                if (!_jobInfoPool.Contains(handle.ID, handle.Generation))
                 {
                     completedCount++;
                 }
@@ -568,26 +674,25 @@ public sealed unsafe partial class JobScheduler : IJobScheduler, IDisposable
                 return;
             }
 
-            spin.SpinOnce(sleepThreshold);
+            spin.SpinOnce(_SLEEP_THRESHOLD);
         }
     }
 
     public JobHandle WaitAny(params ReadOnlySpan<JobHandle> handles)
     {
-        var sleepThreshold = _SLEEP_THRESHOLD * handles.Length;
         var spin = new SpinWait();
 
         while (true)
         {
             foreach (var handle in handles)
             {
-                if (!_jobInfoPool.Contains(handle._id, handle._generation))
+                if (!_jobInfoPool.Contains(handle.ID, handle.Generation))
                 {
                     return handle;
                 }
             }
 
-            spin.SpinOnce(sleepThreshold);
+            spin.SpinOnce(_SLEEP_THRESHOLD);
         }
     }
 

Misaki.HighPerformance.Jobs/WorkerThread.cs

@@ -1,9 +1,11 @@
-using System.Collections.Concurrent;
+using System.Collections.Concurrent;
 
 namespace Misaki.HighPerformance.Jobs;
 
 internal class WorkerThread : IDisposable
 {
+    private const int _MAX_STEAL_ATTEMPTS = 8;
+
     private readonly int _index;
     private readonly Thread _thread;
     private readonly ConcurrentQueue<JobHandle> _localQueue;
@@ -29,54 +31,55 @@ internal class WorkerThread : IDisposable
 
     public void Start() => _thread.Start();
 
-    private JobHandle FindJob()
+    private bool TryFindJob(out JobHandle handle)
     {
-        var handle = JobHandle.Invalid;
+        // 1. Check own local queue first
-        if (_localQueue.TryDequeue(out handle)
+        if (_localQueue.TryDequeue(out handle))
-            || _scheduler.TryStealJob(-1, out handle))
         {
-            return handle;
+            return true;
         }
 
-        while (true)
+        // 2. Check global queue
+        if (_scheduler.TryStealJob(-1, out handle))
+        {
+            return true;
+        }
+
+        // 3. Bounded random work stealing from other workers
+        for (var i = 0; i < _MAX_STEAL_ATTEMPTS; i++)
         {
             var randomIndex = _random.Next(0, _scheduler.WorkerCount);
-            if (_scheduler.TryStealJob(randomIndex, out handle))
+            if (randomIndex != _index && _scheduler.TryStealJob(randomIndex, out handle))
             {
-                return handle;
+                return true;
             }
         }
+
+        handle = JobHandle.Invalid;
+        return false;
     }
 
     private unsafe void WorkLoop()
     {
         while (!_scheduler.IsCancellationRequested)
         {
-            var spinner = new SpinWait();
+            // Wait for work signal directly — the semaphore already acts as
-            for (var i = 0; i < 25; i++)
+            // both a notification and a count of available work items.
-            {
-                spinner.SpinOnce(-1);
-
-                if (_scheduler.HasWork())
-                {
-                    // Instead of goto, we still need to go through the WaitForWork to claim a release.
-                    // This causes lock and lots of branches inside the SemaphoreSlim, which lost 0.03ms.
-                    // goto DoWork;
-                    break;
-                }
-            }
-
             try
             {
                 _scheduler.WaitForWork();
             }
             catch (OperationCanceledException)
+            {
+                break;
+            }
+
+            // After being signaled, try to find and execute a job.
+            if (!TryFindJob(out var handle))
             {
                 continue;
             }
 
-            //DoWork:
-            var handle = FindJob();
             ref var jobInfo = ref _scheduler.GetJobInfoReference(handle, out var exist);
 
             if (exist)

Misaki.HighPerformance.Mathematics.SPMD/IJobSPMD.cs

@@ -0,0 +1,80 @@
+using Misaki.HighPerformance.Jobs;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+
+namespace Misaki.HighPerformance.Mathematics.SPMD;
+
+public interface IJobSPMD<TNumber>
+    where TNumber : unmanaged, INumber<TNumber>, IMinMaxValue<TNumber>, IBitwiseOperators<TNumber, TNumber, TNumber>
+{
+    void Execute<TLane>(int baseIndex, int threadIndex)
+        where TLane : ISPMD<TLane, TNumber>;
+}
+
+internal struct SPMDJobWrapper<T, TNumber> : IJobParallelFor
+    where T : unmanaged, IJobSPMD<TNumber>
+    where TNumber : unmanaged, INumber<TNumber>, IMinMaxValue<TNumber>, IBitwiseOperators<TNumber, TNumber, TNumber>
+{
+    public T innerJob;
+    public int totalCount;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public void Execute(int loopIndex, int threadIndex)
+    {
+        var baseIndex = loopIndex * WideLane<TNumber>.LaneWidth;
+        var remaining = totalCount - baseIndex;
+
+        if (remaining >= WideLane<TNumber>.LaneWidth)
+        {
+            innerJob.Execute<WideLane<TNumber>>(baseIndex, threadIndex);
+        }
+        else
+        {
+            for (var i = 0; i < remaining; i++)
+            {
+                innerJob.Execute<ScalarLane<TNumber>>(baseIndex + i, threadIndex);
+            }
+        }
+    }
+}
+
+public static class IJobParallelForSPMDExtensions
+{
+    public static void Run<T, TNumber>(this ref T job, int totalCount, int threadIndex)
+        where T : unmanaged, IJobSPMD<TNumber>
+        where TNumber : unmanaged, INumber<TNumber>, IMinMaxValue<TNumber>, IBitwiseOperators<TNumber, TNumber, TNumber>
+    {
+        var iterations = (totalCount + WideLane<TNumber>.LaneWidth - 1) / WideLane<TNumber>.LaneWidth;
+        for (var loopIndex = 0; loopIndex < iterations; loopIndex++)
+        {
+            var baseIndex = loopIndex * WideLane<TNumber>.LaneWidth;
+            var remaining = totalCount - baseIndex;
+
+            if (remaining >= WideLane<TNumber>.LaneWidth)
+            {
+                job.Execute<WideLane<TNumber>>(baseIndex, threadIndex);
+            }
+            else
+            {
+                for (var i = 0; i < remaining; i++)
+                {
+                    job.Execute<ScalarLane<TNumber>>(baseIndex + i, threadIndex);
+                }
+            }
+        }
+    }
+
+    public static JobHandle ScheduleParallelSPDM<T, TNumber>(this JobScheduler jobScheduler, ref T job, int totalCount, int batchSize, int threadIndex, JobHandle dependency)
+        where T : unmanaged, IJobSPMD<TNumber>
+        where TNumber : unmanaged, INumber<TNumber>, IMinMaxValue<TNumber>, IBitwiseOperators<TNumber, TNumber, TNumber>
+    {
+        var warper = new SPMDJobWrapper<T, TNumber>
+        {
+            innerJob = job,
+            totalCount = totalCount,
+        };
+
+        var iterations = (totalCount + WideLane<TNumber>.LaneWidth - 1) / WideLane<TNumber>.LaneWidth;
+        return jobScheduler.ScheduleParallel(ref warper, iterations, batchSize, threadIndex, dependency);
+    }
+}

Misaki.HighPerformance.Mathematics.SPMD/ISPMD.cs

@@ -0,0 +1,589 @@
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using static System.Runtime.InteropServices.JavaScript.JSType;
+
+namespace Misaki.HighPerformance.Mathematics.SPMD;
+
+/// <summary>
+/// Common marker interface for SPMD lane types.
+/// </summary>
+public interface ISPMD
+{
+    /// <summary>
+    /// Gets the number of lanes (vector width) for the SPMD implementation.
+    /// </summary>
+    static abstract int LaneWidth
+    {
+        get;
+    }
+}
+
+/// <summary>
+/// Represents a single-lane or multi-lane (vectorized) SPMD value and the operations supported on it.
+/// </summary>
+/// <typeparam name="TSelf">The concrete SPMD lane type implementing this interface.</typeparam>
+/// <typeparam name="TNumber">The underlying numeric element type.</typeparam>
+public interface ISPMD<TSelf, TNumber> : ISPMD
+    where TSelf : ISPMD<TSelf, TNumber>
+    where TNumber : unmanaged, INumber<TNumber>, IMinMaxValue<TNumber>, IBitwiseOperators<TNumber, TNumber, TNumber>
+{
+    /// <summary>
+    /// Gets a lane value where all lanes are set to numeric zero.
+    /// </summary>
+    static abstract TSelf Zero
+    {
+        get;
+    }
+
+    /// <summary>
+    /// Gets a lane value where all lanes are set to numeric one.
+    /// </summary>
+    static abstract TSelf One
+    {
+        get;
+    }
+
+    /// <summary>
+    /// Gets a lane value where all lanes are set to the minimum representable value of the underlying numeric type.
+    /// </summary>
+    static abstract TSelf MinValue
+    {
+        get;
+    }
+
+    /// <summary>
+    /// Gets a lane value where all lanes are set to the maximum representable value of the underlying numeric type.
+    /// </summary>
+    static abstract TSelf MaxValue
+    {
+        get;
+    }
+
+    /// <summary>
+    /// Gets the element value for the specified lane index.
+    /// </summary>
+    /// <param name="index">The zero-based lane index.</param>
+    TNumber this[int index]
+    {
+        get;
+    }
+
+    /// <summary>
+    /// Creates a lane value where all lanes are set to the specified value.
+    /// </summary>
+    /// <param name="value">The value to set for all lanes.</param>
+    /// <returns>The created lane value.</returns>
+    static abstract TSelf Create(TNumber value);
+    /// <summary>
+    /// Creates a new instance of the type from the specified sequence of numeric values.
+    /// </summary>
+    /// <param name="values">A parameter array of read-only spans containing the numeric values to use for initialization.</param>
+    /// <returns>A new instance of the type initialized with the provided numeric values.</returns>
+    static abstract TSelf Create(params ReadOnlySpan<TNumber> values);
+    /// <summary>
+    /// Creates a lane value from the specified vector.
+    /// </summary>
+    /// <param name="value">The vector to create the lane value from.</param>
+    /// <returns>The lane value built from the vector.</returns>
+    static abstract TSelf Create(Vector<TNumber> value);
+    /// <summary>
+    /// Creates a lane value with a sequence starting from the specified value with the given step.
+    /// </summary>
+    /// <param name="start">The starting value.</param>
+    /// <param name="step">The step value for the sequence.</param>
+    /// <returns>The lane value containing the arithmetic sequence.</returns>
+    /// <remarks>
+    /// Implementations may rely on vector creation helpers and assume that the resulting sequence length matches <see cref="LaneWidth"/>.
+    /// </remarks>
+    static abstract TSelf Sequence(TNumber start, TNumber step);
+    /// <summary>
+    /// Loads a lane value from the specified reference.
+    /// </summary>
+    /// <param name="value">The reference to load from.</param>
+    /// <returns>The loaded lane value.</returns>
+    static abstract TSelf Load(ref TNumber value);
+    /// <summary>
+    /// Loads a lane value from the specified pointer.
+    /// </summary>
+    /// <param name="pValue">The pointer to load from.</param>
+    /// <returns>The loaded lane value.</returns>
+    /// <remarks>
+    /// Unsafe pointer overloads are provided for scenarios where sequential lane data is already contiguous in memory.
+    /// </remarks>
+    static abstract unsafe TSelf Load(TNumber* pValue);
+
+    /// <summary>
+    /// Stores the lane value to the specified reference.
+    /// </summary>
+    /// <param name="destination">The reference to store to.</param>
+    void Store(ref TNumber destination);
+    /// <summary>
+    /// Stores the lane value to the specified pointer.
+    /// </summary>
+    /// <param name="pDestination">The pointer to store to.</param>
+    unsafe void Store(TNumber* pDestination);
+    /// <summary>
+    /// Compresses the data specified by the given mask and stores the compressed result in the provided destination
+    /// variable.
+    /// </summary>
+    /// <param name="mask">A mask value that determines which elements are included in the compression operation.</param>
+    /// <param name="destination">A reference to the variable where the compressed data will be stored.</param>
+    /// <returns>The number of elements written to the destination as a result of the compression. Returns 0 if no elements are compressed.</returns>
+    /// <remarks>
+    /// Implementations may use hardware-specific shuffle tables to reorder the selected lanes before storing, falling back to a scalar loop otherwise.
+    /// </remarks>
+    int CompressStore(TSelf mask, ref TNumber destination);
+    /// <summary>
+    /// Compresses the data specified by the given mask and stores the compressed result in the provided destination
+    /// variable.
+    /// </summary>
+    /// <param name="mask">A mask value that determines which elements are included in the compression operation.</param>
+    /// <param name="pDestination">A pointer to the variable where the compressed data will be stored.</param>
+    /// <returns>The number of elements written to the destination as a result of the compression. Returns 0 if no elements are compressed.</returns>
+    /// <remarks>
+    /// Implementations may use hardware-specific shuffle tables to reorder the selected lanes before storing, falling back to a scalar loop otherwise.
+    /// </remarks>
+    unsafe int CompressStore(TSelf mask, TNumber* pDestination);
+
+    /// <summary>
+    /// Converts the lane value to a vector.
+    /// </summary>
+    /// <returns>The backing vector representation.</returns>
+    Vector<TNumber> AsVector();
+
+    /// <summary>
+    /// Adds two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane-wise sum.</returns>
+    static abstract TSelf operator +(TSelf a, TSelf b);
+    /// <summary>
+    /// Adds a lane value and a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The lane value with the scalar added to each element.</returns>
+    static abstract TSelf operator +(TSelf a, TNumber b);
+    /// <summary>
+    /// Subtracts two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane-wise difference.</returns>
+    static abstract TSelf operator -(TSelf a, TSelf b);
+    /// <summary>
+    /// Subtracts a scalar from a lane value element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The lane value with the scalar subtracted from each element.</returns>
+    static abstract TSelf operator -(TSelf a, TNumber b);
+    /// <summary>
+    /// Multiplies two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane-wise product.</returns>
+    static abstract TSelf operator *(TSelf a, TSelf b);
+    /// <summary>
+    /// Multiplies a lane value by a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The lane value scaled by the scalar.</returns>
+    static abstract TSelf operator *(TSelf a, TNumber b);
+    /// <summary>
+    /// Divides two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane-wise quotient.</returns>
+    static abstract TSelf operator /(TSelf a, TSelf b);
+    /// <summary>
+    /// Divides a lane value by a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The lane value divided by the scalar.</returns>
+    static abstract TSelf operator /(TSelf a, TNumber b);
+    /// <summary>
+    /// Computes the modulus of two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane-wise modulus.</returns>
+    static abstract TSelf operator %(TSelf a, TSelf b);
+    /// <summary>
+    /// Computes the modulus of a lane value and a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The lane value modulus scalar.</returns>
+    static abstract TSelf operator %(TSelf a, TNumber b);
+
+    /// <summary>
+    /// Negates the lane value element-wise.
+    /// </summary>
+    /// <param name="a">The lane value to negate.</param>
+    /// <returns>The negated lane value.</returns>
+    static abstract TSelf operator -(TSelf a);
+
+    /// <summary>
+    /// Computes the bitwise AND of two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The result of the bitwise AND.</returns>
+    static abstract TSelf operator &(TSelf a, TSelf b);
+    /// <summary>
+    /// Computes the bitwise AND of a lane value and a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The result of the bitwise AND.</returns>
+    static abstract TSelf operator &(TSelf a, TNumber b);
+    /// <summary>
+    /// Computes the bitwise OR of two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The result of the bitwise OR.</returns>
+    static abstract TSelf operator |(TSelf a, TSelf b);
+    /// <summary>
+    /// Computes the bitwise OR of a lane value and a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The result of the bitwise OR.</returns>
+    static abstract TSelf operator |(TSelf a, TNumber b);
+    /// <summary>
+    /// Computes the bitwise XOR of two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The result of the bitwise XOR.</returns>
+    static abstract TSelf operator ^(TSelf a, TSelf b);
+    /// <summary>
+    /// Computes the bitwise XOR of a lane value and a scalar element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <param name="b">The scalar value.</param>
+    /// <returns>The result of the bitwise XOR.</returns>
+    static abstract TSelf operator ^(TSelf a, TNumber b);
+    /// <summary>
+    /// Computes the bitwise NOT of a lane value element-wise.
+    /// </summary>
+    /// <param name="a">The lane value.</param>
+    /// <returns>The bitwise complement of the lane value.</returns>
+    static abstract TSelf operator ~(TSelf a);
+
+    /// <summary>
+    /// Computes the absolute value of the lane value element-wise.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The absolute lane value.</returns>
+    static abstract TSelf Abs(TSelf value);
+    /// <summary>
+    /// Computes the floor of the lane value element-wise.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The lane value with each element rounded toward negative infinity.</returns>
+    static abstract TSelf Floor(TSelf value);
+    /// <summary>
+    /// Computes the fractional part of the lane value element-wise.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The fractional lane value.</returns>
+    static abstract TSelf Frac(TSelf value);
+    /// <summary>
+    /// Computes the square root of the lane value element-wise.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The square root lane value.</returns>
+    static abstract TSelf Sqrt(TSelf value);
+    /// <summary>
+    /// Performs linear interpolation between two lane values.
+    /// </summary>
+    /// <param name="a">The start lane value.</param>
+    /// <param name="b">The end lane value.</param>
+    /// <param name="t">The interpolation factor.</param>
+    /// <returns>The interpolated lane value.</returns>
+    static abstract TSelf Lerp(TSelf a, TSelf b, TSelf t);
+    /// <summary>
+    /// Computes a * b + c element-wise.
+    /// </summary>
+    /// <param name="a">The first multiplier.</param>
+    /// <param name="b">The second multiplier.</param>
+    /// <param name="c">The addend.</param>
+    /// <returns>The result of the fused multiply-add operation.</returns>
+    /// <remarks>
+    /// Float and double implementations should use fused multiply-add instructions when available for both accuracy and performance.
+    /// </remarks>
+    static abstract TSelf MultipleAdd(TSelf a, TSelf b, TSelf c);
+    /// <summary>
+    /// Returns the minimum of the two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane value containing the minimum of each element.</returns>
+    static abstract TSelf Min(TSelf a, TSelf b);
+    /// <summary>
+    /// Returns the maximum of the two lane values element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The lane value containing the maximum of each element.</returns>
+    static abstract TSelf Max(TSelf a, TSelf b);
+    /// <summary>
+    /// Clamps each element of the lane value between the specified minimum and maximum values.
+    /// </summary>
+    /// <param name="value">The lane value to clamp.</param>
+    /// <param name="min">The inclusive minimum.</param>
+    /// <param name="max">The inclusive maximum.</param>
+    /// <returns>The clamped lane value.</returns>
+    static abstract TSelf Clamp(TSelf value, TSelf min, TSelf max);
+    /// <summary>
+    /// Saturates each element in the lane value to the 0..1 range.
+    /// </summary>
+    /// <param name="value">The lane value to saturate.</param>
+    /// <returns>The saturated lane value.</returns>
+    static abstract TSelf Saturate(TSelf value);
+    /// <summary>
+    /// Computes the sine of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The sine of each lane element.</returns>
+    /// <remarks>
+    /// Implementations may rely on vectorized math intrinsics for float/double and approximate values for other types.
+    /// </remarks>
+    static abstract TSelf Sin(TSelf value);
+    /// <summary>
+    /// Computes the cosine of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The cosine of each lane element.</returns>
+    /// <remarks>
+    /// Implementations may rely on vectorized math intrinsics for float/double and approximate values for other types.
+    /// </remarks>
+    static abstract TSelf Cos(TSelf value);
+    /// <summary>
+    /// Computes both sine and cosine of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>A tuple containing sine and cosine lane values.</returns>
+    /// <remarks>
+    /// Implementations returning both sin and cos simultaneously can reuse intermediate values for better performance.
+    /// </remarks>
+    static abstract (TSelf sin, TSelf cos) SinCos(TSelf value);
+    /// <summary>
+    /// Computes the tangent of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The tangent of each lane element.</returns>
+    /// <remarks>
+    /// Many implementations use polynomial approximations and assume the input is reduced to [-pi/4, pi/4] for accuracy.
+    /// </remarks>
+    static abstract TSelf Tan(TSelf value);
+    /// <summary>
+    /// Computes the arcsine of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The arcsine of each lane element.</returns>
+    /// <remarks>
+    /// Implementations typically assume input is within [-1, 1] and may use polynomial approximations for performance.
+    /// </remarks>
+    static abstract TSelf Asin(TSelf value);
+    /// <summary>
+    /// Computes the arccosine of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The arccosine of each lane element.</returns>
+    /// <remarks>
+    /// Input is expected to be in [-1, 1]; implementations often rely on approximation polynomials combined with range reduction.
+    /// </remarks>
+    static abstract TSelf Acos(TSelf value);
+    /// <summary>
+    /// Computes the arctangent of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The arctangent of each lane element.</returns>
+    /// <remarks>
+    /// Polynomial approximations with restricted input ranges are commonly used for performance-sensitive implementations.
+    /// </remarks>
+    static abstract TSelf Atan(TSelf value);
+    /// <summary>
+    /// Computes the arctangent of y/x for each lane element.
+    /// </summary>
+    /// <param name="y">The numerator lane value.</param>
+    /// <param name="x">The denominator lane value.</param>
+    /// <returns>The arctangent of each lane pair.</returns>
+    /// <remarks>
+    /// Implementations often rely on quadrant-aware polynomial routines and assume inputs are finite to avoid NaNs.
+    /// </remarks>
+    static abstract TSelf Atan2(TSelf y, TSelf x);
+    /// <summary>
+    /// Raises each lane element to the specified power.
+    /// </summary>
+    /// <param name="x">The base lane value.</param>
+    /// <param name="y">The exponent lane value. Cannot be negative.</param>
+    /// <returns>The power result for each lane.</returns>
+    static abstract TSelf Pow(TSelf x, TSelf y);
+    /// <summary>
+    /// Computes the exponential of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The exponential of each lane element.</returns>
+    /// <remarks>
+    /// Float and double implementations typically call into vectorized exp intrinsics; other types may fall back to scalar paths.
+    /// </remarks>
+    static abstract TSelf Exp(TSelf value);
+    /// <summary>
+    /// Computes 2 raised to each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The base-2 exponential of each lane element.</returns>
+    /// <remarks>
+    /// This can be implemented via <see cref="Exp(TSelf)"/> when no dedicated base-2 intrinsic exists.
+    /// </remarks>
+    static abstract TSelf Exp2(TSelf value);
+    /// <summary>
+    /// Computes the natural logarithm of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The natural logarithm of each lane element.</returns>
+    /// <remarks>
+    /// Vectorized logarithm instructions may only exist for floating-point types; other types should mimic the scalar behavior.
+    /// </remarks>
+    static abstract TSelf Log(TSelf value);
+    /// <summary>
+    /// Computes the base-2 logarithm of each lane element.
+    /// </summary>
+    /// <param name="value">The source lane value.</param>
+    /// <returns>The base-2 logarithm of each lane element.</returns>
+    /// <remarks>
+    /// If a dedicated base-2 intrinsic is unavailable, the implementation may compute <c>Log(value)/Log(2)</c>.
+    /// </remarks>
+    static abstract TSelf Log2(TSelf value);
+    /// <summary>
+    /// Computes the ceiling of each lane element.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The smallest integral value greater than or equal to each element.</returns>
+    /// <remarks>
+    /// Implementations should use <see cref="Vector"/> helpers for floating-point types when available.
+    /// </remarks>
+    static abstract TSelf Ceil(TSelf value);
+    /// <summary>
+    /// Rounds each lane element to the nearest integer value.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The rounded lane value.</returns>
+    /// <remarks>
+    /// Implementations should prefer vectorized round intrinsics for floating-point implementations.
+    /// </remarks>
+    static abstract TSelf Round(TSelf value);
+    /// <summary>
+    /// Truncates each lane element toward zero.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The truncated lane value.</returns>
+    /// <remarks>
+    /// Floating-point truncation typically maps to <see cref="Vector.Truncate(Vector{TNumber})"/>.
+    /// </remarks>
+    static abstract TSelf Trunc(TSelf value);
+    /// <summary>
+    /// Returns the sign of each lane element.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>-1, 0, or 1 per lane.</returns>
+    static abstract TSelf Sign(TSelf value);
+    /// <summary>
+    /// Copies the sign of the second lane value to the magnitude of the first.
+    /// </summary>
+    /// <param name="magnitude">The magnitude lane value.</param>
+    /// <param name="sign">The sign lane value.</param>
+    /// <returns>The result of merging magnitude with sign.</returns>
+    static abstract TSelf CopySign(TSelf magnitude, TSelf sign);
+    /// <summary>
+    /// Computes the reciprocal of each lane element.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The reciprocal lane value.</returns>
+    /// <remarks>
+    /// Fast paths may use <c>Sse.Reciprocal</c> or <c>Avx.Reciprocal</c> when <c>TNumber</c> is <c>float</c>.
+    /// </remarks>
+    static abstract TSelf Rcp(TSelf value);
+    /// <summary>
+    /// Computes the reciprocal square root of each lane element.
+    /// </summary>
+    /// <param name="x">The lane value.</param>
+    /// <returns>The reciprocal square root lane value.</returns>
+    /// <remarks>
+    /// Float implementations may prefer hardware reciprocal-sqrt intrinsics and fallback to <c>Create(TNumber.One)/Sqrt(x)</c> otherwise.
+    /// </remarks>
+    static abstract TSelf Rsqrt(TSelf value);
+
+    /// <summary>
+    /// Selects values from two lane values based on a condition mask.
+    /// </summary>
+    /// <param name="conditionMask">The condition mask.</param>
+    /// <param name="ifTrue">The value to select if true.</param>
+    /// <param name="ifFalse">The value to select if false.</param>
+    /// <returns>The selected lane value.</returns>
+    static abstract TSelf Select(TSelf conditionMask, TSelf ifTrue, TSelf ifFalse);
+    /// <summary>
+    /// Compares two lane values for greater than element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The mask representing the greater than comparison result.</returns>
+    static abstract TSelf GreaterThan(TSelf a, TSelf b);
+    /// <summary>
+    /// Compares two lane values for greater than or equal element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The mask representing the greater than or equal comparison result.</returns>
+    static abstract TSelf GreaterThanOrEqual(TSelf a, TSelf b);
+    /// <summary>
+    /// Compares two lane values for less than element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The mask representing the less than comparison result.</returns>
+    static abstract TSelf LessThan(TSelf a, TSelf b);
+    /// <summary>
+    /// Compares two lane values for less than or equal element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The mask representing the less than or equal comparison result.</returns>
+    static abstract TSelf LessThanOrEqual(TSelf a, TSelf b);
+    /// <summary>
+    /// Compares two lane values for equality element-wise.
+    /// </summary>
+    /// <param name="a">The first lane value.</param>
+    /// <param name="b">The second lane value.</param>
+    /// <returns>The mask representing the equality comparison result.</returns>
+    static abstract TSelf Equal(TSelf a, TSelf b);
+
+    /// <summary>
+    /// Checks if any lane in the mask is true.
+    /// </summary>
+    /// <param name="mask">The mask to check.</param>
+    /// <returns>True if any lane is true; otherwise, false.</returns>
+    static abstract bool Any(TSelf mask);
+    /// <summary>
+    /// Checks if all lanes in the mask are true.
+    /// </summary>
+    /// <param name="mask">The mask to check.</param>
+    /// <returns>True if all lanes are true; otherwise, false.</returns>
+    static abstract bool All(TSelf mask);
+    /// <summary>
+    /// Checks if no lanes in the mask are true.
+    /// </summary>
+    /// <param name="mask">The mask to check.</param>
+    /// <returns>True if no lanes are true; otherwise, false.</returns>
+    static abstract bool None(TSelf mask);
+}

Misaki.HighPerformance.Mathematics.SPMD/Misaki.HighPerformance.Mathematics.SPMD.csproj

@@ -0,0 +1,14 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <TargetFramework>net10.0</TargetFramework>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+    <AllowUnsafeBlocks>true</AllowUnsafeBlocks>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="..\Misaki.HighPerformance.Jobs\Misaki.HighPerformance.Jobs.csproj" />
+  </ItemGroup>
+
+</Project>

Misaki.HighPerformance.Mathematics.SPMD/ScalerLane.cs

@@ -0,0 +1,564 @@
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace Misaki.HighPerformance.Mathematics.SPMD;
+
+[StructLayout(LayoutKind.Sequential)]
+public readonly unsafe struct ScalarLane<T> : ISPMD<ScalarLane<T>, T>
+    where T : unmanaged, INumber<T>, IMinMaxValue<T>, IBitwiseOperators<T, T, T>
+{
+    public readonly T value;
+
+    public static int LaneWidth
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => 1;
+    }
+
+    public static ScalarLane<T> Zero
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(T.Zero);
+    }
+
+    public static ScalarLane<T> One
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(T.One);
+    }
+
+    public static ScalarLane<T> MinValue
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(T.MinValue);
+    }
+
+    public static ScalarLane<T> MaxValue
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(T.MaxValue);
+    }
+
+    public readonly T this[int index]
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => value;
+    }
+
+    public ScalarLane(T value)
+    {
+        this.value = value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Create(T value) => new(value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Create(params ReadOnlySpan<T> values) => new(values[0]);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Create(Vector<T> value) => new(value[0]);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Sequence(T start, T step) => new(start);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Load(ref T value) => new(value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Load(T* pValue) => new(*pValue);
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly void Store(ref T destination) => destination = value;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly void Store(T* pDestination) => *pDestination = value;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public int CompressStore(ScalarLane<T> mask, ref T destination)
+    {
+        return CompressStore(mask, (T*)Unsafe.AsPointer(in destination));
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public int CompressStore(ScalarLane<T> mask, T* pDestination)
+    {
+        if (mask.value != T.Zero)
+        {
+            *pDestination = value;
+            return 1;
+        }
+
+        return 0;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly Vector<T> AsVector() => Vector.Create(value);
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator +(ScalarLane<T> a, ScalarLane<T> b) => new(a.value + b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator +(ScalarLane<T> a, T b) => new(a.value + b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator -(ScalarLane<T> a, ScalarLane<T> b) => new(a.value - b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator -(ScalarLane<T> a, T b) => new(a.value - b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator *(ScalarLane<T> a, ScalarLane<T> b) => new(a.value * b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator *(ScalarLane<T> a, T b) => new(a.value * b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator /(ScalarLane<T> a, ScalarLane<T> b) => new(a.value / b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator /(ScalarLane<T> a, T b) => new(a.value / b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator %(ScalarLane<T> a, ScalarLane<T> b) => new(a.value % b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator %(ScalarLane<T> a, T b) => new(a.value % b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator -(ScalarLane<T> a) => new(-a.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator &(ScalarLane<T> a, ScalarLane<T> b) => new(a.value & b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator &(ScalarLane<T> a, T b) => new(a.value & b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator |(ScalarLane<T> a, ScalarLane<T> b) => new(a.value | b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator |(ScalarLane<T> a, T b) => new(a.value | b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator ^(ScalarLane<T> a, ScalarLane<T> b) => new(a.value ^ b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator ^(ScalarLane<T> a, T b) => new(a.value ^ b);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> operator ~(ScalarLane<T> a) => new(~a.value);
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Abs(ScalarLane<T> value) => new(T.Abs(value.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Floor(ScalarLane<T> value)
+    {
+        // Note: INumber<T> does not provide Floor method, so we need to handle float and double specifically.
+        // This is acceptable for performance because JIT generates specialized code for each T as long as they are struct.
+        // Which mean for ScalarLane<float>, typeof(T) == typeof(float) is always true and jit will optimize away the other branches.
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Floor(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Floor(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Frac(ScalarLane<T> value) => new(value.value - T.CreateTruncating(value.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Sqrt(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Sqrt(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Sqrt(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Lerp(ScalarLane<T> a, ScalarLane<T> b, ScalarLane<T> t) => new(a.value + (b.value - a.value) * t.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> MultipleAdd(ScalarLane<T> a, ScalarLane<T> b, ScalarLane<T> c) => new(T.MultiplyAddEstimate(a.value, b.value, c.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Min(ScalarLane<T> a, ScalarLane<T> b) => new(T.Min(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Max(ScalarLane<T> a, ScalarLane<T> b) => new(T.Max(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Clamp(ScalarLane<T> value, ScalarLane<T> min, ScalarLane<T> max) => new(T.Clamp(value.value, min.value, max.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Saturate(ScalarLane<T> value) => Clamp(value, new(T.Zero), new(T.One));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Sin(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Sin(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Sin(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Cos(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Cos(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Cos(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static (ScalarLane<T> sin, ScalarLane<T> cos) SinCos(ScalarLane<T> value) => (Sin(value), Cos(value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Tan(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Tan(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Tan(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Asin(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Asin(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Asin(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Acos(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Acos(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Acos(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Atan(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Atan(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Atan(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Atan2(ScalarLane<T> y, ScalarLane<T> x)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var fy = Unsafe.As<ScalarLane<T>, float>(ref y);
+            var fx = Unsafe.As<ScalarLane<T>, float>(ref x);
+            var result = MathF.Atan2(fy, fx);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var dy = Unsafe.As<ScalarLane<T>, double>(ref y);
+            var dx = Unsafe.As<ScalarLane<T>, double>(ref x);
+            var result = Math.Atan2(dy, dx);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return y;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Pow(ScalarLane<T> x, ScalarLane<T> y)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var fx = Unsafe.As<ScalarLane<T>, float>(ref x);
+            var fy = Unsafe.As<ScalarLane<T>, float>(ref y);
+            var result = MathF.Pow(fx, fy);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var dx = Unsafe.As<ScalarLane<T>, double>(ref x);
+            var dy = Unsafe.As<ScalarLane<T>, double>(ref y);
+            var result = Math.Pow(dx, dy);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return x;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Exp(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Exp(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Log(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Exp2(ScalarLane<T> value)
+    {
+        return Pow(new ScalarLane<T>(T.CreateChecked(2)), value);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Log(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Log(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Log(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Log2(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Log2(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Log2(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Ceil(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Ceiling(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Ceiling(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(decimal))
+        {
+            var d = Unsafe.As<ScalarLane<T>, decimal>(ref value);
+            var result = Math.Ceiling(d);
+            return Unsafe.As<decimal, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Round(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Round(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Round(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(decimal))
+        {
+            var d = Unsafe.As<ScalarLane<T>, decimal>(ref value);
+            var result = Math.Round(d);
+            return Unsafe.As<decimal, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Trunc(ScalarLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            var f = Unsafe.As<ScalarLane<T>, float>(ref value);
+            var result = MathF.Truncate(f);
+            return Unsafe.As<float, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            var d = Unsafe.As<ScalarLane<T>, double>(ref value);
+            var result = Math.Truncate(d);
+            return Unsafe.As<double, ScalarLane<T>>(ref result);
+        }
+        else if (typeof(T) == typeof(decimal))
+        {
+            var d = Unsafe.As<ScalarLane<T>, decimal>(ref value);
+            var result = Math.Truncate(d);
+            return Unsafe.As<decimal, ScalarLane<T>>(ref result);
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Sign(ScalarLane<T> value) => new((value.value > T.Zero) ? T.One : (value.value < T.Zero) ? ~T.Zero : T.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> CopySign(ScalarLane<T> magnitude, ScalarLane<T> sign) => new(T.CopySign(magnitude.value, sign.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Rcp(ScalarLane<T> value) => new(T.One / value.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Rsqrt(ScalarLane<T> value) => Sqrt(Rcp(value));
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Select(ScalarLane<T> conditionMask, ScalarLane<T> ifTrue, ScalarLane<T> ifFalse) => new(conditionMask.value != T.Zero ? ifTrue.value : ifFalse.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> GreaterThan(ScalarLane<T> a, ScalarLane<T> b) => new(a.value > b.value ? ~T.Zero : T.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> GreaterThanOrEqual(ScalarLane<T> a, ScalarLane<T> b) => new(a.value >= b.value ? ~T.Zero : T.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> LessThan(ScalarLane<T> a, ScalarLane<T> b) => new(a.value < b.value ? ~T.Zero : T.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> LessThanOrEqual(ScalarLane<T> a, ScalarLane<T> b) => new(a.value <= b.value ? ~T.Zero : T.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static ScalarLane<T> Equal(ScalarLane<T> a, ScalarLane<T> b) => new(a.value == b.value ? ~T.Zero : T.Zero);
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool Any(ScalarLane<T> mask) => mask.value != T.Zero;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool All(ScalarLane<T> mask) => mask.value != T.Zero;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool None(ScalarLane<T> mask) => mask.value == T.Zero;
+
+    public override string ToString()
+    {
+        return value.ToString() ?? string.Empty;
+    }
+}

Misaki.HighPerformance.Mathematics.SPMD/ShuffleTableGenerator.cs

@@ -0,0 +1,206 @@
+using System.Runtime.InteropServices;
+
+namespace Misaki.HighPerformance.Mathematics.SPMD;
+
+public static unsafe class ShuffleTableGenerator
+{
+    public static uint* ComputeShuffleTable512_32Bit()
+    {
+        const nuint entryCount = 512;
+        const int elementCount = 16;
+
+        // Align to 64 bytes for AVX-512 performance
+        var table = (uint*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(uint), 64);
+
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            // We are filling 16 integers for this mask
+            var pRow = table + (mask * elementCount);
+            var outputIndex = 0;
+            // 1. Pack the valid indices to the front
+            for (var bit = 0; bit < 16; bit++)
+            {
+                // Check if the i-th bit is set
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (uint)bit; // Write the Source Index
+                    outputIndex++;
+                }
+            }
+            // 2. Fill the remaining slots (Pad with 0 or similar)
+            // It doesn't strictly matter what these are, as we won't read them,
+            // but filling with 0 is clean.
+            while (outputIndex < 16)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+
+    public static ulong* ComputeShuffleTable512_64Bit()
+    {
+        const nuint entryCount = 256;
+        const int elementCount = 8;
+
+        // Align to 64 bytes for AVX-512 performance
+        var table = (ulong*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(ulong), 64);
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            // We are filling 8 integers for this mask
+            var pRow = table + (mask * elementCount);
+            var outputIndex = 0;
+
+            // 1. Pack the valid indices to the front
+            for (var bit = 0; bit < 8; bit++)
+            {
+                // Check if the i-th bit is set
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (ulong)bit; // Write the Source Index
+                    outputIndex++;
+                }
+            }
+
+            // 2. Fill the remaining slots (Pad with 0 or similar)
+            while (outputIndex < 8)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+
+    public static uint* ComputeShuffleTable256_32Bit()
+    {
+        const nuint entryCount = 256;
+        const nuint elementCount = 8;
+
+        // Align to 32 bytes for AVX performance
+        var table = (uint*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(uint), 32);
+
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            // We are filling 8 integers for this mask
+            var pRow = table + (mask * elementCount);
+
+            var outputIndex = 0;
+
+            for (var bit = 0; bit < 8; bit++)
+            {
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (uint)bit;
+                    outputIndex++;
+                }
+            }
+
+            while (outputIndex < 8)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+
+    public static ulong* ComputeShuffleTable256_64Bit()
+    {
+        const nuint entryCount = 16;
+        const nuint elementCount = 4;
+
+        var table = (ulong*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(ulong), 32);
+
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            var pRow = table + (mask * elementCount);
+            var outputIndex = 0;
+
+            // We only check 4 bits because there are only 4 ulongs in a Vector256
+            for (var bit = 0; bit < 4; bit++)
+            {
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (ulong)bit;
+                    outputIndex++;
+                }
+            }
+
+            // Fill remaining slots with 0 (or a specific 'clear' index)
+            while (outputIndex < 4)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+
+    public static uint* ComputeShuffleTable128_32Bit()
+    {
+        const nuint entryCount = 16;
+        const nuint elementCount = 4;
+
+        var table = (uint*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(uint), 16);
+
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            var pRow = table + (mask * elementCount);
+            var outputIndex = 0;
+
+            for (var bit = 0; bit < 4; bit++)
+            {
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (uint)bit;
+                    outputIndex++;
+                }
+            }
+
+            while (outputIndex < 4)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+
+    public static ulong* ComputeShuffleTable128_64Bit()
+    {
+        const nuint entryCount = 8;
+        const nuint elementCount = 2;
+
+        var table = (ulong*)NativeMemory.AlignedAlloc(entryCount * elementCount * sizeof(ulong), 16);
+
+        for (var mask = 0u; mask < entryCount; mask++)
+        {
+            var pRow = table + (mask * elementCount);
+            var outputIndex = 0;
+
+            for (var bit = 0; bit < 2; bit++)
+            {
+                if ((mask & (1 << bit)) != 0)
+                {
+                    pRow[outputIndex] = (byte)bit;
+                    outputIndex++;
+                }
+            }
+
+            while (outputIndex < 2)
+            {
+                pRow[outputIndex] = 0;
+                outputIndex++;
+            }
+        }
+
+        return table;
+    }
+}

Misaki.HighPerformance.Mathematics.SPMD/WideLane.cs

@@ -0,0 +1,794 @@
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+
+namespace Misaki.HighPerformance.Mathematics.SPMD;
+
+[StructLayout(LayoutKind.Sequential)]
+public readonly unsafe struct WideLane<T> : ISPMD<WideLane<T>, T>
+    where T : unmanaged, INumber<T>, IMinMaxValue<T>, IBitwiseOperators<T, T, T>
+{
+    private static readonly Vector<T> s_indices;
+
+    private static readonly uint* s_shuffleTable512_32bit;
+    private static readonly ulong* s_shuffleTable512_64bit;
+    private static readonly uint* s_shuffleTable256_32bit;
+    private static readonly ulong* s_shuffleTable256_64bit;
+    private static readonly uint* s_shuffleTable128_32bit;
+    private static readonly ulong* s_shuffleTable128_64bit;
+
+    public readonly Vector<T> value;
+
+    public static int LaneWidth
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => Vector<T>.Count;
+    }
+
+    public static WideLane<T> Zero
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(Vector<T>.Zero);
+    }
+
+    public static WideLane<T> One
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => new(Vector<T>.One);
+    }
+
+    public static WideLane<T> MinValue
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => Create(T.MinValue);
+    }
+
+    public static WideLane<T> MaxValue
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => Create(T.MaxValue);
+    }
+
+    public readonly T this[int index]
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => value[index];
+    }
+
+    static WideLane()
+    {
+        var pValues = stackalloc T[LaneWidth];
+        for (var i = 0; i < LaneWidth; i++)
+        {
+            pValues[i] = T.CreateChecked(i);
+        }
+
+        s_indices = Vector.Load(pValues);
+
+        s_shuffleTable512_32bit = ShuffleTableGenerator.ComputeShuffleTable512_32Bit();
+        s_shuffleTable512_64bit = ShuffleTableGenerator.ComputeShuffleTable512_64Bit();
+        s_shuffleTable256_32bit = ShuffleTableGenerator.ComputeShuffleTable256_32Bit();
+        s_shuffleTable256_64bit = ShuffleTableGenerator.ComputeShuffleTable256_64Bit();
+        s_shuffleTable128_32bit = ShuffleTableGenerator.ComputeShuffleTable128_32Bit();
+        s_shuffleTable128_64bit = ShuffleTableGenerator.ComputeShuffleTable128_64Bit();
+    }
+
+    public WideLane(Vector<T> value)
+    {
+        this.value = value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static Vector<T> VectorFloor(Vector<T> vector)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<Vector<T>, Vector<float>>(ref vector);
+            var floored = Vector.Floor(v);
+            return Unsafe.As<Vector<float>, Vector<T>>(ref floored);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<Vector<T>, Vector<double>>(ref vector);
+            var floored = Vector.Floor(v);
+            return Unsafe.As<Vector<double>, Vector<T>>(ref floored);
+        }
+
+        return vector;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static Vector<T> VectorTruncate(Vector<T> vector)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<Vector<T>, Vector<float>>(ref vector);
+            var truncated = Vector.Truncate(v);
+            return Unsafe.As<Vector<float>, Vector<T>>(ref truncated);
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<Vector<T>, Vector<double>>(ref vector);
+            var truncated = Vector.Truncate(v);
+            return Unsafe.As<Vector<double>, Vector<T>>(ref truncated);
+        }
+
+        return vector;
+    }
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Create(T value) => new(Vector.Create(value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Create(params ReadOnlySpan<T> values) => new(Vector.Create(values));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Create(Vector<T> value) => new(value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Sequence(T start, T step) => new(Vector.Create(start) + (Vector.Create(step) * s_indices));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Load(ref T value) => new(Vector.LoadUnsafe(ref value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Load(T* pValue) => new(Vector.Load(pValue));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> CastFrom<U>(WideLane<U> value)
+        where U : unmanaged, INumber<U>, IMinMaxValue<U>, IBitwiseOperators<U, U, U>
+    {
+        return new(Unsafe.As<WideLane<U>, Vector<T>>(ref value));
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly void Store(ref T destination) => value.StoreUnsafe(ref destination);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly void Store(T* pDestination) => value.Store(pDestination);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public int CompressStore(WideLane<T> mask, ref T destination)
+    {
+        return CompressStore(mask, (T*)Unsafe.AsPointer(in destination));
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public int CompressStore(WideLane<T> mask, T* pDestination)
+    {
+        var size = sizeof(T);
+
+        if (LaneWidth == Vector512<T>.Count && Vector512.IsHardwareAccelerated)
+        {
+            if (size == 4)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector512<uint>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector512<uint>>(ref mask);
+
+                var moveMask = m.ExtractMostSignificantBits();
+                // Offset is (moveMask * 16) because each control vector has 16 elements
+                var shuffle = Vector512.Load(s_shuffleTable512_32bit + (moveMask * 16));
+                var compressed = Vector512.Shuffle(vec, shuffle);
+
+                compressed.Store((uint*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+
+            if (size == 8)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector512<ulong>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector512<ulong>>(ref mask);
+
+                var moveMask = m.ExtractMostSignificantBits();
+                // Offset is (moveMask * 8) because each control vector has 8 elements
+                var shuffle = Vector512.Load(s_shuffleTable512_64bit + (moveMask * 8));
+                var compressed = Vector512.Shuffle(vec, shuffle);
+
+                compressed.Store((ulong*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+        }
+        else if (LaneWidth == Vector256<T>.Count && Vector256.IsHardwareAccelerated)
+        {
+            if (size == 4)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector256<uint>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector256<uint>>(ref mask);
+
+                var moveMask = m.ExtractMostSignificantBits();
+                // Offset is (moveMask * 8) because each control vector has 8 elements
+                var shuffle = Vector256.Load(s_shuffleTable256_32bit + (moveMask * 8));
+                var compressed = Vector256.Shuffle(vec, shuffle);
+
+                compressed.Store((uint*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+
+            if (size == 8)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector256<ulong>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector256<ulong>>(ref mask);
+
+                // For 64-bit, ExtractMostSignificantBits only populates 4 bits (0-15)
+                var moveMask = m.ExtractMostSignificantBits();
+
+                // Offset is (moveMask * 4) because each control vector has 4 elements
+                var shuffle = Vector256.Load(s_shuffleTable256_64bit + (moveMask * 4));
+                var compressed = Vector256.Shuffle(vec, shuffle);
+
+                compressed.Store((ulong*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+        }
+        else if (LaneWidth == Vector128<T>.Count && Vector128.IsHardwareAccelerated)
+        {
+            if (size == 4)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector128<uint>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector128<uint>>(ref mask);
+
+                var moveMask = m.ExtractMostSignificantBits();
+                // Offset is (moveMask * 4) because each control vector has 4 elements
+                var shuffle = Vector128.Load(s_shuffleTable128_32bit + (moveMask * 4));
+                var compressed = Vector128.Shuffle(vec, shuffle);
+
+                compressed.Store((uint*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+
+            if (size == 8)
+            {
+                ref var vec = ref Unsafe.As<WideLane<T>, Vector128<ulong>>(ref Unsafe.AsRef(in this));
+                var m = Unsafe.As<WideLane<T>, Vector128<ulong>>(ref mask);
+                var moveMask = m.ExtractMostSignificantBits();
+                // Offset is (moveMask * 2) because each control vector has 2 elements
+                var shuffle = Vector128.Load(s_shuffleTable128_64bit + (moveMask * 2));
+                var compressed = Vector128.Shuffle(vec, shuffle);
+                compressed.Store((ulong*)pDestination);
+                return BitOperations.PopCount(moveMask);
+            }
+        }
+
+        // This is slow, but correct on ANY hardware.
+        // Check sign bit of the mask lane
+        var count = 0;
+        for (var i = 0; i < LaneWidth; i++)
+        {
+            if (mask.value[i] == ~T.Zero)
+            {
+                pDestination[count++] = value[i];
+            }
+        }
+
+        return count;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public readonly Vector<T> AsVector() => value;
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator +(WideLane<T> a, WideLane<T> b) => new(a.value + b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator +(WideLane<T> a, T b) => new(a.value + Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator -(WideLane<T> a, WideLane<T> b) => new(a.value - b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator -(WideLane<T> a, T b) => new(a.value - Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator *(WideLane<T> a, WideLane<T> b) => new(a.value * b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator *(WideLane<T> a, T b) => new(a.value * Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator /(WideLane<T> a, WideLane<T> b) => new(a.value / b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator /(WideLane<T> a, T b) => new(a.value / Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator %(WideLane<T> a, WideLane<T> b) => new(a.value - VectorFloor(a.value / b.value) * b.value);
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator %(WideLane<T> a, T b)
+    {
+        var vb = Vector.Create(b);
+        return new(a.value - VectorFloor(a.value / vb) * vb);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator -(WideLane<T> a) => new(-a.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator &(WideLane<T> a, WideLane<T> b) => new(a.value & b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator &(WideLane<T> a, T b) => new(a.value & Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator |(WideLane<T> a, WideLane<T> b) => new(a.value | b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator |(WideLane<T> a, T b) => new(a.value | Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator ^(WideLane<T> a, WideLane<T> b) => new(a.value ^ b.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator ^(WideLane<T> a, T b) => new(a.value ^ Vector.Create(b));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> operator ~(WideLane<T> a) => new(~a.value);
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Abs(WideLane<T> value) => new(Vector.Abs(value.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Floor(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var floored = Vector.Floor(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref floored));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var floored = Vector.Floor(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref floored));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Frac(WideLane<T> value) => new(value.value - VectorFloor(value.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Sqrt(WideLane<T> value) => new(Vector.SquareRoot(value.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Lerp(WideLane<T> a, WideLane<T> b, WideLane<T> t) => new(a.value + (b.value - a.value) * t.value);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> MultipleAdd(WideLane<T> a, WideLane<T> b, WideLane<T> c)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var va = ref Unsafe.As<WideLane<T>, Vector<float>>(ref a);
+            ref var vb = ref Unsafe.As<WideLane<T>, Vector<float>>(ref b);
+            ref var vc = ref Unsafe.As<WideLane<T>, Vector<float>>(ref c);
+            var result = Vector.FusedMultiplyAdd(va, vb, vc);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var va = ref Unsafe.As<WideLane<T>, Vector<double>>(ref a);
+            ref var vb = ref Unsafe.As<WideLane<T>, Vector<double>>(ref b);
+            ref var vc = ref Unsafe.As<WideLane<T>, Vector<double>>(ref c);
+            var result = Vector.FusedMultiplyAdd(va, vb, vc);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+        else
+        {
+            return new((a.value * b.value) + c.value);
+        }
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Min(WideLane<T> a, WideLane<T> b) => new(Vector.Min(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Max(WideLane<T> a, WideLane<T> b) => new(Vector.Max(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Clamp(WideLane<T> value, WideLane<T> min, WideLane<T> max) => new(Vector.Clamp(value.value, min.value, max.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Saturate(WideLane<T> value) => Clamp(value, Create(T.Zero), Create(T.One));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Sin(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Sin(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Sin(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Cos(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Cos(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Cos(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static (WideLane<T> sin, WideLane<T> cos) SinCos(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var (sin, cos) = Vector.SinCos(v);
+            return (new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref sin)), new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref cos)));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var (sin, cos) = Vector.SinCos(v);
+            return (new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref sin)), new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref cos)));
+        }
+
+        return (value, value);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Tan(WideLane<T> value)
+    {
+        // 1. Range Reduction
+        // Transform value into range [-pi/4, pi/4]. 
+        // This is complex to do right (Payne-Hanek), but for games
+        // a simple approximation: value = value - (PI * Round(value / PI)) is good enough.
+
+        var pi = Create(T.CreateChecked(Math.PI));
+        var x = value - pi * Round(value / pi);
+
+        // 2. The Approximation (Remez Polynomial)
+        // tan(value) ~= value + c1*value^3 + c2*value^5
+        // Factored (Horner's Method) for fewer ops: value * (1 + value^2 * (c1 + c2*value^2))
+
+        var x2 = x * x;
+        var vc1 = Create(T.CreateChecked(0.3333314036)); // 1/3
+        var vc2 = Create(T.CreateChecked(0.1333923995)); // 2/15
+
+        // x2 * (c1 + c2 * x2)
+        var poly = MultipleAdd(x2, vc2, vc1);
+        // value * (1 + x2 * poly)
+        return MultipleAdd(x, MultipleAdd(x2, poly, One), Zero);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Asin(WideLane<T> value)
+    {
+        // asin(value) = pi/2 - acos(value)
+
+        var piOver2 = Create(T.CreateChecked(Math.PI / 2));
+        return piOver2 - Acos(value);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Acos(WideLane<T> value)
+    {
+        // 0 <= value <= 1 : acos(value) = sqrt(1 - value) * (c0 + c1*value + c2*value^2 + c3*value^3)
+        // value < 0 : acos(value) = pi - acos(-value)
+
+        var x = Abs(value);
+
+        var c0 = Create(T.CreateChecked(1.5707288f)); // pi/2
+        var c1 = Create(T.CreateChecked(-0.2121144f));
+        var c2 = Create(T.CreateChecked(0.0742610f));
+        var c3 = Create(T.CreateChecked(-0.0187293f));
+
+        var term1 = MultipleAdd(x, c3, c2);
+        var term2 = MultipleAdd(x, term1, c1);
+        var poly = MultipleAdd(x, term2, c0);
+
+        var sqrtTerm = Sqrt(One - x);
+        var result = poly * sqrtTerm;
+
+        var pi = Create(T.CreateChecked(Math.PI));
+        var isNegative = LessThan(value, Zero);
+
+        return Select(isNegative, pi - result, result);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Atan(WideLane<T> value)
+    {
+        // atan(value) = value * (c1 + c2*value^2)
+
+        var c1 = Create(T.CreateChecked(0.97239411f));
+        var c2 = Create(T.CreateChecked(-0.19194795f));
+
+        var x2 = value * value;
+        var poly = MultipleAdd(x2, c2, c1);
+        return value * poly;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Atan2(WideLane<T> y, WideLane<T> x)
+    {
+        var absX = Abs(x);
+        var absY = Abs(y);
+
+        // 1. Determine the ratio (input to Atan)
+        // If |value| > |y|, we are in the "shallow" region, ratio = y/value
+        // If |y| > |value|, we are in the "steep" region, ratio = value/y (and we transform result)
+        var yGtX = GreaterThan(absY, absX);
+
+        // Select numerator and denominator to ensure ratio is always in [-1, 1]
+        var num = Select(yGtX, absX, absY);
+        var den = Select(yGtX, absY, absX);
+
+        var t = num / den; // t is now in [0, 1]
+        var t2 = t * t;
+
+        // 2. Polynomial Approximation (Odd function: value * (c1 + c2*value^2))
+        var c1 = Create(T.CreateChecked(0.97239411f));
+        var c2 = Create(T.CreateChecked(-0.19194795f));
+
+        // (c1 + c2 * t2)
+        var poly = MultipleAdd(c2, t2, c1);
+
+        // result = t * poly
+        var result = t * poly;
+
+        // 3. Reconstruct the angle
+        // If we swapped value/y (yGtX), the identity is: atan(value/y) = PI/2 - atan(y/value)
+        var halfPi = Create(T.CreateChecked(1.570796327f));
+        result = Select(yGtX, halfPi - result, result);
+
+        // 4. Adjust for Quadrants (Signs)
+        // If value < 0, we are in quadrants 2 or 3, so we need to add PI
+        var pi = Create(T.CreateChecked(3.141592654f));
+        var xLtZero = LessThan(x, Zero);
+        result = Select(xLtZero, pi - result, result);
+
+        // If y < 0, the result should be negative (standard atan2 convention)
+        // NOTE: This sign flip strategy depends on exact polynomial range mapping, 
+        // but typically just copy the sign of Y to the result.
+        var yLtZero = LessThan(y, Zero);
+        // If original Y was negative, negate the result
+        // (This works because our ratio logic effectively computed atan(|y|/|value|) above)
+        var negativeResult = -result;
+        return Select(yLtZero, negativeResult, result);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Pow(WideLane<T> x, WideLane<T> y) => Exp(y * Log(x));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Exp(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Exp(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Exp(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Exp2(WideLane<T> value)
+    {
+        return Pow(Create(T.CreateChecked(2)), value);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Log(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Log(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Log(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Log2(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Log2(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Log2(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Ceil(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Ceiling(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Ceiling(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Round(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Round(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Round(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Trunc(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<float>>(ref value);
+            var result = Vector.Truncate(v);
+            return new WideLane<T>(Unsafe.As<Vector<float>, Vector<T>>(ref result));
+        }
+        else if (typeof(T) == typeof(double))
+        {
+            ref var v = ref Unsafe.As<WideLane<T>, Vector<double>>(ref value);
+            var result = Vector.Truncate(v);
+            return new WideLane<T>(Unsafe.As<Vector<double>, Vector<T>>(ref result));
+        }
+
+        return value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Sign(WideLane<T> value) => Select(
+            GreaterThan(value, Zero),
+            One,
+            Select(
+                LessThan(value, Zero),
+                ~Zero,
+                Zero));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> CopySign(WideLane<T> magnitude, WideLane<T> sign) => new(Vector.CopySign(magnitude.value, sign.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Rcp(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            if (Sse.IsSupported && LaneWidth == Vector128<float>.Count)
+            {
+                var vf = Unsafe.As<WideLane<T>, Vector128<float>>(ref value);
+                var result = Sse.Reciprocal(vf);
+                return Unsafe.As<Vector128<float>, WideLane<T>>(ref result);
+            }
+            else if (Avx.IsSupported && LaneWidth == Vector256<float>.Count)
+            {
+                var vf = Unsafe.As<WideLane<T>, Vector256<float>>(ref value);
+                var result = Avx.Reciprocal(vf);
+                return Unsafe.As<Vector256<float>, WideLane<T>>(ref result);
+            }
+        }
+
+        return Create(T.One) / value;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Rsqrt(WideLane<T> value)
+    {
+        if (typeof(T) == typeof(float))
+        {
+            if (Sse.IsSupported && LaneWidth == Vector128<float>.Count)
+            {
+                var vf = Unsafe.As<WideLane<T>, Vector128<float>>(ref value);
+                var result = Sse.ReciprocalSqrt(vf);
+                return Unsafe.As<Vector128<float>, WideLane<T>>(ref result);
+            }
+            else if (Avx.IsSupported && LaneWidth == Vector256<float>.Count)
+            {
+                var vf = Unsafe.As<WideLane<T>, Vector256<float>>(ref value);
+                var result = Avx.ReciprocalSqrt(vf);
+                return Unsafe.As<Vector256<float>, WideLane<T>>(ref result);
+            }
+        }
+
+        return Create(T.One) / Sqrt(value);
+    }
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Select(WideLane<T> conditionMask, WideLane<T> ifTrue, WideLane<T> ifFalse)
+        => new(Vector.ConditionalSelect(
+            conditionMask.value,
+            ifTrue.value,
+            ifFalse.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> GreaterThan(WideLane<T> a, WideLane<T> b) => new(Vector.GreaterThan(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> GreaterThanOrEqual(WideLane<T> a, WideLane<T> b) => new(Vector.GreaterThanOrEqual(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> LessThan(WideLane<T> a, WideLane<T> b) => new(Vector.LessThan(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> LessThanOrEqual(WideLane<T> a, WideLane<T> b) => new(Vector.LessThanOrEqual(a.value, b.value));
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static WideLane<T> Equal(WideLane<T> a, WideLane<T> b) => new(Vector.Equals(a.value, b.value));
+
+
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool Any(WideLane<T> mask) => !Vector.EqualsAll(mask.value, Vector<T>.Zero);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool All(WideLane<T> mask) => Vector.EqualsAll(mask.value, Vector<T>.AllBitsSet);
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static bool None(WideLane<T> mask) => Vector.EqualsAll(mask.value, Vector<T>.Zero);
+
+
+    public override string ToString()
+    {
+        return value.ToString();
+    }
+}

Misaki.HighPerformance.Mathematics/AutoSIMDAttribute.cs

@@ -0,0 +1,9 @@
+using System;
+using System.Collections.Generic;
+using System.Text;
+
+namespace Misaki.HighPerformance.Mathematics;
+
+internal class AutoSIMDAttribute
+{
+}

Misaki.HighPerformance.Mathematics/math.cs

@@ -2503,7 +2503,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float tan(float x)
     {
-        return (float)System.Math.Tan(x);
+        return MathF.Tan(x);
     }
 
     /// <summary>Returns the componentwise tangent of a float2 vector.</summary>
@@ -2540,7 +2540,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double tan(double x)
     {
-        return System.Math.Tan(x);
+        return Math.Tan(x);
     }
 
     /// <summary>Returns the componentwise tangent of a double2 vector.</summary>
@@ -2577,7 +2577,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float tanh(float x)
     {
-        return (float)System.Math.Tanh(x);
+        return MathF.Tanh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic tangent of a float2 vector.</summary>
@@ -2614,7 +2614,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double tanh(double x)
     {
-        return System.Math.Tanh(x);
+        return Math.Tanh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic tangent of a double2 vector.</summary>
@@ -2651,7 +2651,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float atan(float x)
     {
-        return (float)System.Math.Atan(x);
+        return MathF.Atan(x);
     }
 
     /// <summary>Returns the componentwise arctangent of a float2 vector.</summary>
@@ -2688,7 +2688,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double atan(double x)
     {
-        return System.Math.Atan(x);
+        return Math.Atan(x);
     }
 
     /// <summary>Returns the componentwise arctangent of a double2 vector.</summary>
@@ -2726,7 +2726,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float atan2(float y, float x)
     {
-        return (float)System.Math.Atan2(y, x);
+        return MathF.Atan2(y, x);
     }
 
     /// <summary>Returns the componentwise 2-argument arctangent of a pair of floats2 vectors.</summary>
@@ -2767,7 +2767,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double atan2(double y, double x)
     {
-        return System.Math.Atan2(y, x);
+        return Math.Atan2(y, x);
     }
 
     /// <summary>Returns the 2-argument arctangent of a pair of double2 vectors.</summary>
@@ -2807,7 +2807,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float cos(float x)
     {
-        return (float)System.Math.Cos(x);
+        return MathF.Cos(x);
     }
 
     /// <summary>Returns the componentwise cosine of a float2 vector.</summary>
@@ -2844,7 +2844,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double cos(double x)
     {
-        return System.Math.Cos(x);
+        return Math.Cos(x);
     }
 
     /// <summary>Returns the componentwise cosine of a double2 vector.</summary>
@@ -2881,7 +2881,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float cosh(float x)
     {
-        return (float)System.Math.Cosh(x);
+        return MathF.Cosh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic cosine of a float2 vector.</summary>
@@ -2918,7 +2918,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double cosh(double x)
     {
-        return System.Math.Cosh(x);
+        return Math.Cosh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic cosine of a double2 vector.</summary>
@@ -2955,7 +2955,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float acos(float x)
     {
-        return (float)System.Math.Acos((float)x);
+        return MathF.Acos(x);
     }
 
     /// <summary>Returns the componentwise arccosine of a float2 vector.</summary>
@@ -2992,7 +2992,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double acos(double x)
     {
-        return System.Math.Acos(x);
+        return Math.Acos(x);
     }
 
     /// <summary>Returns the componentwise arccosine of a double2 vector.</summary>
@@ -3029,7 +3029,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float sin(float x)
     {
-        return (float)System.Math.Sin((float)x);
+        return MathF.Sin(x);
     }
 
     /// <summary>Returns the componentwise sine of a float2 vector.</summary>
@@ -3066,7 +3066,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double sin(double x)
     {
-        return System.Math.Sin(x);
+        return Math.Sin(x);
     }
 
     /// <summary>Returns the componentwise sine of a double2 vector.</summary>
@@ -3103,7 +3103,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float sinh(float x)
     {
-        return (float)System.Math.Sinh((float)x);
+        return MathF.Sinh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic sine of a float2 vector.</summary>
@@ -3140,7 +3140,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double sinh(double x)
     {
-        return System.Math.Sinh(x);
+        return Math.Sinh(x);
     }
 
     /// <summary>Returns the componentwise hyperbolic sine of a double2 vector.</summary>
@@ -3177,7 +3177,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float asin(float x)
     {
-        return (float)System.Math.Asin((float)x);
+        return MathF.Asin(x);
     }
 
     /// <summary>Returns the componentwise arcsine of a float2 vector.</summary>
@@ -3214,7 +3214,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double asin(double x)
     {
-        return System.Math.Asin(x);
+        return Math.Asin(x);
     }
 
     /// <summary>Returns the componentwise arcsine of a double2 vector.</summary>
@@ -3251,7 +3251,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float floor(float x)
     {
-        return (float)System.Math.Floor((float)x);
+        return MathF.Floor(x);
     }
 
     /// <summary>Returns the result of rounding each component of a float2 vector value down to the nearest value less or equal to the original value.</summary>
@@ -3288,7 +3288,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double floor(double x)
     {
-        return System.Math.Floor(x);
+        return Math.Floor(x);
     }
 
     /// <summary>Returns the result of rounding each component of a double2 vector value down to the nearest value less or equal to the original value.</summary>
@@ -3325,7 +3325,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float ceil(float x)
     {
-        return (float)System.Math.Ceiling((float)x);
+        return MathF.Ceiling(x);
     }
 
     /// <summary>Returns the result of rounding each component of a float2 vector value up to the nearest value greater or equal to the original value.</summary>
@@ -3362,7 +3362,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double ceil(double x)
     {
-        return System.Math.Ceiling(x);
+        return Math.Ceiling(x);
     }
 
     /// <summary>Returns the result of rounding each component of a double2 vector value up to the nearest integral value greater or equal to the original value.</summary>
@@ -3399,7 +3399,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float round(float x)
     {
-        return (float)System.Math.Round((float)x);
+        return MathF.Round(x);
     }
 
     /// <summary>Returns the result of rounding each component of a float2 vector value to the nearest integral value.</summary>
@@ -3436,7 +3436,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double round(double x)
     {
-        return System.Math.Round(x);
+        return Math.Round(x);
     }
 
     /// <summary>Returns the result of rounding each component of a double2 vector value to the nearest integral value.</summary>
@@ -3473,7 +3473,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float trunc(float x)
     {
-        return (float)System.Math.Truncate((float)x);
+        return MathF.Truncate(x);
     }
 
     /// <summary>Returns the result of a componentwise truncation of a float2 value to an integral float2 value.</summary>
@@ -3500,7 +3500,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float4 trunc(float4 x)
     {
-        return Vector128.Truncate(x.AsVector128()).Asfloat4();
+        return Vector128.Truncate(x.__v).Asfloat4();
     }
 
 
@@ -3510,7 +3510,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double trunc(double x)
     {
-        return System.Math.Truncate(x);
+        return Math.Truncate(x);
     }
 
     /// <summary>Returns the result of a componentwise truncation of a double2 value to an integral double2 value.</summary>
@@ -3537,7 +3537,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double4 trunc(double4 x)
     {
-        return Vector256.Truncate(x.AsVector256()).Asdouble4();
+        return Vector256.Truncate(x.__v).Asdouble4();
     }
 
 
@@ -3805,7 +3805,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float pow(float x, float y)
     {
-        return (float)System.Math.Pow((float)x, (float)y);
+        return MathF.Pow(x, (float)y);
     }
 
     /// <summary>Returns the componentwise result of raising x to the power y.</summary>
@@ -3846,7 +3846,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double pow(double x, double y)
     {
-        return System.Math.Pow(x, y);
+        return Math.Pow(x, y);
     }
 
     /// <summary>Returns the componentwise result of raising x to the power y.</summary>
@@ -3886,7 +3886,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float exp(float x)
     {
-        return (float)System.Math.Exp((float)x);
+        return MathF.Exp(x);
     }
 
     /// <summary>Returns the componentwise base-e exponential of x.</summary>
@@ -3923,7 +3923,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double exp(double x)
     {
-        return System.Math.Exp(x);
+        return Math.Exp(x);
     }
 
     /// <summary>Returns the componentwise base-e exponential of x.</summary>
@@ -3960,7 +3960,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float exp2(float x)
     {
-        return (float)System.Math.Exp((float)x * 0.69314718f);
+        return MathF.Exp(x * 0.69314718f);
     }
 
     /// <summary>Returns the componentwise base-2 exponential of x.</summary>
@@ -3997,7 +3997,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double exp2(double x)
     {
-        return System.Math.Exp(x * 0.693147180559945309);
+        return Math.Exp(x * 0.693147180559945309);
     }
 
     /// <summary>Returns the componentwise base-2 exponential of x.</summary>
@@ -4034,7 +4034,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float exp10(float x)
     {
-        return (float)System.Math.Exp((float)x * 2.30258509f);
+        return MathF.Exp(x * 2.30258509f);
     }
 
     /// <summary>Returns the componentwise base-10 exponential of x.</summary>
@@ -4071,7 +4071,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double exp10(double x)
     {
-        return System.Math.Exp(x * 2.302585092994045684);
+        return Math.Exp(x * 2.302585092994045684);
     }
 
     /// <summary>Returns the componentwise base-10 exponential of x.</summary>
@@ -4108,7 +4108,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float log(float x)
     {
-        return (float)System.Math.Log((float)x);
+        return MathF.Log(x);
     }
 
     /// <summary>Returns the componentwise natural logarithm of a float2 vector.</summary>
@@ -4145,7 +4145,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double log(double x)
     {
-        return System.Math.Log(x);
+        return Math.Log(x);
     }
 
     /// <summary>Returns the componentwise natural logarithm of a double2 vector.</summary>
@@ -4182,7 +4182,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float log2(float x)
     {
-        return (float)System.Math.Log((float)x, 2.0f);
+        return MathF.Log(x, 2.0f);
     }
 
     /// <summary>Returns the componentwise base-2 logarithm of a float2 vector.</summary>
@@ -4219,7 +4219,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double log2(double x)
     {
-        return System.Math.Log(x, 2.0);
+        return Math.Log(x, 2.0);
     }
 
     /// <summary>Returns the componentwise base-2 logarithm of a double2 vector.</summary>
@@ -4255,7 +4255,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float log10(float x)
     {
-        return (float)System.Math.Log10((float)x);
+        return MathF.Log10(x);
     }
 
     /// <summary>Returns the componentwise base-10 logarithm of a float2 vector.</summary>
@@ -4292,7 +4292,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double log10(double x)
     {
-        return System.Math.Log10(x);
+        return Math.Log10(x);
     }
 
     /// <summary>Returns the componentwise base-10 logarithm of a double2 vector.</summary>
@@ -4519,7 +4519,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float sqrt(float x)
     {
-        return (float)System.Math.Sqrt((float)x);
+        return MathF.Sqrt(x);
     }
 
     /// <summary>Returns the componentwise square root of a float2 vector.</summary>
@@ -4556,7 +4556,7 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double sqrt(double x)
     {
-        return System.Math.Sqrt(x);
+        return Math.Sqrt(x);
     }
 
     /// <summary>Returns the componentwise square root of a double2 vector.</summary>
@@ -4727,8 +4727,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float2 normalizesafe(float2 x, float2 defaultvalue = new float2())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
     /// <summary>
@@ -4741,8 +4741,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float3 normalizesafe(float3 x, float3 defaultvalue = new float3())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
     /// <summary>
@@ -4755,8 +4755,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float4 normalizesafe(float4 x, float4 defaultvalue = new float4())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
 
@@ -4770,8 +4770,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double2 normalizesafe(double2 x, double2 defaultvalue = new double2())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
     /// <summary>
@@ -4784,8 +4784,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double3 normalizesafe(double3 x, double3 defaultvalue = new double3())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
     /// <summary>
@@ -4798,8 +4798,8 @@ public static partial class math
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static double4 normalizesafe(double4 x, double4 defaultvalue = new double4())
     {
-        var len = math.dot(x, x);
+        var len = dot(x, x);
-        return math.select(defaultvalue, x * math.rsqrt(len), len > FLT_MIN_NORMAL);
+        return select(defaultvalue, x * rsqrt(len), len > FLT_MIN_NORMAL);
     }
 
 
@@ -8665,7 +8665,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.z - d1.y;
         if (y1 * y1 < cutoff)
@@ -8705,7 +8705,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.x + d1.z;
         if (y1 * y1 < cutoff)
@@ -8745,7 +8745,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.y + d1.x;
         if (y1 * y1 < cutoff)
@@ -8785,7 +8785,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.x - d1.z;
         if (y1 * y1 < cutoff)
@@ -8825,7 +8825,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.y - d1.x;
         if (y1 * y1 < cutoff)
@@ -8864,7 +8864,7 @@ public static partial class math
         var d1 = qv * qv.wwww * float4(2f); //xw, yw, zw, ww
         var d2 = qv * qv.yzxw * float4(2f); //xy, yz, zx, ww
         var d3 = qv * qv;
-        var euler = Misaki.HighPerformance.Mathematics.float3.zero;
+        var euler = Mathematics.float3.zero;
 
         var y1 = d2.z + d1.y;
         if (y1 * y1 < cutoff)
@@ -8898,24 +8898,24 @@ public static partial class math
     /// <param name="order">The order in which the rotations are applied.</param>
     /// <returns>The Euler angle representation of the quaternion in the specified order.</returns>
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public static float3 Euler(quaternion q, math.RotationOrder order = math.RotationOrder.Default)
+    public static float3 Euler(quaternion q, math.RotationOrder order = RotationOrder.Default)
     {
         switch (order)
         {
-            case math.RotationOrder.XYZ:
+            case RotationOrder.XYZ:
                 return EulerXYZ(q);
-            case math.RotationOrder.XZY:
+            case RotationOrder.XZY:
                 return EulerXZY(q);
-            case math.RotationOrder.YXZ:
+            case RotationOrder.YXZ:
                 return EulerYXZ(q);
-            case math.RotationOrder.YZX:
+            case RotationOrder.YZX:
                 return EulerYZX(q);
-            case math.RotationOrder.ZXY:
+            case RotationOrder.ZXY:
                 return EulerZXY(q);
-            case math.RotationOrder.ZYX:
+            case RotationOrder.ZYX:
                 return EulerZYX(q);
             default:
-                return Misaki.HighPerformance.Mathematics.float3.zero;
+                return Mathematics.float3.zero;
         }
     }
 

Misaki.HighPerformance.Test/Benchmark/MathematicsBenchmark.cs

@@ -1,7 +1,8 @@
-#define NOISE_BENCHMARK
+#define ADD_BENCHMARK
 
 using BenchmarkDotNet.Attributes;
 using Misaki.HighPerformance.Mathematics;
+using Misaki.HighPerformance.Mathematics.SPMD;
 using System.Numerics;
 using System.Runtime.Intrinsics;
 
@@ -9,7 +10,7 @@ namespace Misaki.HighPerformance.Test.Benchmark;
 
 public class MathematicsBenchmark
 {
-#if VECTOR_BENCHMARK
+#if ADD_BENCHMARK
     private Vector4 _va = new Vector4(1, 2, 1, 2);
     private Vector4 _vb = new Vector4(3, 4, 3, 4);
 
@@ -39,61 +40,52 @@ public class MathematicsBenchmark
     }
 #endif
 
-#if NOISE_BENCHMARK
+#if FMA_BENCHMARK
+    private Vector4 _va = new Vector4(1, 2, 1, 2);
+    private Vector4 _vb = new Vector4(3, 4, 3, 4);
+    private Vector4 _vc = new Vector4(5, 6, 5, 6);
 
-    private const int _SIZE = 32;
+    private Vector128<float> _va128 = Vector128.Create(1f, 2f, 1f, 2f);
+    private Vector128<float> _vb128 = Vector128.Create(3f, 4f, 3f, 4f);
+    private Vector128<float> _vc128 = Vector128.Create(5f, 6f, 5f, 6f);
+
+    private float4 _fa = new float4(1, 2, 1, 2);
+    private float4 _fb = new float4(3, 4, 3, 4);
+    private float4 _fc = new float4(5, 6, 5, 6);
 
     [Benchmark]
-    public unsafe void VectorNoise()
+    public Vector4 Vector4()
     {
-        var buf = stackalloc float[_SIZE * _SIZE];
+        for (var i = 0; i < 10; i++)
-        var job = new Misaki.HighPerformance.Test.Jobs.NoiseJobVector
         {
-            buffers = buf,
+            _va = _vb * _vc + _va;
-            width = _SIZE,
-            height = _SIZE,
-        };
-
-        for (var i = 0; i < _SIZE * _SIZE; i++)
-        {
-            job.Execute(i, 0);
         }
+
+        return _va;
     }
 
     [Benchmark]
-    public unsafe void MathNoise()
+    public Vector128<float> VectorFMA()
     {
-        var buf = stackalloc float[_SIZE * _SIZE];
+        for (var i = 0; i < 10; i++)
-        var job = new Misaki.HighPerformance.Test.Jobs.NoiseJobMath
         {
-            buffers = buf,
+            _va128 = System.Runtime.Intrinsics.X86.Fma.MultiplyAdd(_vb128, _vc128, _va128);
-            width = _SIZE,
-            height = _SIZE,
-        };
-
-        for (var i = 0; i < _SIZE * _SIZE; i++)
-        {
-            job.Execute(i, 0);
         }
+
+        return _va128;
     }
 
     [Benchmark]
-    // This is 10x faster than VectorNoise and MathNoise, but writing a burst like compiler to compile MathNoise into this is incredibly hard.
+    public float4 floatFMA()
-    public unsafe void MathVNoise()
     {
-        var buf = stackalloc float[_SIZE * _SIZE];
+        for (var i = 0; i < 10; i++)
-        var job = new Misaki.HighPerformance.Test.Jobs.NoiseJobMathV
         {
-            buffers = buf,
+            _fa = _fb * _fc + _fa;
-            width = _SIZE,
-            height = _SIZE,
-        };
-
-        for (var i = 0; i < _SIZE * _SIZE / 8; i++)
-        {
-            job.Execute(i, 0);
         }
+
+        return _fa;
     }
+
 #endif
 
 #if MATRIX_BENCHMARK

Misaki.HighPerformance.Test/Benchmark/SPMDBenchmark.cs

@@ -0,0 +1,98 @@
+using BenchmarkDotNet.Attributes;
+using Misaki.HighPerformance.Jobs;
+using Misaki.HighPerformance.Mathematics.SPMD;
+using System.Runtime.InteropServices;
+
+namespace Misaki.HighPerformance.Test.Benchmark;
+
+public unsafe class SPMDBenchmark
+{
+    private const int _SIZE = 512;
+
+    private JobScheduler _scheduler = null!;
+    private float* _buf;
+
+    [GlobalSetup]
+    public void Setup()
+    {
+        _scheduler = new JobScheduler(Environment.ProcessorCount);
+        _buf = (float*)NativeMemory.Alloc(sizeof(float) * _SIZE * _SIZE);
+    }
+
+    [GlobalCleanup]
+    public void Cleanup()
+    {
+        _scheduler.Dispose();
+        NativeMemory.Free(_buf);
+    }
+
+    [Benchmark]
+    public void VectorNoiseSingleThread()
+    {
+        var job = new Jobs.NoiseJobVector
+        {
+            buffers = _buf,
+            width = _SIZE,
+            height = _SIZE,
+        };
+
+        job.Run(_SIZE * _SIZE, 0);
+    }
+
+    //[Benchmark]
+    public void VectorNoise()
+    {
+        var job = new Jobs.NoiseJobVector
+        {
+            buffers = _buf,
+            width = _SIZE,
+            height = _SIZE,
+        };
+
+        var handle = _scheduler.ScheduleParallel(ref job, _SIZE * _SIZE, 64);
+        _scheduler.WaitComplete(handle);
+    }
+
+    //[Benchmark]
+    public void MathNoise()
+    {
+        var job = new Jobs.NoiseJobMath
+        {
+            buffers = _buf,
+            width = _SIZE,
+            height = _SIZE,
+        };
+
+        var handle = _scheduler.ScheduleParallel(ref job, _SIZE * _SIZE, 64);
+        _scheduler.WaitComplete(handle);
+    }
+
+    //[Benchmark(Baseline = true)]
+    public void ManualSPMDNoise()
+    {
+        var job = new Jobs.NoiseJobMathV
+        {
+            buffers = _buf,
+            width = _SIZE,
+            height = _SIZE,
+        };
+
+        var iterations = (_SIZE * _SIZE + 8 - 1) / 8;
+        var handle = _scheduler.ScheduleParallel(ref job, iterations, 64);
+        _scheduler.WaitComplete(handle);
+    }
+
+    [Benchmark(Baseline = true)]
+    public void SPMDNoise()
+    {
+        var job = new Jobs.NoiseJobMathSPMD
+        {
+            buffers = _buf,
+            width = _SIZE,
+            height = _SIZE,
+        };
+
+        var handle = _scheduler.ScheduleParallelSPDM<Jobs.NoiseJobMathSPMD, float>(ref job, _SIZE * _SIZE, 64, -1, JobHandle.Invalid);
+        _scheduler.WaitComplete(handle);
+    }
+}

Misaki.HighPerformance.Test/Jobs/NoiseJobVector.cs

@@ -1,9 +1,9 @@
 using Misaki.HighPerformance.Jobs;
 using Misaki.HighPerformance.Mathematics;
+using Misaki.HighPerformance.Mathematics.SPMD;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.Intrinsics;
-using System.Runtime.Intrinsics.X86;
 
 namespace Misaki.HighPerformance.Test.Jobs;
 
@@ -16,7 +16,7 @@ internal unsafe struct NoiseJobVector : IJobParallelFor
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static float Frac(float x)
     {
-        return x - MathF.Truncate(x);
+        return x - MathF.Floor(x);
     }
 
     private static Vector2 GradientNoiseDirect(Vector2 uv)
@@ -101,7 +101,7 @@ internal unsafe struct NoiseJobMathV : IJobParallelFor
     private static Vector256<float> Mod289(Vector256<float> x)
     {
         var div = x / Vector256.Create(289.0f);
-        var flr = Vector256.Floor(div);
+        var flr = Vector256.Truncate(div);
         return x - (flr * Vector256.Create(289.0f));
     }
 
@@ -119,12 +119,13 @@ internal unsafe struct NoiseJobMathV : IJobParallelFor
         var hy = Mod289(iy);
 
         var p = hx * Vector256.Create(34.0f) + Vector256.Create(1.0f);
-        p = Mod289(p * hx + hy);
+        p = Mod289(p * hx) + hy;
+        var pPrev = p;
         p = p * Vector256.Create(34.0f) + Vector256.Create(1.0f);
-        p = Mod289(p * hx);
+        p = Mod289(p * pPrev);
 
         var r = (p / 41.0f);
-        r = (r - Vector256.Floor(r)) * 2.0f - Vector256<float>.One;
+        r = (r - Vector256.Truncate(r)) * 2.0f - Vector256<float>.One;
 
         var gx = r - Vector256.Floor(r + Vector256.Create(0.5f));
         var gy = Vector256.Abs(r) - Vector256.Create(0.5f);
@@ -153,10 +154,10 @@ internal unsafe struct NoiseJobMathV : IJobParallelFor
         var d10 = GradDot(ipX + Vector256<float>.One, ipY, fpX - Vector256<float>.One, fpY);
         var d11 = GradDot(ipX + Vector256<float>.One, ipY + Vector256<float>.One, fpX - Vector256<float>.One, fpY - Vector256<float>.One);
 
-        var lerpX1 = d00 + (d10 - d00) * uX;
+        var lerpY1 = d00 + (d10 - d00) * uY;
-        var lerpX2 = d01 + (d11 - d01) * uX;
+        var lerpY2 = d01 + (d11 - d01) * uY;
 
-        return lerpX1 + (lerpX2 - lerpX1) * uY;
+        return lerpY1 + (lerpY2 - lerpY1) * uX;
     }
 
     public void Execute(int loopIndex, int threadIndex)
@@ -164,15 +165,17 @@ internal unsafe struct NoiseJobMathV : IJobParallelFor
         // ---------------------------------------------------------
         // IMPORTANT: Loop Stride is now 8!
         // ---------------------------------------------------------
-        int baseIndex = loopIndex * 8;
+        var baseIndex = loopIndex * 8;
 
         // Safety check
         if (baseIndex + 7 >= width * height)
+        {
             return;
+        }
 
         // Calculate Coords
-        int y = baseIndex / width;
+        var y = baseIndex / width;
-        int x = baseIndex % width;
+        var x = baseIndex % width;
 
         // Sequence: 0, 1, 2, 3, 4, 5, 6, 7
         var vSeqX = Vector256.Create(0f, 1f, 2f, 3f, 4f, 5f, 6f, 7f);
@@ -185,6 +188,81 @@ internal unsafe struct NoiseJobMathV : IJobParallelFor
         var result = GradientNoiseAVX(vBaseX / vWidth, vBaseY / vHeight);
 
         // Store 8 floats (32 bytes)
-        Avx.Store(buffers + baseIndex, result);
+        result.Store(buffers + baseIndex);
+    }
+}
+
+internal unsafe struct NoiseJobMathSPMD : IJobSPMD<float>
+{
+    public float* buffers;
+    public int width;
+    public int height;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static T GradDot<T>(T ix, T iy, T fx, T fy)
+        where T : ISPMD<T, float>
+    {
+        var c289 = T.Create(289f);
+        var c34 = T.Create(34f);
+        var c1 = T.Create(1f);
+        var c41 = T.Create(41f);
+        var c2 = T.Create(2f);
+        var half = T.Create(0.5f);
+
+        ix %= c289;
+        iy %= c289;
+        var x = (c34 * ix + c1) * ix % c289 + iy;
+        x = (c34 * x + c1) * x % c289;
+        x = T.Frac(x / c41) * c2 - c1;
+
+        var gx = x - T.Floor(x + half);
+        var gy = T.Abs(x) - half;
+
+        // normalize
+        var len = T.Sqrt(gx * gx + gy * gy);
+        gx /= len;
+        gy /= len;
+
+        return gx * fx + gy * fy;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static T Noise<T>(T uvX, T uvY)
+        where T : ISPMD<T, float>
+    {
+        var c1 = T.Create(1f);
+        var c6 = T.Create(6f);
+        var c10 = T.Create(10f);
+        var c15 = T.Create(15f);
+
+        var ipX = T.Floor(uvX);
+        var ipY = T.Floor(uvY);
+        var fpX = uvX - ipX;
+        var fpY = uvY - ipY;
+
+        var d00 = GradDot(ipX, ipY, fpX, fpY);
+        var d01 = GradDot(ipX, ipY + c1, fpX, fpY - c1);
+        var d10 = GradDot(ipX + c1, ipY, fpX - c1, fpY);
+        var d11 = GradDot(ipX + c1, ipY + c1, fpX - c1, fpY - c1);
+
+        // fade
+        var uX = fpX * fpX * fpX * (fpX * (fpX * c6 - c15) + c10);
+        var uY = fpY * fpY * fpY * (fpY * (fpY * c6 - c15) + c10);
+
+        return T.Lerp(T.Lerp(d00, d10, uY), T.Lerp(d01, d11, uY), uX);
+    }
+
+    public readonly void Execute<TLane>(int baseIndex, int threadIndex)
+        where TLane : ISPMD<TLane, float>
+    {
+        var indices = TLane.Sequence(baseIndex, 1f);
+        var w = TLane.Create(width);
+        var h = TLane.Create(height);
+
+        var uvX = (indices % w) / w;
+        var uvY = TLane.Floor(indices / w) / h;
+        
+        var result = Noise(uvX, uvY);
+        result.Store(buffers + baseIndex);
     }
 }

Misaki.HighPerformance.Test/Misaki.HighPerformance.Test.csproj

@@ -24,6 +24,7 @@
     <ProjectReference Include="..\Misaki.HighPerformance.Image\Misaki.HighPerformance.Image.csproj" />
     <ProjectReference Include="..\Misaki.HighPerformance.Jobs\Misaki.HighPerformance.Jobs.csproj" />
     <ProjectReference Include="..\Misaki.HighPerformance.LowLevel\Misaki.HighPerformance.LowLevel.csproj" />
+    <ProjectReference Include="..\Misaki.HighPerformance.Mathematics.SPMD\Misaki.HighPerformance.Mathematics.SPMD.csproj" />
     <ProjectReference Include="..\Misaki.HighPerformance.Mathematics\Misaki.HighPerformance.Mathematics.csproj" />
     <ProjectReference Include="..\Misaki.HighPerformance\Misaki.HighPerformance.csproj" />
     <ProjectReference Include="..\Misaki.HighPerformance.Analyzer\Misaki.HighPerformance.Analyzer\Misaki.HighPerformance.Analyzer.csproj" OutputItemType="Analyzer" ReferenceOutputAssembly="false" />

Misaki.HighPerformance.Test/Program.cs

@@ -1,9 +1,15 @@
 using Misaki.HighPerformance;
+using Misaki.HighPerformance.Jobs;
 using Misaki.HighPerformance.LowLevel;
 using Misaki.HighPerformance.LowLevel.Utilities;
+using Misaki.HighPerformance.Mathematics.SPMD;
+using Misaki.HighPerformance.Test.UnitTest.Jobs;
+using System.Numerics;
+using System.Runtime.Intrinsics;
+using System.Text;
 
-BenchmarkDotNet.Running.BenchmarkRunner.Run<Misaki.HighPerformance.Test.Benchmark.MathematicsBenchmark>();
+BenchmarkDotNet.Running.BenchmarkRunner.Run<Misaki.HighPerformance.Test.Benchmark.SPMDBenchmark>();
-
+//return;
 //using Misaki.HighPerformance.Collections;
 //using Misaki.HighPerformance.LowLevel.Buffer;
 //using Misaki.HighPerformance.LowLevel.Collections;

Misaki.HighPerformance.Test/UnitTest/Jobs/CompressStoreTest.cs

@@ -0,0 +1,114 @@
+using Misaki.HighPerformance.Mathematics.SPMD;
+using System.Numerics;
+
+namespace Misaki.HighPerformance.Test.UnitTest.Jobs;
+
+public static class CompressStoreTest
+{
+    public static void Run()
+    {
+        Console.WriteLine("--- Testing CompressStore (Double) ---");
+
+        // Test 1: Simple Pattern (True, False, True, False...)
+        TestPattern_Double(
+            input: new double[] { 1, 2, 3, 4, 5, 6, 7, 8 },
+            // Mask: Keep only even numbers (values > 0)
+            // We simulate a mask by comparing against 0 or -1
+            keepPattern: new bool[] { true, false, true, false, true, false, true, false }
+        );
+
+        // Test 2: All True
+        TestPattern_Double(
+            input: new double[] { 10, 20, 30, 40, 50, 60, 70, 80 },
+            keepPattern: new bool[] { true, true, true, true, true, true, true, true }
+        );
+
+        // Test 3: All False
+        TestPattern_Double(
+            input: new double[] { 10, 20, 30, 40, 50, 60, 70, 80 },
+            keepPattern: new bool[] { false, false, false, false, false, false, false, false }
+        );
+
+        // Test 4: Sparse (First and Last only)
+        TestPattern_Double(
+            input: new double[] { 1, 2, 3, 4, 5, 6, 7, 8 },
+            keepPattern: new bool[] { true, false, false, false, false, false, false, true }
+        );
+    }
+
+    private unsafe static void TestPattern_Double(double[] input, bool[] keepPattern)
+    {
+        // 1. Setup Input Vector
+        // Handle case where Vector<T> is smaller than 8 (e.g. 2 or 4)
+        var vecSize = Vector<double>.Count;
+        var safeInput = new double[vecSize];
+        var safeMaskVal = new double[vecSize];
+
+        // Expected Output Calculation
+        var expected = new double[vecSize];
+        var expectedCount = 0;
+
+        for (var i = 0; i < vecSize; i++)
+        {
+            safeInput[i] = input[i];
+            // If we want to keep it, make mask "GreaterThan" true
+            // We'll compare X > 0. 
+            // If keep=true, val=1. If keep=false, val=-1.
+            safeMaskVal[i] = keepPattern[i] ? 1 : -1;
+
+            if (keepPattern[i])
+            {
+                expected[expectedCount++] = input[i];
+            }
+        }
+
+        // 2. Create WideLanes
+        var vInput = WideLane<double>.Load(ref safeInput.AsSpan().GetPinnableReference());
+
+        // Create Mask: greater than 0
+        var vMaskVal = WideLane<double>.Load(ref safeMaskVal.AsSpan().GetPinnableReference());
+        var vZero = WideLane<double>.Create(0);
+        var vMask = WideLane<double>.GreaterThan(vMaskVal, vZero);
+
+        // 3. Run CompressStore
+        var outputBuffer = new double[vecSize];
+        var actualCount = 0;
+
+        fixed (double* ptr = outputBuffer)
+        {
+            actualCount = vInput.CompressStore(vMask, ptr);
+        }
+
+        // 4. Verify
+        var pass = actualCount == expectedCount;
+        for (var i = 0; i < expectedCount; i++)
+        {
+            if (outputBuffer[i] != expected[i])
+                pass = false;
+        }
+
+        // 5. Report
+        var hardware = (vecSize == 4) ? "AVX2 (256-bit)" : (vecSize == 2) ? "SSE/NEON (128-bit)" : "Scalar";
+        Console.Write($"[{hardware}] Pattern: ");
+        for (var i = 0; i < vecSize; i++)
+            Console.Write(keepPattern[i] ? "1" : "0");
+
+        if (pass)
+        {
+            Console.WriteLine($" -> PASS (Count: {actualCount})");
+        }
+        else
+        {
+            Console.WriteLine($" -> FAIL!");
+            Console.WriteLine($"   Expected Count: {expectedCount}, Actual: {actualCount}");
+            Console.Write("   Expected Data: ");
+            foreach (var d in expected)
+                Console.Write($"{d} ");
+            Console.WriteLine();
+            Console.Write("   Actual Data:   ");
+            foreach (var d in outputBuffer)
+                Console.Write($"{d} ");
+            Console.WriteLine();
+        }
+    }
+}

Misaki.HighPerformance.Test/UnitTest/Jobs/TestJobSystem.cs

@@ -1,19 +1,28 @@
-using Misaki.HighPerformance.Jobs;
+using Misaki.HighPerformance.Jobs;
 using Misaki.HighPerformance.LowLevel.Buffer;
 using Misaki.HighPerformance.LowLevel.Collections;
 using Misaki.HighPerformance.LowLevel.Utilities;
+using Misaki.HighPerformance.Mathematics.SPMD;
+using System.Runtime.InteropServices;
 
 namespace Misaki.HighPerformance.Test.UnitTest.Jobs;
 
 [TestClass]
+[DoNotParallelize]
 public unsafe class TestJobSystem
 {
     private JobScheduler _jobScheduler = null!;
 
+    public TestContext TestContext
+    {
+        get;
+        set;
+    }
+
     [TestInitialize]
     public void Initialize()
     {
-        _jobScheduler = new JobScheduler(Environment.ProcessorCount);
+        _jobScheduler = new JobScheduler(3);
     }
 
     [TestCleanup]
@@ -251,4 +260,102 @@ public unsafe class TestJobSystem
 
         Assert.AreEqual(JobState.Completed, _jobScheduler.GetJobStatus(completedHandle));
     }
+
+    [TestMethod]
+    public void RaceConditionTest()
+    {
+        const int jobCount = 20000;
+        
+        var pExecutedCount = (int*)NativeMemory.Alloc(sizeof(int));
+        *pExecutedCount = 0;
+
+        var startSignal = false;
+
+        // 1. Create a "Gatekeeper" vectorJob that spins/blocks a worker thread until signaled.
+        //    This allows us to control exactly when the dependency completes.
+        var rootJob = new WaitJob { pSignal = &startSignal };
+        var rootHandle = _jobScheduler.Schedule(ref rootJob);
+
+        // 2. Start a background task to flood the scheduler with dependencies on the Gatekeeper.
+        using var barrier = new Barrier(2);
+        var scheduleTask = Task.Run(() =>
+        {
+            var depJob = new IncrementJob { pCounter = pExecutedCount };
+            barrier.SignalAndWait(TestContext.CancellationTokenSource.Token); // Synchronize start with main thread
+
+            for (var i = 0; i < jobCount; i++)
+            {
+                // CONTENTION POINT: 
+                // Trying to add a dependency to 'rootHandle'. 
+                // Eventually, this will happen exactly while 'rootHandle' is transitioning to Completed.
+                _jobScheduler.Schedule(ref depJob, rootHandle);
+            }
+        }, TestContext.CancellationTokenSource.Token);
+
+        barrier.SignalAndWait(TestContext.CancellationTokenSource.Token); // Wait for scheduler task to be ready
+
+        // Allow the scheduling loop to get a head start and queue some readers
+        Thread.Sleep(5);
+
+        // 3. Open the gate.
+        // This triggers the Gatekeeper to complete. It will change its State and iterate its dependency list.
+        // This happens CONCURRENTLY with the loop above adding more items to that same list.
+        startSignal = true;
+
+        scheduleTask.Wait(TestContext.CancellationTokenSource.Token);
+
+        // 4. Validate results
+        // If the lock-free logic works, every single dependent vectorJob must eventually execute.
+        // If there is a race (e.g., missed notification), pExecutedCount will stick below jobCount.
+        var spin = new SpinWait();
+        var timeout = DateTime.Now.AddSeconds(10);
+
+        while (Volatile.Read(ref *pExecutedCount) < jobCount)
+        {
+            if (DateTime.Now > timeout)
+            {
+                break;
+            }
+
+            spin.SpinOnce();
+        }
+
+        // Ensure the root vectorJob is officially cleaned up
+        _jobScheduler.WaitComplete(rootHandle);
+
+        Assert.AreEqual(jobCount, *pExecutedCount, "Race condition detected: Some dependent jobs failed to execute (Wait timeout).");
+
+        NativeMemory.Free(pExecutedCount);
+    }
+
+    [TestMethod]
+    public void SPMDCorrectness()
+    {
+        const int size = 8;
+
+        var vectorBuf = stackalloc float[size * size];
+        var vs = new Span<float>(vectorBuf, size * size);
+        var vectorJob = new Misaki.HighPerformance.Test.Jobs.NoiseJobVector
+        {
+            buffers = vectorBuf,
+            width = size,
+            height = size,
+        };
+
+        vectorJob.Run(size * size, -1);
+
+        var spmdBuf = stackalloc float[size * size];
+        var ss = new Span<float>(spmdBuf, size * size);
+        var spmdJob = new Misaki.HighPerformance.Test.Jobs.NoiseJobMath
+        {
+            buffers = spmdBuf,
+            width = size,
+            height = size,
+        };
+
+        spmdJob.Run(size * size, -1);
+
+        var eq = vs.SequenceCompareTo(ss);
+        Assert.AreEqual(0, eq);
+    }
 }

Misaki.HighPerformance.Test/UnitTest/Jobs/TestJobs.cs

@@ -1,4 +1,4 @@
-using Misaki.HighPerformance.Jobs;
+using Misaki.HighPerformance.Jobs;
 
 namespace Misaki.HighPerformance.Test.UnitTest.Jobs;
 
@@ -71,3 +71,27 @@ internal unsafe struct ParallelMultiplyJob : IJobParallelFor
         inout[loopIndex] *= multiplier;
     }
 }
+
+public unsafe struct WaitJob : IJob
+{
+    public bool* pSignal;
+
+    public void Execute(int loopIndex)
+    {
+        var spin = new SpinWait();
+        while (!Volatile.Read(ref *pSignal))
+        {
+            spin.SpinOnce();
+        }
+    }
+}
+
+public unsafe struct IncrementJob : IJob
+{
+    public int* pCounter;
+
+    public void Execute(int loopIndex)
+    {
+        Interlocked.Increment(ref *pCounter);
+    }
+}

Misaki.HighPerformance.sln

@@ -1,7 +1,7 @@
 
 Microsoft Visual Studio Solution File, Format Version 12.00
 # Visual Studio Version 18
-VisualStudioVersion = 18.3.11218.70 d18.3
+VisualStudioVersion = 18.3.11218.70
 MinimumVisualStudioVersion = 10.0.40219.1
 Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "Misaki.HighPerformance", "Misaki.HighPerformance\Misaki.HighPerformance.csproj", "{275B2E80-9B2A-4567-A157-F147A6B28A0F}"
 EndProject
@@ -25,6 +25,8 @@ Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "Misaki.HighPerformance.Anal
 EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Analyzer", "Analyzer", "{457CB43B-38FA-4221-BCC2-BE866D0A2A06}"
 EndProject
+Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "Misaki.HighPerformance.Mathematics.SPMD", "Misaki.HighPerformance.Mathematics.SPMD\Misaki.HighPerformance.Mathematics.SPMD.csproj", "{09DF7B3B-F36D-4925-B086-E8827E228FFC}"
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug|Any CPU = Debug|Any CPU
@@ -71,6 +73,10 @@ Global
 		{446213EF-4E44-4EEF-970A-536F0AFC7AEB}.Debug|Any CPU.Build.0 = Debug|Any CPU
 		{446213EF-4E44-4EEF-970A-536F0AFC7AEB}.Release|Any CPU.ActiveCfg = Release|Any CPU
 		{446213EF-4E44-4EEF-970A-536F0AFC7AEB}.Release|Any CPU.Build.0 = Release|Any CPU
+		{09DF7B3B-F36D-4925-B086-E8827E228FFC}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
+		{09DF7B3B-F36D-4925-B086-E8827E228FFC}.Debug|Any CPU.Build.0 = Debug|Any CPU
+		{09DF7B3B-F36D-4925-B086-E8827E228FFC}.Release|Any CPU.ActiveCfg = Release|Any CPU
+		{09DF7B3B-F36D-4925-B086-E8827E228FFC}.Release|Any CPU.Build.0 = Release|Any CPU
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE

Misaki.HighPerformance/Result.cs

@@ -0,0 +1,398 @@
+using System.Runtime.CompilerServices;
+
+namespace Misaki.HighPerformance;
+
+public readonly struct Result
+{
+    private readonly string? _message;
+    private readonly bool _isSuccess;
+
+    public readonly string? Message => _message;
+    public readonly bool IsSuccess => _isSuccess;
+    public readonly bool IsFailure => !IsSuccess;
+
+    public Result(bool success, string? message = null)
+    {
+        _isSuccess = success;
+        _message = message;
+    }
+
+    public static Result Success()
+    {
+        return new Result(true);
+    }
+
+    public static Result Failure(string? message = null)
+    {
+        return new Result(false, message);
+    }
+
+    public static Result Failure(Error status)
+    {
+        return new Result(false, status.ToString());
+    }
+
+    public static Result<T> Success<T>(T value)
+    {
+        return Result<T>.Success(value);
+    }
+
+    public static Result<T> Failure<T>(string? message = null)
+    {
+        return Result<T>.Failure(message);
+    }
+
+    public static Result<T> Failure<T>(Error status)
+    {
+        return Result<T>.Failure(status.ToString());
+    }
+
+    public void Deconstruct(out bool success, out string? message)
+    {
+        success = IsSuccess;
+        message = Message;
+    }
+
+    public override string ToString() => IsSuccess ? "OK" : $"Error: {Message}";
+
+    public static implicit operator bool(Result result) => result.IsSuccess;
+}
+
+public readonly struct Result<T>
+{
+    private readonly T _value;
+    private readonly string? _message;
+    private readonly bool _isSuccess;
+
+    /// <summary>
+    /// Gets the value. Undefined if the result is a failure.
+    /// </summary>
+    public T Value
+    {
+        get
+        {
+#if DEBUG
+            if (IsFailure)
+            {
+                throw new InvalidOperationException($"Cannot access Value when Result is a failure. {_message}");
+            }
+#endif
+            return _value;
+        }
+    }
+
+    public readonly string? Message => _message;
+    public readonly bool IsSuccess => _isSuccess;
+    public readonly bool IsFailure => !IsSuccess;
+
+    public Result(bool success, T value, string? message = null)
+    {
+        _isSuccess = success;
+        _value = value;
+        _message = message;
+    }
+
+    public static Result<T> Success(T value)
+    {
+        return new Result<T>(true, value);
+    }
+
+    public static Result<T> Failure(string? message = null)
+    {
+        return new Result<T>(false, default!, message);
+    }
+
+    public void Deconstruct(out bool success, out T value, out string? message)
+    {
+        success = IsSuccess;
+        value = Value;
+        message = Message;
+    }
+
+    public override string ToString() => IsSuccess ? $"OK: {Value}" : $"Error: {Message}";
+
+    public static implicit operator Result<T>(T? data) => data is not null ? Success(data) : Failure(null);
+    public static implicit operator Result<T>(Result result) => result.IsSuccess ? Success(default!) : Failure(result.Message);
+    public static implicit operator bool(Result<T> result) => result.IsSuccess;
+}
+
+public enum Error : byte
+{
+    None,
+    NotFound,
+    InvalidArgument,
+    InvalidState,
+    InternalError,
+    PermissionDenied,
+    NotSupported,
+    OutOfMemory,
+    Timeout,
+    Cancelled,
+    UnknownError,
+
+    Success = None,
+}
+
+public readonly struct Result<T, E>
+    where E : struct, Enum
+{
+    private readonly T _value;
+    private readonly E _error;
+
+    /// <summary>
+    /// Gets the value. Undefined if the result is a failure.
+    /// </summary>
+    public T Value
+    {
+        get
+        {
+#if DEBUG
+            if (IsFailure)
+            {
+                throw new InvalidOperationException($"Cannot access Value when Result is a failure. Error: {_error}");
+            }
+#endif
+            return _value;
+        }
+    }
+
+    public E Error => _error;
+    public bool IsSuccess => EqualityComparer<E>.Default.Equals(_error, default);
+    public bool IsFailure => !IsSuccess;
+
+    public Result(T value, E status)
+    {
+        _value = value;
+        _error = status;
+    }
+
+    public static Result<T, E> Success(T value)
+    {
+        return new Result<T, E>(value, default);
+    }
+
+    public static Result<T, E> Failure(E status)
+    {
+        return new Result<T, E>(default!, status);
+    }
+
+    public void Deconstruct(out T value, out E status)
+    {
+        value = Value;
+        status = Error;
+    }
+
+    public override string ToString() => $"Value: {_value}, Status: {_error}";
+
+    public static implicit operator Result<T, E>(T data) => new(data, default);
+    public static implicit operator Result<T, E>(E status) => new(default!, status);
+    public static implicit operator bool(Result<T, E> result) => result.IsSuccess;
+}
+
+public readonly ref struct RefResult<T, E>
+    where E : struct, Enum
+{
+    private readonly ref T _value;
+    private readonly E _error;
+
+    /// <summary>
+    /// Gets a reference to the value. Undefined if the result is a failure.
+    /// </summary>
+    public ref T Value
+    {
+        get
+        {
+#if DEBUG
+            if (IsFailure)
+            {
+                throw new InvalidOperationException($"Cannot access Value when Result is a failure. Error: {_error}");
+            }
+#endif
+            return ref _value;
+        }
+    }
+
+    public E Error => _error;
+    public bool IsSuccess => EqualityComparer<E>.Default.Equals(_error, default);
+    public bool IsFailure => !IsSuccess;
+
+    public RefResult(ref T value, E error)
+    {
+        _value = ref value;
+        _error = error;
+    }
+
+    public static RefResult<T, E> Success(ref T value)
+    {
+        return new RefResult<T, E>(ref value, default);
+    }
+
+    public static RefResult<T, E> Failure(E error)
+    {
+        return new RefResult<T, E>(ref Unsafe.NullRef<T>(), error);
+    }
+
+    public override string ToString() => $"Value: {_value}, Status: {_error}";
+
+    public static implicit operator RefResult<T, E>(E error) => new(ref Unsafe.NullRef<T>(), error);
+    public static implicit operator bool(RefResult<T, E> result) => result.IsSuccess;
+}
+
+public static class ResultExtensions
+{
+    public static void ThrowIfFailed(this Error result, [CallerArgumentExpression(nameof(result))] string? op = null)
+    {
+        if (result != Error.None)
+        {
+            throw new InvalidOperationException($"{op} failed: {result}");
+        }
+    }
+
+    public static void ThrowIfFailed(this Result result, [CallerArgumentExpression(nameof(result))] string? op = null)
+    {
+        if (!result.IsSuccess)
+        {
+            throw new InvalidOperationException($"{op} failed: {result.Message}");
+        }
+    }
+
+    public static T GetValueOrThrow<T>(this Result<T> result, [CallerArgumentExpression(nameof(result))] string? op = null)
+    {
+        if (!result.IsSuccess)
+        {
+            throw new InvalidOperationException($"{op} failed: {result.Message}");
+        }
+
+        return result.Value;
+    }
+
+    public static T GetValueOrThrow<T, S>(this Result<T, S> result, [CallerArgumentExpression(nameof(result))] string? op = null)
+        where S : struct, Enum
+    {
+        if (!result.IsSuccess)
+        {
+            throw new InvalidOperationException($"{op} failed: status {result.Error}");
+        }
+
+        return result.Value;
+    }
+
+    public static T? GetValueOrDefault<T>(this Result<T> result, T? defaultValue = default)
+    {
+        return result.IsSuccess ? result.Value : defaultValue;
+    }
+
+    public static T? GetValueOrDefault<T, S>(this Result<T, S> result, T? defaultValue = default)
+        where S : struct, Enum
+    {
+        return result.IsSuccess ? result.Value : defaultValue;
+    }
+
+    public static bool TryGetValue<T>(this Result<T> result, out T value)
+    {
+        if (result.IsSuccess)
+        {
+            value = result.Value;
+            return true;
+        }
+
+        value = default!;
+        return false;
+    }
+
+    public static bool TryGetValue<T, S>(this Result<T, S> result, out T value)
+        where S : struct, Enum
+    {
+        if (result.IsSuccess)
+        {
+            value = result.Value;
+            return true;
+        }
+
+        value = default!;
+        return false;
+    }
+
+    public static Result OnSuccess(this Result result, Action action)
+    {
+        if (result.IsSuccess)
+        {
+            action();
+        }
+
+        return result;
+    }
+
+    public static Result<T> OnSuccess<T>(this Result<T> result, Action<T> action)
+    {
+        if (result.IsSuccess)
+        {
+            action(result.Value);
+        }
+
+        return result;
+    }
+
+    public static Result<T, E> OnSuccess<T, E>(this Result<T, E> result, Action<T> action)
+        where E : struct, Enum
+    {
+        if (result.IsSuccess)
+        {
+            action(result.Value);
+        }
+
+        return result;
+    }
+
+    public static Result OnFailed(this Result result, Action<string?> action)
+    {
+        if (result.IsFailure)
+        {
+            action(result.Message);
+        }
+
+        return result;
+    }
+
+    public static Result<T> OnFailed<T>(this Result<T> result, Action<string?> action)
+    {
+        if (result.IsFailure)
+        {
+            action(result.Message);
+        }
+
+        return result;
+    }
+
+    public static Result<T, E> OnFailed<T, E>(this Result<T, E> result, Action<E> action)
+        where E : struct, Enum
+    {
+        if (result.IsFailure)
+        {
+            action(result.Error);
+        }
+
+        return result;
+    }
+
+    public static Result<U> Then<T, U>(this Result<T> result, Func<T, Result<U>> func)
+    {
+        if (result.IsFailure)
+        {
+            return Result<U>.Failure(result.Message);
+        }
+
+        return func(result.Value);
+    }
+
+    public static Result<U, E> Then<T, U, E>(this Result<T, E> result, Func<T, Result<U, E>> func)
+        where E : struct, Enum
+    {
+        if (result.IsFailure)
+        {
+            return Result<U, E>.Failure(result.Error);
+        }
+
+        return func(result.Value);
+    }
+}