[GPU] Run the command stream on its own thread and speed up draw execution

ASTRO BOT crowd scene: ~25 FPS -> ~55-58 FPS.

- The PM4 command stream is interpreted on its own thread. Draws, dispatches,
  barriers, end-of-pipe writes and the interpreter's stores are handed to the
  render thread in stream order through a short queue (64 items); callbacks
  that return a result wait for the work before them.
- Graphics program resolution runs on a worker thread ahead of the draws, from
  copy-on-write register bank snapshots, and is validated against the GPU
  ownership state before use.
- The resource materialization cache rebases buffer base addresses instead of
  missing on every moved descriptor.
- DCC metadata fills that cannot be absorbed run on the host, so binding the
  surface no longer waits for a GPU readback.
- Cheaper per-draw work: cached pipeline lookups no longer wait on program
  preparation, redundant dynamic state is skipped, image bindings and target
  requests are memoized, fewer allocations.
- Guest thread stack/TLS slots are chosen and mapped under one lock, so two
  threads can no longer share a stack.
- Wave lowering split into native wave32, native wave64 and emulated wave64.

src/SharpEmu.Core/Cpu/Native/DirectExecutionBackend.cs

@@ -5604,6 +5604,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		return false;
 	}
 
+	// Map does not reject a range that is already mapped: it reuses and zero-fills it. Two threads
+	// creating guest threads at once could both find the same slot free and share one stack, so
+	// choosing and mapping a slot happen under one lock.
+	private static readonly object GuestThreadRegionGate = new();
+
 	private static bool TryMapGuestThreadRegion(
 		IVirtualMemory virtualMemory,
 		ulong baseAddress,
@@ -5611,6 +5616,20 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		ProgramHeaderFlags protection,
 		out ulong mappedBase,
 		out string? error)
+	{
+		lock (GuestThreadRegionGate)
+		{
+			return TryMapGuestThreadRegionLocked(virtualMemory, baseAddress, size, protection, out mappedBase, out error);
+		}
+	}
+
+	private static bool TryMapGuestThreadRegionLocked(
+		IVirtualMemory virtualMemory,
+		ulong baseAddress,
+		ulong size,
+		ProgramHeaderFlags protection,
+		out ulong mappedBase,
+		out string? error)
 	{
 		for (int i = 0; i < GuestThreadRegionSlots; i++)
 		{
@@ -5645,6 +5664,17 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		IVirtualMemory virtualMemory,
 		out ulong tlsBase,
 		out string? error)
+	{
+		lock (GuestThreadRegionGate)
+		{
+			return TryMapGuestThreadTlsRegionLocked(virtualMemory, out tlsBase, out error);
+		}
+	}
+
+	private static bool TryMapGuestThreadTlsRegionLocked(
+		IVirtualMemory virtualMemory,
+		out ulong tlsBase,
+		out string? error)
 	{
 		for (int i = 0; i < GuestThreadRegionSlots; i++)
 		{

src/SharpEmu.Libs/Agc/AgcExports.CommandStream.cs

@@ -23,6 +23,21 @@ public static partial class AgcExports
 
     internal sealed record RetainedTargetlessDraw(RegisterBanks Banks, TargetlessDrawArguments Arguments);
 
+    // A draw or dispatch the translation deferred, with the banks it was issued under.
+    internal abstract record DeferredWork;
+
+    // Takes the deferred work of a command stream interpreted on its own thread.
+    internal interface IDeferredWorkSink
+    {
+        // True on the thread that interprets the stream; elsewhere the translation runs work itself.
+        bool IsProducerThread { get; }
+
+        // The non-draw items handed over so far; a draw's resolution waits until they ran.
+        long Gate { get; }
+
+        void EnqueueDeferred(DeferredWork work, bool isDraw);
+    }
+
     // The snapshots the submit-time prepass captured for one submission.
     private sealed record SubmittedGeometrySnapshots(
         Dictionary<ulong, SubmittedIndexSnapshot>? IndexSnapshots,
@@ -160,9 +175,18 @@ public static partial class AgcExports
             {
                 _pipelines = new ShaderPipelineCache(_context, pipelineHost, GuestGpu.Current, CreateShaderHeaderRegistry(_context));
                 _executor = new RenderExecutor(renderHost, _pipelines);
+                if (host is IConcurrentResolveHost concurrentHost)
+                {
+                    _parallel = new ParallelProgramPrefetch(_executor, renderHost, concurrentHost);
+                    _executor.Prefetch = _parallel;
+                }
             }
         }
 
+        // Resolves the programs of deferred draws ahead of them, for a host that answers reads from
+        // another thread; draws are deferred whenever it exists.
+        private readonly ParallelProgramPrefetch? _parallel;
+
         // The context bank of the queue whose slice runs now; the Metal host rebuilds its records from it.
         internal ContextRegisters? CurrentContextRegisters => _current?.TypedRegisters?.Context;
 
@@ -196,17 +220,194 @@ public static partial class AgcExports
             }
         }
 
+        private sealed record PendingDraw(
+            ulong SubmitId,
+            SubmittedDcbState State,
+            RegisterBanks Banks,
+            bool Indexed,
+            DrawIndexedArguments IndexedArguments,
+            DrawAutoArguments AutoArguments,
+            SubmittedVertexSnapshot? VertexSnapshot,
+            SubmittedIndexSnapshot? IndexSnapshot,
+            ParallelProgramPrefetch.Job? Job) : DeferredWork;
+
+        private sealed record PendingDispatch(
+            ulong SubmitId,
+            SubmittedDcbState State,
+            RegisterBanks Banks,
+            uint GroupsX,
+            uint GroupsY,
+            uint GroupsZ,
+            uint DispatchInitiator,
+            ulong IndirectArgumentsAddress) : DeferredWork;
+
+        // With the command stream on its own thread, its draws and dispatches reach the render
+        // thread through the sink, which runs them with RunDeferred in stream order.
+        internal IDeferredWorkSink? Sink { get; set; }
+
+        private bool Producing => Sink is { IsProducerThread: true };
+
+        // The queue state of the work the render thread runs now; _current belongs to the thread
+        // interpreting the stream.
+        private SubmittedDcbState? _executing;
+
+        public void RunDeferred(DeferredWork work)
+        {
+            switch (work)
+            {
+                case PendingDraw draw:
+                {
+                    var executing = _executing;
+                    _executing = draw.State;
+                    RecordKnownColorTargets(draw.State, draw.Banks);
+                    if (_parallel is not null)
+                    {
+                        _parallel.Current = draw.Job;
+                    }
+
+                    try
+                    {
+                        RunExecutorDraw(draw.SubmitId, draw.State, draw.Banks, draw.Indexed, draw.IndexedArguments, draw.AutoArguments,
+                            clearSnapshots: false);
+                    }
+                    finally
+                    {
+                        _executing = executing;
+                        if (_parallel is not null)
+                        {
+                            _parallel.Current = null;
+                        }
+                    }
+
+                    break;
+                }
+                case PendingDispatch dispatch:
+                {
+                    var executing = _executing;
+                    _executing = dispatch.State;
+                    try
+                    {
+                        ExecuteDispatch(_executor!, dispatch.SubmitId, dispatch.Banks, dispatch.GroupsX, dispatch.GroupsY, dispatch.GroupsZ,
+                            dispatch.DispatchInitiator, dispatch.IndirectArgumentsAddress);
+                    }
+                    finally
+                    {
+                        _executing = executing;
+                    }
+
+                    break;
+                }
+                default:
+                    throw _host.Fatal($"The deferred work is unknown: {work.GetType().Name}.");
+            }
+        }
+
+        // The resolution worker, so the render thread can order it with the stream's other work.
+        internal ParallelProgramPrefetch? Prefetch => _parallel;
+
+        private readonly Queue<PendingDraw> _pendingDraws = new();
+
+        public bool HasPendingDraws => _pendingDraws.Count != 0;
+
+        private bool _draining;
+
+        // Runs every deferred draw in issue order with the banks and snapshots it was issued with.
+        // A draw calls back into the host, whose entry points drain too: those calls run inside
+        // the draw and must not start the next one.
+        public void DrainPendingDraws()
+        {
+            if (_draining)
+            {
+                return;
+            }
+
+            _draining = true;
+            try
+            {
+                DrainPendingDrawsCore();
+            }
+            finally
+            {
+                _draining = false;
+            }
+        }
+
+        private void DrainPendingDrawsCore()
+        {
+            while (_pendingDraws.TryDequeue(out var pending))
+            {
+                var current = _current;
+                var state = pending.State;
+                _current = state;
+                state.CurrentVertexSnapshot = pending.VertexSnapshot;
+                state.CurrentIndexSnapshot = pending.IndexSnapshot;
+                if (_parallel is not null)
+                {
+                    _parallel.Current = pending.Job;
+                }
+
+                try
+                {
+                    RunExecutorDraw(pending.SubmitId, state, pending.Banks, pending.Indexed, pending.IndexedArguments, pending.AutoArguments);
+                }
+                finally
+                {
+                    _current = current;
+                    if (_parallel is not null)
+                    {
+                        _parallel.Current = null;
+                    }
+                }
+            }
+
+            // What follows a drain may change guest memory and cache state the worker reads.
+            _parallel?.WaitIdle();
+        }
+
         private bool TryRunExecutorDraw(ulong submitId, SubmittedDcbState state, bool indexed, in DrawIndexedArguments indexedArguments, in DrawAutoArguments autoArguments)
         {
-            if (_executor is not { } executor)
+            if (_executor is null)
             {
                 return false;
             }
 
             var banks = RequireTypedRegisters(state);
-            RecordKnownColorTargets(state, banks);
+            var producing = Producing;
+            if (!producing)
+            {
+                // A draw run by the render thread records its targets there.
+                RecordKnownColorTargets(state, banks);
+            }
+
             state.FrameDrawCount++;
             state.SawIndexedDraw |= indexed;
+            if (_parallel is not null || producing)
+            {
+                var snapshot = banks.Snapshot();
+                var pending = new PendingDraw(submitId, state, snapshot, indexed, indexedArguments, autoArguments,
+                    state.CurrentVertexSnapshot, state.CurrentIndexSnapshot, _parallel?.Submit(snapshot, producing ? Sink!.Gate : 0));
+                if (producing)
+                {
+                    Sink!.EnqueueDeferred(pending, isDraw: true);
+                }
+                else
+                {
+                    _pendingDraws.Enqueue(pending);
+                }
+
+                state.CurrentIndexSnapshot = null;
+                state.CurrentVertexSnapshot = null;
+                return true;
+            }
+
+            RunExecutorDraw(submitId, state, banks, indexed, indexedArguments, autoArguments);
+            return true;
+        }
+
+        private void RunExecutorDraw(ulong submitId, SubmittedDcbState state, RegisterBanks banks, bool indexed, in DrawIndexedArguments indexedArguments, in DrawAutoArguments autoArguments,
+            bool clearSnapshots = true)
+        {
+            var executor = _executor!;
             var drawStarted = DcbParseProfile.Begin();
             try
             {
@@ -222,17 +423,19 @@ public static partial class AgcExports
             finally
             {
                 DcbParseProfile.RecordDraw(drawStarted);
-                state.CurrentIndexSnapshot = null;
-                state.CurrentVertexSnapshot = null;
+                // The snapshot fields belong to the interpreting thread when it is another one.
+                if (clearSnapshots)
+                {
+                    state.CurrentIndexSnapshot = null;
+                    state.CurrentVertexSnapshot = null;
+                }
             }
-
-            return true;
         }
 
         // Keeps the banks the draw was issued under; the flip replays it into the display buffer.
         internal void RetainTargetlessDraw(RegisterBanks banks, in TargetlessDrawArguments arguments)
         {
-            var state = RequireCurrent();
+            var state = _executing ?? RequireCurrent();
             if (RenderTrace.Enabled)
             {
                 RenderTrace.Write($"TargetlessRetention submit={arguments.SubmitId} replaced={state.RetainedTargetlessDraw is not null} previousSubmit={state.RetainedTargetlessDraw?.Arguments.SubmitId} export=0x{banks.Shader.Vertex.ExportAddress:X16} pixel=0x{banks.Shader.Pixel.Address:X16}");
@@ -363,22 +566,35 @@ public static partial class AgcExports
             {
                 var banks = RequireTypedRegisters(state);
                 state.FrameDispatchCount++;
-                var executorStarted = DcbParseProfile.Begin();
-                try
+                if (Producing)
                 {
-                    executor.Dispatch(submitId, banks, endX, endY, endZ, dispatchInitiator, indirectArgumentsAddress);
-                }
-                finally
-                {
-                    DcbParseProfile.RecordDispatch(executorStarted);
+                    var snapshot = banks.Snapshot();
+                    Sink!.EnqueueDeferred(new PendingDispatch(submitId, state, snapshot, endX, endY, endZ, dispatchInitiator, indirectArgumentsAddress),
+                        isDraw: false);
+                    return;
                 }
 
+                ExecuteDispatch(executor, submitId, banks, endX, endY, endZ, dispatchInitiator, indirectArgumentsAddress);
                 return;
             }
 
             throw _host.Fatal("The command stream has no render executor.");
         }
 
+        private static void ExecuteDispatch(RenderExecutor executor, ulong submitId, RegisterBanks banks, uint endX, uint endY, uint endZ,
+            uint dispatchInitiator, ulong indirectArgumentsAddress)
+        {
+            var executorStarted = DcbParseProfile.Begin();
+            try
+            {
+                executor.Dispatch(submitId, banks, endX, endY, endZ, dispatchInitiator, indirectArgumentsAddress);
+            }
+            finally
+            {
+                DcbParseProfile.RecordDispatch(executorStarted);
+            }
+        }
+
             // The retained draw runs into the display buffer with the target words it was last bound with.
         private void ReplayRetainedTargetlessDraw(RenderExecutor executor, SubmittedDcbState state, int handle, int displayBufferIndex)
         {

src/SharpEmu.Libs/Gpu/Buffers/GuestBufferCache.cs

@@ -333,6 +333,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
             throw SubmissionScheduler.Fatal("The host DMA write range is invalid.");
         }
 
+        GpuWritebackEpoch.Advance();
         if (!_backing.TryWriteBacking(guestAddress, data))
         {
             throw SubmissionScheduler.Fatal($"Could not write the required direct backing: addr=0x{guestAddress:X16} size=0x{data.Length:X16}");
@@ -413,26 +414,71 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
         if (images.OverlapsDccMetadata(guestAddress, size) && TryWriteFillToBacking(guestAddress, size, value))
         {
             _gpuModifiedRanges.Remove(guestAddress, size);
-            var firstPage = guestAddress & ~(TrackerLayout.PageBytes - 1);
-            for (var page = firstPage; page < guestAddress + size; page += TrackerLayout.PageBytes)
-            {
-                if (!_gpuModifiedRanges.Overlaps(page, TrackerLayout.PageBytes))
-                {
-                    _tracker.ClearGpuDirtyPages(page, TrackerLayout.PageBytes);
-                }
-            }
+            Interlocked.Increment(ref _gpuModifiedVersion);
+            ClearGpuDirtyPagesOfRemovedRange(guestAddress, size);
+        }
+    }
+
+    // After a range left the GPU-modified set: the pages inside it are clean in one tracker walk,
+    // and only a page the range shares at either end can still hold another GPU-modified range.
+    private void ClearGpuDirtyPagesOfRemovedRange(ulong guestAddress, ulong size)
+    {
+        const ulong pageBytes = TrackerLayout.PageBytes;
+        var end = guestAddress + size;
+        var firstPage = guestAddress & ~(pageBytes - 1);
+        var lastPage = (end - 1) & ~(pageBytes - 1);
+        var startPartial = guestAddress != firstPage;
+        var endPartial = end != lastPage + pageBytes;
+        var innerBegin = startPartial ? firstPage + pageBytes : firstPage;
+        var innerEnd = endPartial ? lastPage : end;
+        if (innerBegin < innerEnd)
+        {
+            _tracker.ClearGpuDirtyPages(innerBegin, innerEnd - innerBegin);
+        }
+
+        if (startPartial)
+        {
+            ClearGpuDirtyPageIfUnmodified(firstPage);
+        }
+
+        if (endPartial && (lastPage != firstPage || !startPartial))
+        {
+            ClearGpuDirtyPageIfUnmodified(lastPage);
+        }
+    }
+
+    private void ClearGpuDirtyPageIfUnmodified(ulong page)
+    {
+        if (!_gpuModifiedRanges.Overlaps(page, TrackerLayout.PageBytes))
+        {
+            _tracker.ClearGpuDirtyPages(page, TrackerLayout.PageBytes);
         }
     }
 
     public void FillDccMetadata(ulong guestAddress, ulong size, uint value)
     {
         if (guestAddress == 0 || (guestAddress & 3) != 0 || size == 0 || (size & 3) != 0 || size > ulong.MaxValue - guestAddress ||
-            HasGpuDirtyBytes(guestAddress, size) || RequireImageCache().QueryRegion(guestAddress, size).ImageBytes)
+            HasGpuDirtyBytes(guestAddress, size))
         {
             FillBuffer(guestAddress, size, value, false);
             return;
         }
 
+        // Images over the range read it again; only GPU-owned image bytes need the general fill.
+        // The buffer copy is filled on the GPU rather than uploaded from the written bytes.
+        var images = RequireImageCache();
+        var region = images.QueryRegion(guestAddress, size);
+        if (region.GpuImageBytes)
+        {
+            FillBuffer(guestAddress, size, value, false);
+            return;
+        }
+
+        if (region.ImageBytes)
+        {
+            images.InvalidateMemory(guestAddress, size);
+        }
+
         var (destination, destinationOffset) = ObtainBuffer(guestAddress, size, false, false, FindBuffer(guestAddress, size));
         destination.Fill(destinationOffset, size, value);
         if (!TryWriteFillToBacking(guestAddress, size, value))
@@ -443,21 +489,31 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
 
     private bool TryWriteFillToBacking(ulong guestAddress, ulong size, uint value)
     {
-        var values = new uint[(int)Math.Min(size / sizeof(uint), 4096)];
-        Array.Fill(values, value);
-        var bytes = MemoryMarshal.AsBytes<uint>(values);
-        for (ulong offset = 0; offset < size;)
+        // Metadata fills repeat every frame: the pattern buffer is pooled.
+        var count = (int)Math.Min(size / sizeof(uint), 4096);
+        var values = System.Buffers.ArrayPool<uint>.Shared.Rent(count);
+        try
         {
-            var chunk = (int)Math.Min(size - offset, (ulong)bytes.Length);
-            if (!_backing.TryWriteBacking(guestAddress + offset, bytes[..chunk]))
+            values.AsSpan(0, count).Fill(value);
+            var bytes = MemoryMarshal.AsBytes(values.AsSpan(0, count));
+            for (ulong offset = 0; offset < size;)
             {
-                return false;
+                var chunk = (int)Math.Min(size - offset, (ulong)bytes.Length);
+                GpuWritebackEpoch.Advance();
+                if (!_backing.TryWriteBacking(guestAddress + offset, bytes[..chunk]))
+                {
+                    return false;
+                }
+
+                offset += (ulong)chunk;
             }
 
-            offset += (ulong)chunk;
+            return true;
+        }
+        finally
+        {
+            System.Buffers.ArrayPool<uint>.Shared.Return(values);
         }
-
-        return true;
     }
 
     public const ulong MaxHostCopyWords = 64 * 1024;
@@ -1241,6 +1297,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
             var copies = pending.Copies;
             readback.Complete(pending.Ticket, (index, bytes) =>
             {
+                GpuWritebackEpoch.Advance();
                 if (!_backing.TryWriteBacking(copies[index].Address, bytes))
                 {
                     throw SubmissionScheduler.Fatal($"Could not write the required direct backing: addr=0x{copies[index].Address:X16} size=0x{(ulong)bytes.Length:X16}");
@@ -1249,6 +1306,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
             foreach (var copy in copies)
             {
                 _gpuModifiedRanges.Remove(copy.Address, copy.Size);
+                Interlocked.Increment(ref _gpuModifiedVersion);
             }
 
             _tracker.ClearGpuDirtyPages(pending.WindowBegin, pending.WindowEnd - pending.WindowBegin);
@@ -1456,6 +1514,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
             foreach (var copy in copies)
             {
                 _gpuModifiedRanges.Remove(copy.Address, copy.Size);
+                Interlocked.Increment(ref _gpuModifiedVersion);
             }
 
             return;
@@ -1492,6 +1551,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
         foreach (var copy in copies)
         {
             _gpuModifiedRanges.Remove(copy.Address, copy.Size);
+            Interlocked.Increment(ref _gpuModifiedVersion);
         }
     }
 
@@ -1590,6 +1650,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
 
         readback.Read(pieces, waitTick, (index, bytes) =>
         {
+            GpuWritebackEpoch.Advance();
             if (!_backing.TryWriteBacking(copies[index].Address, bytes))
             {
                 throw SubmissionScheduler.Fatal($"Could not write the required direct backing: addr=0x{copies[index].Address:X16} size=0x{(ulong)bytes.Length:X16}");
@@ -1622,6 +1683,7 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
             var placement = copy.Placement;
             var offset = baseOffset + placement.DataOffset;
             _download.Invalidate(offset, placement.DataSize);
+            GpuWritebackEpoch.Advance();
             if (!_backing.TryWriteBacking(copy.Address, _download.Mapped.Slice((int)offset, (int)placement.DataSize)))
             {
                 throw SubmissionScheduler.Fatal($"Could not write the required direct backing: addr=0x{copy.Address:X16} size=0x{placement.DataSize:X16}");
@@ -1713,20 +1775,79 @@ public sealed unsafe class GuestBufferCache : IGuestBufferStore, IDisposable
         profileScope.SwitchPhase(isWritten
             ? RenderPhaseProfile.Phase.BufferDirtySyncWritten
             : isTexelBuffer ? RenderPhaseProfile.Phase.BufferDirtySyncTexel : RenderPhaseProfile.Phase.BufferDirtySyncUpload);
-        var copies = new List<BufferCopy>();
-        var totalSize = 0UL;
-        GpuBuffer? source = null;
-        _tracker.ForEachUploadRange(
-            guestAddress,
-            size,
-            isWritten,
-            (address, bytes) =>
+        var collector = _uploadCollector;
+        if (collector is null || collector.InUse)
+        {
+            // The first call on a thread keeps its collector; a nested one gets its own.
+            collector = new UploadCollector();
+            _uploadCollector ??= collector;
+        }
+
+        collector.Begin(this, buffer, guestAddress, size);
+        try
+        {
+            return SynchronizeBufferCore(collector, buffer, guestAddress, size, isWritten, isTexelBuffer, preserveCpuWriteHotPages, startedAt);
+        }
+        finally
+        {
+            collector.End();
+        }
+    }
+
+    // The ranges of one synchronization. The delegates are bound once, so a call that has dirty
+    // pages to upload allocates no closure and no list.
+    private sealed class UploadCollector
+    {
+        private GuestBufferCache _cache = null!;
+        private GpuBuffer _buffer = null!;
+        private ulong _guestAddress;
+        private ulong _size;
+        public readonly List<BufferCopy> Copies = new();
+        public ulong TotalSize;
+        public GpuBuffer? Source;
+        public bool InUse;
+        public readonly Action<ulong, ulong> Collect;
+        public readonly Action Upload;
+
+        public UploadCollector()
+        {
+            Collect = (address, bytes) =>
             {
-                copies.Add(new BufferCopy(totalSize, buffer.Offset(address), bytes));
-                totalSize += bytes;
-            },
-            () => source = _uploader.PrepareSource(buffer.CpuAddress, CollectionsMarshal.AsSpan(copies), totalSize, guestAddress, size),
-            preserveCpuWriteHotPages);
+                Copies.Add(new BufferCopy(TotalSize, _buffer.Offset(address), bytes));
+                TotalSize += bytes;
+            };
+            Upload = () => Source = _cache._uploader.PrepareSource(_buffer.CpuAddress, CollectionsMarshal.AsSpan(Copies), TotalSize, _guestAddress, _size);
+        }
+
+        public void Begin(GuestBufferCache cache, GpuBuffer buffer, ulong guestAddress, ulong size)
+        {
+            InUse = true;
+            _cache = cache;
+            _buffer = buffer;
+            _guestAddress = guestAddress;
+            _size = size;
+        }
+
+        public void End()
+        {
+            InUse = false;
+            _cache = null!;
+            _buffer = null!;
+            Copies.Clear();
+            TotalSize = 0;
+            Source = null;
+        }
+    }
+
+    [ThreadStatic] private static UploadCollector? _uploadCollector;
+
+    private bool SynchronizeBufferCore(UploadCollector collector, GpuBuffer buffer, ulong guestAddress, ulong size, bool isWritten,
+        bool isTexelBuffer, bool preserveCpuWriteHotPages, long startedAt)
+    {
+        var copies = collector.Copies;
+        _tracker.ForEachUploadRange(guestAddress, size, isWritten, collector.Collect, collector.Upload, preserveCpuWriteHotPages);
+        var source = collector.Source;
+        var totalSize = collector.TotalSize;
         if (source != null)
         {
             buffer.NoteGpuWrite();

src/SharpEmu.Libs/Gpu/GpuCommands/GpuCommandInterpreter.Adaptations.cs

@@ -26,7 +26,7 @@ public sealed partial class GpuCommandInterpreter
         }
 
         // A compare-swap loop applies once per try and suspends while the compare fails.
-        if (!atomic.TryApply(_host.TryReadGuest, _host.Memory, out var priorValue, out _, out var comparePassed))
+        if (!atomic.TryApply(_host.TryReadGuestOperand, _host.Memory, out var priorValue, out _, out var comparePassed))
         {
             throw _host.Fatal($"The atomic address cannot be accessed: address=0x{atomic.Address:X16} size={atomic.ByteCount}.");
         }
@@ -68,7 +68,7 @@ public sealed partial class GpuCommandInterpreter
 
         if (semaphore.IsSignal)
         {
-            if (!Packets.MemorySemaphorePacket.TrySignal(_host.TryReadGuest, _host.Memory, semaphore.Address, semaphore.WriteSignal, out _))
+            if (!Packets.MemorySemaphorePacket.TrySignal(_host.TryReadGuestOperand, _host.Memory, semaphore.Address, semaphore.WriteSignal, out _))
             {
                 throw _host.Fatal($"The semaphore cannot be signaled: address=0x{semaphore.Address:X16}.");
             }
@@ -77,7 +77,7 @@ public sealed partial class GpuCommandInterpreter
         }
 
         // A wait takes one token, or suspends until a signal adds one.
-        if (!Packets.MemorySemaphorePacket.TryConsume(_host.TryReadGuest, _host.Memory, semaphore.Address, out var priorValue))
+        if (!Packets.MemorySemaphorePacket.TryConsume(_host.TryReadGuestOperand, _host.Memory, semaphore.Address, out var priorValue))
         {
             throw _host.Fatal($"The semaphore cannot be read: address=0x{semaphore.Address:X16}.");
         }

src/SharpEmu.Libs/Gpu/GpuCommands/GpuCommandInterpreter.Registers.cs

@@ -82,8 +82,8 @@ public sealed partial class GpuCommandInterpreter
         var next = packet.PacketAddress + (ulong)packet.Length * sizeof(uint);
         var expectedHeader = PacketHeader.Make(centroidPacketDwords, PacketOpcode.SetContextRegister);
         if (packet.Remaining < packet.Length + centroidPacketDwords ||
-            ReadDword(next) != expectedHeader ||
-            RegisterBankLayout.Normalize(ReadDword(next + sizeof(uint))) != ContextRegisterOffset.PaScCentroidPriority0)
+            ReadDword(next, RenderPhaseProfile.CommandReadKind.Payload) != expectedHeader ||
+            RegisterBankLayout.Normalize(ReadDword(next + sizeof(uint), RenderPhaseProfile.CommandReadKind.Payload)) != ContextRegisterOffset.PaScCentroidPriority0)
         {
             throw _host.Fatal($"The sample location packet is not followed by the centroid priority packet: header=0x{packet.Header:X8} address=0x{packet.PacketAddress:X16} next=0x{next:X16}.");
         }
@@ -102,16 +102,16 @@ public sealed partial class GpuCommandInterpreter
 
         var address = packet.PacketAddress;
         var registerHeader = PacketHeader.Make(3, PacketOpcode.SetContextRegister);
-        if (ReadDword(address + 40) != registerHeader || ReadDword(address + 44) != RegisterBankLayout.DepthInfo ||
-            ReadDword(address + 52) != registerHeader || ReadDword(address + 56) != RegisterBankLayout.DepthView ||
-            ReadDword(address + 64) != registerHeader || ReadDword(address + 68) != RegisterBankLayout.HtileDataBase ||
-            ReadDword(address + 76) != registerHeader || ReadDword(address + 80) != RegisterBankLayout.HtileSurface ||
-            ReadDword(address + 88) != PacketHeader.Make(2, PacketOpcode.Nop))
+        if (ReadDword(address + 40, RenderPhaseProfile.CommandReadKind.Payload) != registerHeader || ReadDword(address + 44, RenderPhaseProfile.CommandReadKind.Payload) != RegisterBankLayout.DepthInfo ||
+            ReadDword(address + 52, RenderPhaseProfile.CommandReadKind.Payload) != registerHeader || ReadDword(address + 56, RenderPhaseProfile.CommandReadKind.Payload) != RegisterBankLayout.DepthView ||
+            ReadDword(address + 64, RenderPhaseProfile.CommandReadKind.Payload) != registerHeader || ReadDword(address + 68, RenderPhaseProfile.CommandReadKind.Payload) != RegisterBankLayout.HtileDataBase ||
+            ReadDword(address + 76, RenderPhaseProfile.CommandReadKind.Payload) != registerHeader || ReadDword(address + 80, RenderPhaseProfile.CommandReadKind.Payload) != RegisterBankLayout.HtileSurface ||
+            ReadDword(address + 88, RenderPhaseProfile.CommandReadKind.Payload) != PacketHeader.Make(2, PacketOpcode.Nop))
         {
             return;
         }
 
-        var sizeXy = ReadDword(address + 92);
+        var sizeXy = ReadDword(address + 92, RenderPhaseProfile.CommandReadKind.Payload);
         if (sizeXy != 0)
         {
             TypedRegisters.CompositeDepthSizeXy = sizeXy;

src/SharpEmu.Libs/Gpu/GpuCommands/GpuCommandInterpreter.Waits.cs

@@ -28,7 +28,7 @@ public sealed partial class GpuCommandInterpreter
             return;
         }
 
-        var value = is64Bit ? ReadQword(address) : ReadDword(address);
+        var value = ReadWaitOperand(address, is64Bit);
         if (!WaitOperation.TryCompare(value, reference, mask, compareFunction, out var satisfied))
         {
             throw _host.Fatal($"The wait compare function is unknown: function={compareFunction} address=0x{address:X16}.");

src/SharpEmu.Libs/Gpu/GpuCommands/GpuCommandInterpreter.cs

@@ -394,7 +394,12 @@ public sealed partial class GpuCommandInterpreter
     {
         RenderPhaseProfile.RecordCommandRead(readKind, sizeof(uint));
         Span<byte> bytes = stackalloc byte[sizeof(uint)];
-        if (!_host.TryReadGuest(address, bytes))
+        // The packets themselves are command data; anything else is an operand earlier work may write.
+        var read = readKind is RenderPhaseProfile.CommandReadKind.Header or RenderPhaseProfile.CommandReadKind.Payload or
+            RenderPhaseProfile.CommandReadKind.RegisterTable
+            ? _host.TryReadGuest(address, bytes)
+            : _host.TryReadGuestOperand(address, bytes);
+        if (!read)
         {
             throw _host.Fatal($"The command stream cannot read guest memory: address=0x{address:X16} size=4.");
         }
@@ -406,7 +411,7 @@ public sealed partial class GpuCommandInterpreter
     {
         RenderPhaseProfile.RecordCommandRead(RenderPhaseProfile.CommandReadKind.Operand64, sizeof(ulong));
         Span<byte> bytes = stackalloc byte[sizeof(ulong)];
-        if (!_host.TryReadGuest(address, bytes))
+        if (!_host.TryReadGuestOperand(address, bytes))
         {
             throw _host.Fatal($"The command stream cannot read guest memory: address=0x{address:X16} size=8.");
         }
@@ -414,10 +419,24 @@ public sealed partial class GpuCommandInterpreter
         return BinaryPrimitives.ReadUInt64LittleEndian(bytes);
     }
 
+    private ulong ReadWaitOperand(ulong address, bool is64Bit)
+    {
+        RenderPhaseProfile.RecordCommandRead(is64Bit ? RenderPhaseProfile.CommandReadKind.Operand64 : RenderPhaseProfile.CommandReadKind.Operand32,
+            is64Bit ? sizeof(ulong) : sizeof(uint));
+        Span<byte> bytes = stackalloc byte[sizeof(ulong)];
+        var size = is64Bit ? sizeof(ulong) : sizeof(uint);
+        if (!_host.TryReadGuestWaitOperand(address, bytes[..size]))
+        {
+            throw _host.Fatal($"The command stream cannot read guest memory: address=0x{address:X16} size={size}.");
+        }
+
+        return is64Bit ? BinaryPrimitives.ReadUInt64LittleEndian(bytes) : BinaryPrimitives.ReadUInt32LittleEndian(bytes);
+    }
+
     internal void ReadBytes(ulong address, Span<byte> destination)
     {
         RenderPhaseProfile.RecordCommandRead(RenderPhaseProfile.CommandReadKind.Other, destination.Length);
-        if (!_host.TryReadGuest(address, destination))
+        if (!_host.TryReadGuestOperand(address, destination))
         {
             throw _host.Fatal($"The command stream cannot read guest memory: address=0x{address:X16} size={destination.Length}.");
         }

src/SharpEmu.Libs/Gpu/GpuCommands/ICommandStreamHost.cs

@@ -84,6 +84,14 @@ public interface ICommandStreamHost
     // Reads guest memory the GPU may have written; the host synchronizes GPU-owned pages first.
     bool TryReadGuest(ulong address, Span<byte> destination);
 
+    // Reads a packet operand (a label, a predicate, indirect arguments): memory earlier work may
+    // write, unlike the packets themselves. A host that defers work completes it first.
+    bool TryReadGuestOperand(ulong address, Span<byte> destination) => TryReadGuest(address, destination);
+
+    // Reads the label a wait polls. A stale value only suspends the wait until a retry, so a host
+    // that runs work behind the stream need not complete it first.
+    bool TryReadGuestWaitOperand(ulong address, Span<byte> destination) => TryReadGuestOperand(address, destination);
+
     // Runs commands other threads posted to this worker. Called before every packet.
     void RunPendingCommands();
 

src/SharpEmu.Libs/Gpu/GpuCommands/Registers/ContextRegisters.cs

@@ -338,8 +338,15 @@ public sealed class ScreenViewportRegisters
 
     public ScreenViewportRegisters Copy()
     {
-        var copy = (ScreenViewportRegisters)MemberwiseClone();
+        var copy = CopySharingViewports();
         copy.Viewports = (ViewportRegisters[])Viewports.Clone();
+        return copy;
+    }
+
+    // A copy that shares the viewport array, the largest part, until RegisterBanks copies it.
+    public ScreenViewportRegisters CopySharingViewports()
+    {
+        var copy = (ScreenViewportRegisters)MemberwiseClone();
         copy.ClipRectangleLeft = (int[])ClipRectangleLeft.Clone();
         copy.ClipRectangleTop = (int[])ClipRectangleTop.Clone();
         copy.ClipRectangleRight = (int[])ClipRectangleRight.Clone();
@@ -388,6 +395,9 @@ public sealed class ContextRegisters
     public EnhancedQualityAntialiasingRegisters EnhancedQualityAntialiasing;
     public ShaderInterfaceRegisters ShaderInterface = new();
 
+    // A copy that shares every nested part; RegisterBanks copies a part before writing it.
+    public ContextRegisters ShallowCopy() => (ContextRegisters)MemberwiseClone();
+
     public uint RenderTargetMaskForSlot(uint slot) => (RenderTargetMask >> (int)(slot * 4)) & 0xFu;
 
     public ContextRegisters Copy()

src/SharpEmu.Libs/Gpu/GpuCommands/Registers/RegisterBanks.cs

@@ -1,6 +1,8 @@
 // Copyright (C) 2026 SharpEmu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+using SharpEmu.Libs.Gpu.Images;
+
 namespace SharpEmu.Libs.Gpu.GpuCommands.Registers;
 
 // The three typed register banks of one queue, with the state the register writers carry between packets.
@@ -9,15 +11,31 @@ public sealed class RegisterBanks
     private ContextRegisters? _savedContext;
     private uint? _savedCompositeDepthSizeXy;
 
-    public RegisterBanks(Func<string, Exception> fatal) => Fatal = fatal;
+    public RegisterBanks(Func<string, Exception> fatal)
+        : this(fatal, new ContextRegisters(), new ShaderProgramRegisters(), new UserConfigRegisters())
+    {
+    }
+
+    // Banks over given parts, so a copy or a snapshot does not first build empty ones.
+    private RegisterBanks(
+        Func<string, Exception> fatal,
+        ContextRegisters context,
+        ShaderProgramRegisters shader,
+        UserConfigRegisters userConfig)
+    {
+        Fatal = fatal;
+        Context = context;
+        Shader = shader;
+        UserConfig = userConfig;
+    }
 
     public Func<string, Exception> Fatal { get; }
 
-    public ContextRegisters Context { get; private set; } = new();
+    public ContextRegisters Context { get; private set; }
 
-    public ShaderProgramRegisters Shader { get; private set; } = new();
+    public ShaderProgramRegisters Shader { get; private set; }
 
-    public UserConfigRegisters UserConfig { get; private set; } = new();
+    public UserConfigRegisters UserConfig { get; private set; }
 
     public bool ContextPushed => _savedContext is not null;
 
@@ -31,11 +49,8 @@ public sealed class RegisterBanks
     // A copy of the live banks; the saved context and the draw-time markers are not part of it.
     public RegisterBanks Clone()
     {
-        var clone = new RegisterBanks(Fatal)
+        var clone = new RegisterBanks(Fatal, Context.Copy(), Shader.Copy(), UserConfig.Copy())
         {
-            Context = Context.Copy(),
-            Shader = Shader.Copy(),
-            UserConfig = UserConfig.Copy(),
             CompositeDepthSizeXy = CompositeDepthSizeXy,
             UserDataMarker = UserDataMarker,
             IndexTypeAndSize = IndexTypeAndSize,
@@ -43,9 +58,281 @@ public sealed class RegisterBanks
         return clone;
     }
 
+    // The parts a snapshot shares with these banks; each is copied before its next write here.
+    private bool _contextShared;
+    private bool _vertexShared;
+    private bool _pixelShared;
+    private bool _computeShared;
+    private bool _userConfigShared;
+
+    // A read-only copy that shares every part with the live banks until they next write it: a
+    // deferred draw takes one per draw, and between draws only a few parts change.
+    public RegisterBanks Snapshot()
+    {
+        var snapshot = new RegisterBanks(Fatal, Context, Shader.ShareStages(), UserConfig)
+        {
+            CompositeDepthSizeXy = CompositeDepthSizeXy,
+            UserDataMarker = UserDataMarker,
+            IndexTypeAndSize = IndexTypeAndSize,
+        };
+        _contextShared = _vertexShared = _pixelShared = _computeShared = _userConfigShared = true;
+        _pixelScalarsShared = _hullScalarsShared = _geometryScalarsShared = _legacyVertexScalarsShared = _exportScalarsShared = _computeScalarsShared = true;
+        return snapshot;
+    }
+
+    // Called by the register writers before they store into a part.
+    public void PrepareContextWrite()
+    {
+        if (_contextShared)
+        {
+            Context = Context.ShallowCopy();
+            _contextShared = false;
+            SetContextPartsShared(true);
+        }
+    }
+
+    // The context is copied without its nested parts, and a part only when a writer asks for it
+    // through these accessors: a draw usually rewrites a few plain fields.
+    private bool _screenViewportShared;
+    private bool _viewportsShared;
+    private bool _sampleLocationsShared;
+    private bool _shaderInterfaceShared;
+    private bool _blendControlsShared;
+    private bool _colorTargetsShared;
+    private bool _colorClearWord1Shared;
+    private bool _unmodeledShared;
+
+    private void SetContextPartsShared(bool shared) =>
+        _screenViewportShared = _viewportsShared = _sampleLocationsShared = _shaderInterfaceShared =
+            _blendControlsShared = _colorTargetsShared = _colorClearWord1Shared = _unmodeledShared = shared;
+
+    private void OwnContext(ContextRegisters context)
+    {
+        Context = context;
+        _contextShared = false;
+        SetContextPartsShared(false);
+    }
+
+    public ScreenViewportRegisters ScreenViewportForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_screenViewportShared)
+            {
+                Context.ScreenViewport = Context.ScreenViewport.CopySharingViewports();
+                _screenViewportShared = false;
+            }
+
+            return Context.ScreenViewport;
+        }
+    }
+
+    public ViewportRegisters[] ViewportsForWrite
+    {
+        get
+        {
+            var screenViewport = ScreenViewportForWrite;
+            if (_viewportsShared)
+            {
+                screenViewport.Viewports = (ViewportRegisters[])screenViewport.Viewports.Clone();
+                _viewportsShared = false;
+            }
+
+            return screenViewport.Viewports;
+        }
+    }
+
+    public SampleLocationRegisters SampleLocationsForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_sampleLocationsShared)
+            {
+                Context.SampleLocations = Context.SampleLocations.Copy();
+                _sampleLocationsShared = false;
+            }
+
+            return Context.SampleLocations;
+        }
+    }
+
+    public ShaderInterfaceRegisters ShaderInterfaceForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_shaderInterfaceShared)
+            {
+                Context.ShaderInterface = Context.ShaderInterface.Copy();
+                _shaderInterfaceShared = false;
+            }
+
+            return Context.ShaderInterface;
+        }
+    }
+
+    public BlendRegisters[] BlendControlsForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_blendControlsShared)
+            {
+                Context.BlendControls = (BlendRegisters[])Context.BlendControls.Clone();
+                _blendControlsShared = false;
+            }
+
+            return Context.BlendControls;
+        }
+    }
+
+    public ColorTargetWords[] ColorTargetsForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_colorTargetsShared)
+            {
+                Context.ColorTargets = (ColorTargetWords[])Context.ColorTargets.Clone();
+                _colorTargetsShared = false;
+            }
+
+            return Context.ColorTargets;
+        }
+    }
+
+    public uint[] ColorClearWord1ForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_colorClearWord1Shared)
+            {
+                Context.ColorClearWord1 = (uint[])Context.ColorClearWord1.Clone();
+                _colorClearWord1Shared = false;
+            }
+
+            return Context.ColorClearWord1;
+        }
+    }
+
+    public Dictionary<uint, uint> UnmodeledTableRegistersForWrite
+    {
+        get
+        {
+            PrepareContextWrite();
+            if (_unmodeledShared)
+            {
+                Context.UnmodeledTableRegisters = new Dictionary<uint, uint>(Context.UnmodeledTableRegisters);
+                _unmodeledShared = false;
+            }
+
+            return Context.UnmodeledTableRegisters;
+        }
+    }
+
+    public void PrepareUserConfigWrite()
+    {
+        if (_userConfigShared)
+        {
+            UserConfig = UserConfig.Copy();
+            _userConfigShared = false;
+        }
+    }
+
+    // COMPUTE_DISPATCH_INITIATOR, the first register of the compute range.
+    private const uint ComputeRangeStart = 0x200;
+
+    // A stage is copied without its user scalar sets, and a set only when its window is written:
+    // a draw usually rewrites one set of one or two stages.
+    private bool _pixelScalarsShared;
+    private bool _hullScalarsShared;
+    private bool _geometryScalarsShared;
+    private bool _legacyVertexScalarsShared;
+    private bool _exportScalarsShared;
+    private bool _computeScalarsShared;
+    private const uint UserScalarWindow = UserScalarRegisters.Capacity;
+
+    private static bool Overlaps(uint first, uint last, uint windowStart) =>
+        first < windowStart + UserScalarWindow && last >= windowStart;
+
+    // The shader bank's offsets: pixel below SPI_SHADER_PGM_LO_VS, the geometry stages up to the
+    // compute range, compute from there.
+    public void PrepareShaderWrite(uint offset, uint count)
+    {
+        var last = offset + Math.Max(count, 1) - 1;
+        if (offset < ShaderRegisterOffset.SpiShaderPgmLoVs)
+        {
+            if (_pixelShared)
+            {
+                Shader.Pixel = Shader.Pixel.ShallowCopy();
+                _pixelShared = false;
+            }
+
+            if (_pixelScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.SpiShaderUserDataPs0))
+            {
+                Shader.Pixel.UserScalars = Shader.Pixel.UserScalars.Copy();
+                _pixelScalarsShared = false;
+            }
+        }
+
+        if (offset < ComputeRangeStart && last >= ShaderRegisterOffset.SpiShaderPgmLoVs)
+        {
+            if (_vertexShared)
+            {
+                Shader.Vertex = Shader.Vertex.ShallowCopy();
+                _vertexShared = false;
+            }
+
+            var vertex = Shader.Vertex;
+            if (_hullScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.SpiShaderUserDataHs0))
+            {
+                vertex.HullUserScalars = vertex.HullUserScalars.Copy();
+                _hullScalarsShared = false;
+            }
+
+            if (_geometryScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.SpiShaderUserDataGs0))
+            {
+                vertex.GeometryUserScalars = vertex.GeometryUserScalars.Copy();
+                _geometryScalarsShared = false;
+            }
+
+            if (_legacyVertexScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.SpiShaderUserDataVs0))
+            {
+                vertex.LegacyVertexUserScalars = vertex.LegacyVertexUserScalars.Copy();
+                _legacyVertexScalarsShared = false;
+            }
+
+            if (_exportScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.SpiShaderUserDataEs0))
+            {
+                vertex.ExportUserScalars = vertex.ExportUserScalars.Copy();
+                _exportScalarsShared = false;
+            }
+        }
+
+        if (last >= ComputeRangeStart)
+        {
+            if (_computeShared)
+            {
+                Shader.Compute = Shader.Compute.ShallowCopy();
+                _computeShared = false;
+            }
+
+            if (_computeScalarsShared && Overlaps(offset, last, ShaderRegisterOffset.ComputeUserData0))
+            {
+                Shader.Compute.UserScalars = Shader.Compute.UserScalars.Copy();
+                _computeScalarsShared = false;
+            }
+        }
+    }
+
     public void Reset()
     {
-        Context = new ContextRegisters();
+        _contextShared = _vertexShared = _pixelShared = _computeShared = _userConfigShared = false;
+        _pixelScalarsShared = _hullScalarsShared = _geometryScalarsShared = _legacyVertexScalarsShared = _exportScalarsShared = _computeScalarsShared = false;
+        OwnContext(new ContextRegisters());
         Shader = new ShaderProgramRegisters();
         UserConfig = new UserConfigRegisters();
         _savedContext = null;
@@ -61,7 +348,7 @@ public sealed class RegisterBanks
         switch (operation)
         {
             case ContextStateOperation.Clear:
-                Context = new ContextRegisters();
+                OwnContext(new ContextRegisters());
                 CompositeDepthSizeXy = null;
                 break;
             case ContextStateOperation.Push:
@@ -73,14 +360,14 @@ public sealed class RegisterBanks
                     throw Fatal("The context state is not pushed.");
                 }
 
-                Context = _savedContext;
+                OwnContext(_savedContext);
                 CompositeDepthSizeXy = _savedCompositeDepthSizeXy;
                 _savedContext = null;
                 _savedCompositeDepthSizeXy = null;
                 break;
             case ContextStateOperation.PushClear:
                 Push();
-                Context = new ContextRegisters();
+                OwnContext(new ContextRegisters());
                 CompositeDepthSizeXy = null;
                 break;
             default:

src/SharpEmu.Libs/Gpu/GpuCommands/Registers/RegisterWriteTable.cs

@@ -28,6 +28,7 @@ public static class RegisterWriteTable
 
     public static uint WriteContextPacket(RegisterBanks banks, in PacketContext packet, uint offset, ReadOnlySpan<uint> values)
     {
+        banks.PrepareContextWrite();
         var consumed = WritePacket(banks, in packet, offset, values, Context, ContextIndirect, "context", emptyIsHandled: true);
         if (Rendering.RenderTrace.Enabled)
         {
@@ -43,6 +44,7 @@ public static class RegisterWriteTable
     // A shader packet without values has no writer; the other banks accept it.
     public static uint WriteShaderPacket(RegisterBanks banks, in PacketContext packet, uint offset, ReadOnlySpan<uint> values)
     {
+        banks.PrepareShaderWrite(offset, (uint)values.Length);
         var consumed = WritePacket(banks, in packet, offset, values, Shader, ShaderIndirect, "shader", emptyIsHandled: false);
         if (Rendering.RenderTrace.Enabled)
         {
@@ -52,15 +54,19 @@ public static class RegisterWriteTable
         return consumed;
     }
 
-    public static uint WriteUserConfigPacket(RegisterBanks banks, in PacketContext packet, uint offset, ReadOnlySpan<uint> values) =>
-        WritePacket(banks, in packet, offset, values, UserConfig, UserConfigIndirect, "user-config", emptyIsHandled: true);
+    public static uint WriteUserConfigPacket(RegisterBanks banks, in PacketContext packet, uint offset, ReadOnlySpan<uint> values)
+    {
+        banks.PrepareUserConfigWrite();
+        return WritePacket(banks, in packet, offset, values, UserConfig, UserConfigIndirect, "user-config", emptyIsHandled: true);
+    }
 
     public static void WriteContextEntry(RegisterBanks banks, uint offset, uint value, ulong tableAddress)
     {
+        banks.PrepareContextWrite();
         if (offset < ContextIndirect.Length && ContextIndirect[offset] is null)
         {
             // Keep unknown in-bank table values so a later renderer implementation can decode them.
-            banks.Context.UnmodeledTableRegisters[offset] = value;
+            banks.UnmodeledTableRegistersForWrite[offset] = value;
         }
         else
         {
@@ -88,6 +94,7 @@ public static class RegisterWriteTable
 
     public static void WriteShaderEntry(RegisterBanks banks, uint offset, uint value, ulong tableAddress)
     {
+        banks.PrepareShaderWrite(offset, 1);
         WriteEntry(banks, offset, value, ShaderIndirect, "shader", tableAddress);
         if (Rendering.RenderTrace.Enabled)
             TracePixelRegisterWrite(banks, offset, value, tableAddress, "table");
@@ -104,8 +111,11 @@ public static class RegisterWriteTable
             $"value=0x{value:X8} shader=0x{banks.Shader.Pixel.Address:X16}");
     }
 
-    public static void WriteUserConfigEntry(RegisterBanks banks, uint offset, uint value, ulong tableAddress) =>
+    public static void WriteUserConfigEntry(RegisterBanks banks, uint offset, uint value, ulong tableAddress)
+    {
+        banks.PrepareUserConfigWrite();
         WriteEntry(banks, offset, value, UserConfigIndirect, "user-config", tableAddress);
+    }
 
     private static uint WritePacket(
         RegisterBanks banks,

src/SharpEmu.Libs/Gpu/GpuCommands/Registers/RegisterWriters.Context.cs

@@ -98,8 +98,8 @@ internal static partial class RegisterWriters
         {
             indirect[offset] = static (banks, register, value) =>
             {
-                banks.Context.ShaderInterface.PixelInterpolatorSettings[register - SpiPsInputCntl0] = value;
-                banks.Context.ShaderInterface.PixelInterpolatorWritten |= 1u << (int)(register - SpiPsInputCntl0);
+                banks.ShaderInterfaceForWrite.PixelInterpolatorSettings[register - SpiPsInputCntl0] = value;
+                banks.ShaderInterfaceForWrite.PixelInterpolatorWritten |= 1u << (int)(register - SpiPsInputCntl0);
             };
         }
 
@@ -160,7 +160,7 @@ internal static partial class RegisterWriters
 
         for (var offset = PaScAaSampleLocations0; offset < PaScAaSampleLocations0 + SampleLocationRegisters.LocationCount; offset++)
         {
-            indirect[offset] = static (banks, register, value) => banks.Context.SampleLocations.Locations[register - PaScAaSampleLocations0] = value;
+            indirect[offset] = static (banks, register, value) => banks.SampleLocationsForWrite.Locations[register - PaScAaSampleLocations0] = value;
         }
 
         foreach (var offset in new[]
@@ -175,36 +175,36 @@ internal static partial class RegisterWriters
             indirect[offset] = IgnoreEntry;
         }
 
-        indirect[SpiVsOutConfig] = static (banks, _, value) => banks.Context.ShaderInterface.VertexOutputConfiguration = value;
+        indirect[SpiVsOutConfig] = static (banks, _, value) => banks.ShaderInterfaceForWrite.VertexOutputConfiguration = value;
         indirect[DbRenderOverride] = static (banks, _, value) => banks.Context.DepthRenderOverride = DepthRenderOverrideRegisters.Decode(value);
-        indirect[SpiShaderPosFormat] = static (banks, _, value) => banks.Context.ShaderInterface.PositionExportFormat = value;
-        indirect[SpiShaderIdxFormat] = static (banks, _, value) => banks.Context.ShaderInterface.IndexExportFormat = value;
-        indirect[PaClVsOutCntl] = static (banks, _, value) => banks.Context.ShaderInterface.VertexOutputControl = value;
-        indirect[GeNggSubgroupCntl] = static (banks, _, value) => banks.Context.ShaderInterface.PrimitiveShaderSubgroupControl = value;
-        indirect[VgtGsInstanceCnt] = static (banks, _, value) => banks.Context.ShaderInterface.GeometryInstanceCount = value;
-        indirect[VgtGsOnchipCntl] = static (banks, _, value) => banks.Context.ShaderInterface.GeometryOnChipControl = value;
-        indirect[VgtHosMaxTessLevel] = static (banks, _, value) => banks.Context.ShaderInterface.MaxTessellationLevel = value;
-        indirect[VgtHosMinTessLevel] = static (banks, _, value) => banks.Context.ShaderInterface.MinTessellationLevel = value;
-        indirect[GeMaxOutputPerSubgroup] = static (banks, _, value) => banks.Context.ShaderInterface.MaxOutputPerSubgroup = value;
-        indirect[VgtEsgsRingItemSize] = static (banks, _, value) => banks.Context.ShaderInterface.ExportRingItemSize = value;
-        indirect[VgtGsMaxVertOut] = static (banks, _, value) => banks.Context.ShaderInterface.GeometryMaxVerticesOut = value;
-        indirect[VgtPrimitiveIdEn] = static (banks, _, value) => banks.Context.ShaderInterface.PrimitiveIdEnable = value;
-        indirect[VgtReuseOff] = static (banks, _, value) => banks.Context.ShaderInterface.VertexReuseOff = value;
+        indirect[SpiShaderPosFormat] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PositionExportFormat = value;
+        indirect[SpiShaderIdxFormat] = static (banks, _, value) => banks.ShaderInterfaceForWrite.IndexExportFormat = value;
+        indirect[PaClVsOutCntl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.VertexOutputControl = value;
+        indirect[GeNggSubgroupCntl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PrimitiveShaderSubgroupControl = value;
+        indirect[VgtGsInstanceCnt] = static (banks, _, value) => banks.ShaderInterfaceForWrite.GeometryInstanceCount = value;
+        indirect[VgtGsOnchipCntl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.GeometryOnChipControl = value;
+        indirect[VgtHosMaxTessLevel] = static (banks, _, value) => banks.ShaderInterfaceForWrite.MaxTessellationLevel = value;
+        indirect[VgtHosMinTessLevel] = static (banks, _, value) => banks.ShaderInterfaceForWrite.MinTessellationLevel = value;
+        indirect[GeMaxOutputPerSubgroup] = static (banks, _, value) => banks.ShaderInterfaceForWrite.MaxOutputPerSubgroup = value;
+        indirect[VgtEsgsRingItemSize] = static (banks, _, value) => banks.ShaderInterfaceForWrite.ExportRingItemSize = value;
+        indirect[VgtGsMaxVertOut] = static (banks, _, value) => banks.ShaderInterfaceForWrite.GeometryMaxVerticesOut = value;
+        indirect[VgtPrimitiveIdEn] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PrimitiveIdEnable = value;
+        indirect[VgtReuseOff] = static (banks, _, value) => banks.ShaderInterfaceForWrite.VertexReuseOff = value;
         indirect[VgtMultiPrimIbResetIndex] = static (banks, _, value) => banks.Context.PrimitiveResetIndex = value;
-        indirect[VgtTessDistribution] = static (banks, _, value) => banks.Context.ShaderInterface.TessellationDistribution = value;
+        indirect[VgtTessDistribution] = static (banks, _, value) => banks.ShaderInterfaceForWrite.TessellationDistribution = value;
         indirect[VgtShaderStagesEn] = static (banks, _, value) => banks.Context.ShaderStages = value;
-        indirect[VgtLsHsConfig] = static (banks, _, value) => banks.Context.ShaderInterface.LocalHullConfiguration = value;
-        indirect[VgtGsOutPrimType] = static (banks, _, value) => banks.Context.ShaderInterface.GeometryOutputPrimitiveType = value;
-        indirect[VgtTfParam] = static (banks, _, value) => banks.Context.ShaderInterface.TessellationFactorParameter = value;
-        indirect[SpiShaderZFormat] = static (banks, _, value) => banks.Context.ShaderInterface.DepthExportFormat = value;
+        indirect[VgtLsHsConfig] = static (banks, _, value) => banks.ShaderInterfaceForWrite.LocalHullConfiguration = value;
+        indirect[VgtGsOutPrimType] = static (banks, _, value) => banks.ShaderInterfaceForWrite.GeometryOutputPrimitiveType = value;
+        indirect[VgtTfParam] = static (banks, _, value) => banks.ShaderInterfaceForWrite.TessellationFactorParameter = value;
+        indirect[SpiShaderZFormat] = static (banks, _, value) => banks.ShaderInterfaceForWrite.DepthExportFormat = value;
         indirect[SpiShaderColFormat] = ShaderColorFormatEntry;
-        indirect[SpiPsInputEna] = static (banks, _, value) => banks.Context.ShaderInterface.PixelInputEnable = value;
-        indirect[SpiPsInputAddr] = static (banks, _, value) => banks.Context.ShaderInterface.PixelInputAddress = value;
-        indirect[SpiPsInControl] = static (banks, _, value) => banks.Context.ShaderInterface.PixelInputControl = value;
-        indirect[SpiBarycCntl] = static (banks, _, value) => banks.Context.ShaderInterface.BarycentricControl = value;
-        indirect[DbShaderControl] = static (banks, _, value) => banks.Context.ShaderInterface.DepthShaderControl = DepthShaderControlRegisters.Decode(value);
-        indirect[CbShaderMask] = static (banks, _, value) => banks.Context.ShaderInterface.ColorShaderMask = value;
-        indirect[PaScShaderControl] = static (banks, _, value) => banks.Context.ShaderInterface.ScanShaderControl = value;
+        indirect[SpiPsInputEna] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PixelInputEnable = value;
+        indirect[SpiPsInputAddr] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PixelInputAddress = value;
+        indirect[SpiPsInControl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.PixelInputControl = value;
+        indirect[SpiBarycCntl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.BarycentricControl = value;
+        indirect[DbShaderControl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.DepthShaderControl = DepthShaderControlRegisters.Decode(value);
+        indirect[CbShaderMask] = static (banks, _, value) => banks.ShaderInterfaceForWrite.ColorShaderMask = value;
+        indirect[PaScShaderControl] = static (banks, _, value) => banks.ShaderInterfaceForWrite.ScanShaderControl = value;
         indirect[CbTargetMask] = static (banks, _, value) => banks.Context.RenderTargetMask = value;
         indirect[PaScScreenScissorTl] = ScreenScissorTopLeftEntry;
         indirect[PaScScreenScissorBr] = ScreenScissorBottomRightEntry;
@@ -248,7 +248,7 @@ internal static partial class RegisterWriters
         indirect[DbStencilRefMaskBack] = StencilMaskBackEntry;
         indirect[PaClClipCntl] = static (banks, _, value) => banks.Context.Clip = ClipControlRegisters.Decode(value);
         indirect[PaSuLineCntl] = LineControlEntry;
-        indirect[PaClVteCntl] = static (banks, _, value) => banks.Context.ScreenViewport.TransformControl = value;
+        indirect[PaClVteCntl] = static (banks, _, value) => banks.ScreenViewportForWrite.TransformControl = value;
         indirect[PaScModeCntl0] = static (banks, _, value) => banks.Context.ScanMode = ScanModeRegisters.Decode(value);
         indirect[PaScAaConfig] = static (banks, _, value) => banks.Context.AntialiasingConfig = AntialiasingConfigRegisters.Decode(value);
         indirect[PaScCentroidPriority0] = CentroidPriorityEntry;
@@ -369,8 +369,8 @@ internal static partial class RegisterWriters
             throw PacketError(banks, "The pixel interpolator packet has an invalid count", in packet, offset, values.Length);
         }
 
-        values.CopyTo(banks.Context.ShaderInterface.PixelInterpolatorSettings);
-        banks.Context.ShaderInterface.PixelInterpolatorWritten |= (uint)((1ul << values.Length) - 1);
+        values.CopyTo(banks.ShaderInterfaceForWrite.PixelInterpolatorSettings);
+        banks.ShaderInterfaceForWrite.PixelInterpolatorWritten |= (uint)((1ul << values.Length) - 1);
         return (uint)values.Length;
     }
 
@@ -410,7 +410,7 @@ internal static partial class RegisterWriters
     // One location at any slot, or all sixteen from the first slot.
     private static uint SampleLocationsPacket(RegisterBanks banks, in PacketContext packet, uint offset, ReadOnlySpan<uint> values)
     {
-        var locations = banks.Context.SampleLocations.Locations;
+        var locations = banks.SampleLocationsForWrite.Locations;
         if (values.Length == 1)
         {
             locations[offset - PaScAaSampleLocations0] = values[0];
@@ -506,19 +506,19 @@ internal static partial class RegisterWriters
 
     private static void HardwareScreenOffsetEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.HardwareOffsetX = RegisterField.Get(value, 0, 0x1FF);
-        banks.Context.ScreenViewport.HardwareOffsetY = RegisterField.Get(value, 16, 0x1FF);
+        banks.ScreenViewportForWrite.HardwareOffsetX = RegisterField.Get(value, 0, 0x1FF);
+        banks.ScreenViewportForWrite.HardwareOffsetY = RegisterField.Get(value, 16, 0x1FF);
     }
 
     private static void WindowOffsetEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.WindowOffsetX = RegisterField.Coordinate16(value);
-        banks.Context.ScreenViewport.WindowOffsetY = RegisterField.Coordinate16(value >> 16);
+        banks.ScreenViewportForWrite.WindowOffsetX = RegisterField.Coordinate16(value);
+        banks.ScreenViewportForWrite.WindowOffsetY = RegisterField.Coordinate16(value >> 16);
     }
 
     private static void ClipRectangleEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var viewport = banks.Context.ScreenViewport;
+        var viewport = banks.ScreenViewportForWrite;
         if (offset == PaScClipRectRule)
         {
             viewport.ClipRectangleRule = (ushort)(value & 0xFFFFu);
@@ -598,7 +598,7 @@ internal static partial class RegisterWriters
 
     private static void CentroidPriorityEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var locations = banks.Context.SampleLocations;
+        var locations = banks.SampleLocationsForWrite;
         switch (offset)
         {
             case PaScCentroidPriority0:
@@ -612,7 +612,7 @@ internal static partial class RegisterWriters
 
     private static void ShaderColorFormatEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var modes = banks.Context.ShaderInterface.TargetOutputModes;
+        var modes = banks.ShaderInterfaceForWrite.TargetOutputModes;
         for (var index = 0; index < modes.Length; index++)
         {
             modes[index] = (byte)((value >> (index * 4)) & 0xFu);
@@ -621,19 +621,19 @@ internal static partial class RegisterWriters
 
     private static void ScreenScissorTopLeftEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.ScreenScissorLeft = RegisterField.Coordinate16(value);
-        banks.Context.ScreenViewport.ScreenScissorTop = RegisterField.Coordinate16(value >> 16);
+        banks.ScreenViewportForWrite.ScreenScissorLeft = RegisterField.Coordinate16(value);
+        banks.ScreenViewportForWrite.ScreenScissorTop = RegisterField.Coordinate16(value >> 16);
     }
 
     private static void ScreenScissorBottomRightEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.ScreenScissorRight = RegisterField.Coordinate16(value);
-        banks.Context.ScreenViewport.ScreenScissorBottom = RegisterField.Coordinate16(value >> 16);
+        banks.ScreenViewportForWrite.ScreenScissorRight = RegisterField.Coordinate16(value);
+        banks.ScreenViewportForWrite.ScreenScissorBottom = RegisterField.Coordinate16(value >> 16);
     }
 
     private static void GenericScissorTopLeftEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var viewport = banks.Context.ScreenViewport;
+        var viewport = banks.ScreenViewportForWrite;
         viewport.GenericScissorLeft = RegisterField.Coordinate15(value);
         viewport.GenericScissorTop = RegisterField.Coordinate15(value >> 16);
         viewport.GenericScissorWindowOffsetEnable = !RegisterField.Bit(value, 31);
@@ -641,13 +641,13 @@ internal static partial class RegisterWriters
 
     private static void GenericScissorBottomRightEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.GenericScissorRight = RegisterField.Coordinate15(value);
-        banks.Context.ScreenViewport.GenericScissorBottom = RegisterField.Coordinate15(value >> 16);
+        banks.ScreenViewportForWrite.GenericScissorRight = RegisterField.Coordinate15(value);
+        banks.ScreenViewportForWrite.GenericScissorBottom = RegisterField.Coordinate15(value >> 16);
     }
 
     private static void WindowScissorTopLeftEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var viewport = banks.Context.ScreenViewport;
+        var viewport = banks.ScreenViewportForWrite;
         viewport.WindowScissorLeft = RegisterField.Coordinate15(value);
         viewport.WindowScissorTop = RegisterField.Coordinate15(value >> 16);
         viewport.WindowScissorWindowOffsetEnable = !RegisterField.Bit(value, 31);
@@ -655,13 +655,13 @@ internal static partial class RegisterWriters
 
     private static void WindowScissorBottomRightEntry(RegisterBanks banks, uint offset, uint value)
     {
-        banks.Context.ScreenViewport.WindowScissorRight = RegisterField.Coordinate15(value);
-        banks.Context.ScreenViewport.WindowScissorBottom = RegisterField.Coordinate15(value >> 16);
+        banks.ScreenViewportForWrite.WindowScissorRight = RegisterField.Coordinate15(value);
+        banks.ScreenViewportForWrite.WindowScissorBottom = RegisterField.Coordinate15(value >> 16);
     }
 
     private static void ViewportScissorTopLeftEntry(RegisterBanks banks, uint offset, uint value)
     {
-        ref var viewport = ref banks.Context.ScreenViewport.Viewports[(offset - PaScViewportScissor0Tl) / 2];
+        ref var viewport = ref banks.ViewportsForWrite[(offset - PaScViewportScissor0Tl) / 2];
         viewport.ScissorLeft = RegisterField.Coordinate15(value);
         viewport.ScissorTop = RegisterField.Coordinate15(value >> 16);
         viewport.ScissorWindowOffsetEnable = !RegisterField.Bit(value, 31);
@@ -669,21 +669,21 @@ internal static partial class RegisterWriters
 
     private static void ViewportScissorBottomRightEntry(RegisterBanks banks, uint offset, uint value)
     {
-        ref var viewport = ref banks.Context.ScreenViewport.Viewports[(offset - PaScViewportScissor0Br) / 2];
+        ref var viewport = ref banks.ViewportsForWrite[(offset - PaScViewportScissor0Br) / 2];
         viewport.ScissorRight = RegisterField.Coordinate15(value);
         viewport.ScissorBottom = RegisterField.Coordinate15(value >> 16);
     }
 
     private static void ViewportZMinEntry(RegisterBanks banks, uint offset, uint value) =>
-        banks.Context.ScreenViewport.Viewports[(offset - PaScViewportZMin0) / 2].MinDepth = RegisterField.AsFloat(value);
+        banks.ViewportsForWrite[(offset - PaScViewportZMin0) / 2].MinDepth = RegisterField.AsFloat(value);
 
     private static void ViewportZMaxEntry(RegisterBanks banks, uint offset, uint value) =>
-        banks.Context.ScreenViewport.Viewports[(offset - PaScViewportZMax0) / 2].MaxDepth = RegisterField.AsFloat(value);
+        banks.ViewportsForWrite[(offset - PaScViewportZMax0) / 2].MaxDepth = RegisterField.AsFloat(value);
 
     private static void ViewportScaleOffsetEntry(RegisterBanks banks, uint offset, uint value)
     {
         var relative = offset - PaClViewportXScale;
-        ref var viewport = ref banks.Context.ScreenViewport.Viewports[relative / ViewportFieldCount];
+        ref var viewport = ref banks.ViewportsForWrite[relative / ViewportFieldCount];
         var number = RegisterField.AsFloat(value);
         switch (relative % ViewportFieldCount)
         {
@@ -718,7 +718,7 @@ internal static partial class RegisterWriters
 
     private static void GuardBandEntry(RegisterBanks banks, uint offset, uint value)
     {
-        var viewport = banks.Context.ScreenViewport;
+        var viewport = banks.ScreenViewportForWrite;
         var number = RegisterField.AsFloat(value);
         switch (offset)
         {
@@ -760,7 +760,7 @@ internal static partial class RegisterWriters
     }
 
     private static void BlendControlEntry(RegisterBanks banks, uint offset, uint value) =>
-        banks.Context.BlendControls[offset - CbBlend0Control] = BlendRegisters.Decode(value);
+        banks.BlendControlsForWrite[offset - CbBlend0Control] = BlendRegisters.Decode(value);
 
     private static void DepthZInfoEntry(RegisterBanks banks, uint offset, uint value)
     {
@@ -804,7 +804,7 @@ internal static partial class RegisterWriters
     }
 
     private static ref ColorTargetWords ColorTarget(RegisterBanks banks, uint offset, uint first, uint stride) =>
-        ref banks.Context.ColorTargets[(offset - first) / stride];
+        ref banks.ColorTargetsForWrite[(offset - first) / stride];
 
     private static void ColorBaseLowEntry(RegisterBanks banks, uint offset, uint value)
     {
@@ -878,7 +878,7 @@ internal static partial class RegisterWriters
     }
 
     private static void ColorClearWord1Entry(RegisterBanks banks, uint offset, uint value) =>
-        banks.Context.ColorClearWord1[(offset - CbColor0ClearWord1) / ColorSlotStride] = value;
+        banks.ColorClearWord1ForWrite[(offset - CbColor0ClearWord1) / ColorSlotStride] = value;
 
     private static void ColorDccBaseLowEntry(RegisterBanks banks, uint offset, uint value)
     {

src/SharpEmu.Libs/Gpu/GpuCommands/Registers/ShaderProgramRegisters.cs

@@ -199,6 +199,9 @@ public sealed class VertexStageRegisters
     public UserScalarRegisters LegacyVertexUserScalars = new();
     public UserScalarRegisters ExportUserScalars = new();
 
+    // The same registers sharing the user scalar sets, which a writer then copies one by one.
+    public VertexStageRegisters ShallowCopy() => (VertexStageRegisters)MemberwiseClone();
+
     public VertexStageRegisters Copy()
     {
         var copy = (VertexStageRegisters)MemberwiseClone();
@@ -217,6 +220,8 @@ public sealed class PixelStageRegisters
     public PixelResource2 Resource2;
     public UserScalarRegisters UserScalars = new();
 
+    public PixelStageRegisters ShallowCopy() => (PixelStageRegisters)MemberwiseClone();
+
     public PixelStageRegisters Copy()
     {
         var copy = (PixelStageRegisters)MemberwiseClone();
@@ -253,6 +258,8 @@ public sealed class ComputeStageRegisters
     public byte SharedVectorRegisters;
     public UserScalarRegisters UserScalars = new();
 
+    public ComputeStageRegisters ShallowCopy() => (ComputeStageRegisters)MemberwiseClone();
+
     public ComputeStageRegisters Copy()
     {
         var copy = (ComputeStageRegisters)MemberwiseClone();
@@ -264,6 +271,9 @@ public sealed class ComputeStageRegisters
 // The shader bank: the program registers of every stage and their user scalars.
 public sealed class ShaderProgramRegisters
 {
+    // A bank holding the same stage objects, for a snapshot that shares them.
+    public ShaderProgramRegisters ShareStages() => (ShaderProgramRegisters)MemberwiseClone();
+
     public VertexStageRegisters Vertex = new();
     public PixelStageRegisters Pixel = new();
     public ComputeStageRegisters Compute = new();

src/SharpEmu.Libs/Gpu/Images/CachedImage.Transfers.cs

@@ -31,10 +31,13 @@ public sealed unsafe partial class CachedImage
         SubresourceRange = new ImageSubresourceRange(ViewFormatRules.FullAspects(Backing.Format), baseLevel, levelCount, baseLayer, layerCount),
     };
 
+    // What a state that needs no barrier returns, so the common case allocates nothing. Callers only read it.
+    private static readonly List<ImageMemoryBarrier2> NoBarriers = [];
+
     // Barriers for the requested state; a repeated write always gets a barrier.
     public (List<ImageMemoryBarrier2> Barriers, PipelineStageFlags SourceStages) GetBarriers(ImageLayout layout, AccessFlags access, PipelineStageFlags stage, SubresourceRange? range)
     {
-        var barriers = new List<ImageMemoryBarrier2>();
+        List<ImageMemoryBarrier2>? barriers = null;
         PipelineStageFlags sourceStages = 0;
         if ((access & (WriteAccess | AccessFlags.ColorAttachmentWriteBit | AccessFlags.DepthStencilAttachmentWriteBit)) != 0)
         {
@@ -83,7 +86,7 @@ public sealed unsafe partial class CachedImage
                     var repeatedWrite = (state.Access & WriteAccess) != 0;
                     if (state.Layout != layout || state.Access != access || repeatedWrite)
                     {
-                        barriers.Add(MakeBarrier(state, layout, access, level, 1, layer, 1));
+                        (barriers ??= []).Add(MakeBarrier(state, layout, access, level, 1, layer, 1));
                         sourceStages |= state.Stage;
                         states[index] = new ImageAccessState(stage, access, layout);
                     }
@@ -101,15 +104,15 @@ public sealed unsafe partial class CachedImage
             var repeatedWrite = (state.Access & WriteAccess) != 0;
             if (state.Layout == layout && state.Access == access && !repeatedWrite)
             {
-                return (barriers, sourceStages);
+                return (NoBarriers, sourceStages);
             }
 
-            barriers.Add(MakeBarrier(state, layout, access, 0, Vk.RemainingMipLevels, 0, Vk.RemainingArrayLayers));
+            (barriers ??= []).Add(MakeBarrier(state, layout, access, 0, Vk.RemainingMipLevels, 0, Vk.RemainingArrayLayers));
             sourceStages |= state.Stage;
         }
 
         Backing.State = new ImageAccessState(stage, access, layout);
-        return (barriers, sourceStages);
+        return (barriers ?? NoBarriers, sourceStages);
     }
 
     public void Transition(ImageLayout layout, AccessFlags access, SubresourceRange? range, CommandBuffer command)

src/SharpEmu.Libs/Gpu/Images/ColorToMultisampleDepthBlit.cs

@@ -303,6 +303,7 @@ public sealed unsafe class ColorToMultisampleDepthBlit : IDisposable
         vk.CmdBeginRenderPass(command, &begin, SubpassContents.Inline);
         vk.CmdBindDescriptorSets(command, PipelineBindPoint.Graphics, _pipelineLayout, 0, 1, &set, 0, null);
         vk.CmdBindPipeline(command, PipelineBindPoint.Graphics, cached.Pipeline);
+        Vulkan.GraphicsDynamicStateEpoch.Advance();
         var viewport = new Viewport(0.0f, 0.0f, extent.Width, extent.Height, 0.0f, 1.0f);
         var scissor = new Rect2D(new Offset2D(0, 0), extent);
         vk.CmdSetViewport(command, 0, 1, &viewport);

src/SharpEmu.Libs/Gpu/Images/GuestImageCache.Lookup.cs

@@ -35,6 +35,25 @@ public sealed partial class GuestImageCache
 
     private void InvalidateLookups() => _lookupGeneration++;
 
+    // Changes whenever an image a lookup could find is created, removed or replaced, so a caller
+    // that remembers a lookup's result can tell when it may have changed.
+    public ulong LookupGeneration => _lookupGeneration;
+
+    // What a remembered lookup still does on reuse: the image counts as used now. False when the
+    // image is gone.
+    public bool TryTouchRemembered(ResourceSlotIdentifier imageIdentifier)
+    {
+        using var held = _lock.Hold();
+        if (_slots.TryGet(imageIdentifier) is not { Registered: true } image)
+        {
+            return false;
+        }
+
+        image.LastAccessTick = _scheduler.CurrentTick;
+        TouchImage(image);
+        return true;
+    }
+
     private static int LookupSlot(in ImageRequest request, bool exactFormat)
     {
         var hash = HashCode.Combine(request.Description.Data.Address, request.Description.Data.Size, request.Description.PixelFormat,

src/SharpEmu.Libs/Gpu/Images/GuestImageCache.Metadata.cs

@@ -287,13 +287,20 @@ public sealed partial class GuestImageCache
 
     public bool IsMetadataCleared(ulong address, uint slice) => IsMetadataCleared(address, slice, out _);
 
+    // DCC metadata of a bound target, or the range of a DCC fill seen before its target: both are
+    // filled on the host, which keeps them CPU-readable instead of GPU-written.
     public bool OverlapsDccMetadata(ulong address, ulong size)
     {
         using var held = _lock.Hold();
         foreach (var (start, metadata) in _surfaceMetadata)
         {
-            if (metadata.Kind == SurfaceMetadataKind.Dcc && metadata.Size != 0 &&
-                address < start + metadata.Size && start < address + size)
+            var extent = metadata.Kind switch
+            {
+                SurfaceMetadataKind.Dcc => metadata.Size,
+                SurfaceMetadataKind.PendingDcc => metadata.FillSize,
+                _ => 0UL,
+            };
+            if (extent != 0 && address < start + extent && start < address + size)
             {
                 return true;
             }

src/SharpEmu.Libs/Gpu/Images/GuestImageCache.Transfers.cs

@@ -761,6 +761,7 @@ public sealed unsafe partial class GuestImageCache
         _scheduler.QueuePriorityCompletionAction(() =>
         {
             download.Invalidate(offset, range.Size);
+            GpuWritebackEpoch.Advance();
             if (!backing.TryWriteBacking(range.Address, download.Mapped.Slice((int)offset, (int)range.Size)))
             {
                 throw SubmissionScheduler.Fatal($"The image readback could not be written to guest memory: address=0x{range.Address:X16} size=0x{range.Size:X}.");

src/SharpEmu.Libs/Gpu/Images/ImageRequestBuilders.ColorTarget.cs

@@ -76,8 +76,56 @@ public static partial class ImageRequestBuilders
 
     public static bool SupportsDccFixedClear(Format format) => PackedClearValue.SupportsDccFixedColor(format);
 
+    // Every draw rebuilds its targets' requests (a whole mip layout) from the same register words,
+    // twice: once for the attachments and once for the pixel outputs. The build reads nothing but
+    // its arguments, so a result is reused while they are equal.
+    private struct ColorTargetMemo
+    {
+        public bool Valid;
+        public ColorTargetWords Words;
+        public uint TargetMask;
+        public uint DrawLayerOffset;
+        public bool IgnoreTargetMask;
+        public ColorTargetResolution? Result;
+    }
+
+    private const int TargetMemoCount = 8;
+
+    [ThreadStatic]
+    private static ColorTargetMemo[]? _colorTargetMemos;
+
+    private static int TargetMemoSlot(ulong address) => (int)((address >> 12) ^ (address >> 20)) & (TargetMemoCount - 1);
+
     // Builds the request for a bound color target. Null when the slot carries no target.
     public static ColorTargetResolution? ColorTarget(in ColorTargetWords words, uint targetMask, uint drawLayerOffset, bool ignoreTargetMask)
+    {
+        if (Rendering.RenderTrace.Enabled)
+        {
+            return ColorTargetCore(in words, targetMask, drawLayerOffset, ignoreTargetMask);
+        }
+
+        var memos = _colorTargetMemos ??= new ColorTargetMemo[TargetMemoCount];
+        ref var memo = ref memos[TargetMemoSlot(words.BaseAddress)];
+        if (memo.Valid && memo.TargetMask == targetMask && memo.DrawLayerOffset == drawLayerOffset &&
+            memo.IgnoreTargetMask == ignoreTargetMask && memo.Words.Equals(words))
+        {
+            return memo.Result;
+        }
+
+        var result = ColorTargetCore(in words, targetMask, drawLayerOffset, ignoreTargetMask);
+        memo = new ColorTargetMemo
+        {
+            Valid = true,
+            Words = words,
+            TargetMask = targetMask,
+            DrawLayerOffset = drawLayerOffset,
+            IgnoreTargetMask = ignoreTargetMask,
+            Result = result,
+        };
+        return result;
+    }
+
+    private static ColorTargetResolution? ColorTargetCore(in ColorTargetWords words, uint targetMask, uint drawLayerOffset, bool ignoreTargetMask)
     {
         var mask = targetMask & 0xF;
         if (ignoreTargetMask && words.BaseAddress != 0 && mask == 0)
@@ -235,8 +283,7 @@ public static partial class ImageRequestBuilders
             pitch = TileGeometry.TexturePitch(transferFormat, width, GuestTileMode.Linear);
         }
 
-        var mipSpans = new TileLevelSpan[TiledSurfaceLayout.MaxLevels];
-        var mipPadded = new TilePaddedSize[TiledSurfaceLayout.MaxLevels];
+        var (mipSpans, mipPadded) = MipLayoutScratch.Rent();
         TiledSurfaceLayout? volumeLayout = null;
         ulong size;
         ulong backingSize = 0;

src/SharpEmu.Libs/Gpu/Images/ImageRequestBuilders.DepthTarget.cs

@@ -59,8 +59,28 @@ public static partial class ImageRequestBuilders
     }
 
     // Builds the request for the bound depth target. Null when no depth or stencil state is active.
-    public static DepthTargetResolution? DepthTarget(in DepthTargetWords depthWords, IImageFormatSupport device) =>
-        DepthTargetCore(in depthWords, device, copyMode: false, writeBuffer: false);
+    // Reused while the words and the device are the same, as ColorTarget's requests are.
+    public static DepthTargetResolution? DepthTarget(in DepthTargetWords depthWords, IImageFormatSupport device)
+    {
+        if (Rendering.RenderTrace.Enabled)
+        {
+            return DepthTargetCore(in depthWords, device, copyMode: false, writeBuffer: false);
+        }
+
+        if (_depthTargetMemo is { } memo && ReferenceEquals(memo.Device, device) && memo.Words.Equals(depthWords))
+        {
+            return memo.Result;
+        }
+
+        var result = DepthTargetCore(in depthWords, device, copyMode: false, writeBuffer: false);
+        _depthTargetMemo = new DepthTargetMemo(depthWords, device, result);
+        return result;
+    }
+
+    private sealed record DepthTargetMemo(DepthTargetWords Words, IImageFormatSupport Device, DepthTargetResolution? Result);
+
+    [ThreadStatic]
+    private static DepthTargetMemo? _depthTargetMemo;
 
     // Builds the source or destination image used by DB_RENDER_OVERRIDE depth/stencil copies.
     public static DepthTargetResolution? DepthTargetCopy(in DepthTargetWords depthWords, IImageFormatSupport device, bool writeBuffer) =>

src/SharpEmu.Libs/Gpu/Images/ImageRequestBuilders.Texture.cs

@@ -89,8 +89,7 @@ public static partial class ImageRequestBuilders
             return;
         }
 
-        var spans = new TileLevelSpan[TiledSurfaceLayout.MaxLevels];
-        var padded = new TilePaddedSize[TiledSurfaceLayout.MaxLevels];
+        var (spans, padded) = MipLayoutScratch.Rent();
         TileGeometry.TryGetTextureSize(
             description.GuestFormat, description.Extent.Width, description.Extent.Height, description.Resources.Levels, description.TileMode,
             out _, spans, padded);

src/SharpEmu.Libs/Gpu/Images/SamplerStore.cs

@@ -48,7 +48,7 @@ public enum SamplerBorderColor : uint
 public sealed unsafe class SamplerStore : IDisposable
 {
     private readonly GpuDeviceInfo _device;
-    private readonly Dictionary<(uint, uint, uint, uint, bool), Sampler> _samplers = new();
+    private readonly System.Collections.Concurrent.ConcurrentDictionary<(uint, uint, uint, uint, bool), Sampler> _samplers = new();
     private readonly object _gate = new();
 
     public SamplerStore(GpuDeviceInfo device) => _device = device;
@@ -57,16 +57,22 @@ public sealed unsafe class SamplerStore : IDisposable
 
     public Sampler GetSampler(in SamplerDescriptorWords words, bool integerView)
     {
+        var key = (words[0], words[1], words[2], words[3], integerView);
+        // Every sampler binding of every draw asks; only the first of each kind creates.
+        if (_samplers.TryGetValue(key, out var found))
+        {
+            return found;
+        }
+
         lock (_gate)
         {
-            var key = (words[0], words[1], words[2], words[3], integerView);
             if (_samplers.TryGetValue(key, out var existing))
             {
                 return existing;
             }
 
             var sampler = CreateSampler(words, integerView);
-            _samplers.Add(key, sampler);
+            _samplers[key] = sampler;
             return sampler;
         }
     }

src/SharpEmu.Libs/Gpu/Images/TextureTransferLayout.cs

@@ -233,8 +233,7 @@ public sealed class TextureTransferLayout
         layout.Pitch = TileGeometry.TexturePitch(format, width, tile);
         if (tile == GuestTileMode.Linear)
         {
-            var levelSpans = new TileLevelSpan[TiledSurfaceLayout.MaxLevels];
-            var paddedSizes = new TilePaddedSize[TiledSurfaceLayout.MaxLevels];
+            var (levelSpans, paddedSizes) = MipLayoutScratch.Rent();
             TileGeometry.TryGetTextureSize(format, width, height, levels, tile, out _, levelSpans, paddedSizes);
             for (uint level = 0; level < levels; level++)
             {

src/SharpEmu.Libs/Gpu/Images/TileGeometry.cs

@@ -21,6 +21,24 @@ public struct TileLevelSpan
 
 public readonly record struct TilePaddedSize(uint Width, uint Height);
 
+// The per-level output arrays of the layout queries, kept per thread: every draw that resolves a
+// target or a texture needs them for a moment and reads them before the next query.
+internal static class MipLayoutScratch
+{
+    [ThreadStatic] private static TileLevelSpan[]? _spans;
+    [ThreadStatic] private static TilePaddedSize[]? _padded;
+
+    // Zeroed, as a new array would be, so a level the query leaves out reads as empty.
+    public static (TileLevelSpan[] Spans, TilePaddedSize[] Padded) Rent()
+    {
+        var spans = _spans ??= new TileLevelSpan[TiledSurfaceLayout.MaxLevels];
+        var padded = _padded ??= new TilePaddedSize[TiledSurfaceLayout.MaxLevels];
+        Array.Clear(spans);
+        Array.Clear(padded);
+        return (spans, padded);
+    }
+}
+
 public enum TileBlockKind : uint
 {
     Standard256B,

src/SharpEmu.Libs/Gpu/Pipelines/IShaderPipelineHost.cs

@@ -36,6 +36,9 @@ internal interface IShaderPipelineHost
 
     bool ComputeWave64Supported { get; }
 
+    // True when a 64-invocation wave64 compute workgroup runs as one 64-lane host subgroup.
+    bool ComputeWave64SubgroupNative => false;
+
     bool GraphicsSubgroupOperationsEnabled { get; }
 
     // The device supports shaderSharedInt64Atomics, so LDS 64-bit atomics can be

src/SharpEmu.Libs/Gpu/Pipelines/PipelineKeys.cs

@@ -45,6 +45,13 @@ public sealed class PipelineStaticParameters : IEquatable<PipelineStaticParamete
 
     public PipelineStaticParameters()
     {
+        Reset();
+    }
+
+    // The state of a new instance, for a lookup object reused across draws.
+    public void Reset()
+    {
+        Array.Clear(_bytes);
         DepthClipEnable = true;
         Samples = 1;
         ColorCount = 1;
@@ -146,6 +153,15 @@ public sealed class PipelineRenderingState : IEquatable<PipelineRenderingState>
     public Format StencilFormat { get; set; } = Format.Undefined;
     public uint ColorCount { get; set; }
 
+    // The state of a new instance, for a lookup object reused across draws.
+    public void Reset()
+    {
+        Array.Clear(ColorFormats);
+        DepthFormat = Format.Undefined;
+        StencilFormat = Format.Undefined;
+        ColorCount = 0;
+    }
+
     public bool Equals(PipelineRenderingState? other) =>
         other is not null && ColorCount == other.ColorCount && DepthFormat == other.DepthFormat && StencilFormat == other.StencilFormat &&
         ColorFormats.AsSpan().SequenceEqual(other.ColorFormats);
@@ -179,6 +195,15 @@ public sealed class PipelineVertexInputState : IEquatable<PipelineVertexInputSta
     public byte BindingCount { get; set; }
     public byte AttributeCount { get; set; }
 
+    // The state of a new instance, for a lookup object reused across draws.
+    public void Reset()
+    {
+        Array.Clear(Bindings);
+        Array.Clear(Attributes);
+        BindingCount = 0;
+        AttributeCount = 0;
+    }
+
     public bool Equals(PipelineVertexInputState? other) =>
         other is not null && BindingCount == other.BindingCount && AttributeCount == other.AttributeCount &&
         Bindings.AsSpan().SequenceEqual(other.Bindings) && Attributes.AsSpan().SequenceEqual(other.Attributes);
@@ -207,8 +232,9 @@ public sealed class PipelineVertexInputState : IEquatable<PipelineVertexInputSta
 public sealed class GraphicsPipelineKey : IEquatable<GraphicsPipelineKey>
 {
     public required PipelineRenderingState Rendering { get; init; }
-    public ulong VertexProgramId { get; init; }
-    public ulong PixelProgramId { get; init; }
+    // Settable for the lookup key reused across draws; a key in the cache is never changed.
+    public ulong VertexProgramId { get; set; }
+    public ulong PixelProgramId { get; set; }
     public required PipelineVertexInputState VertexInput { get; init; }
     public required PipelineStaticParameters StaticParameters { get; init; }
 

src/SharpEmu.Libs/Gpu/Pipelines/ShaderPipelineCache.cs

@@ -27,6 +27,9 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
     private readonly Dictionary<GraphicsPipelineKey, PipelineHandle> _graphicsPipelines = new();
     private readonly Dictionary<ComputePipelineKey, PipelineHandle> _computePipelines = new();
     private readonly object _gate = new();
+    // The pipeline maps: a lookup takes this lock or _gate, an addition both; a cached draw so does
+    // not wait for the program preparation a resolving thread holds _gate for.
+    private readonly object _pipelineGate = new();
     private readonly bool _strictShaders = Environment.GetEnvironmentVariable("SHARPEMU_STRICT_COMPUTE") != "0";
     private readonly HashSet<(ShaderStage Stage, ulong Hash, uint CodeSize)> _reportedShaderSkips = [];
 
@@ -70,6 +73,11 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         return new ShaderSource(registered, hash, userData, userDataBase, stage);
     }
 
+    // Program resolution runs on the render thread and on a thread resolving draws ahead of it;
+    // the program, decode and materialization caches it updates are used under this lock only.
+    // A pipeline lookup takes _pipelineGate instead, so it does not wait for a resolution.
+    private readonly object _resolveGate = new();
+
     public GraphicsPrograms GetGraphicsPrograms(
         VertexStageRegisters vertex,
         PixelStageRegisters pixel,
@@ -78,6 +86,21 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         ReadOnlySpan<ColorComponentMap> targetExportMapping,
         bool pixelActive,
         bool depthBound)
+    {
+        lock (_resolveGate)
+        {
+            return GetGraphicsProgramsCore(vertex, pixel, shaderInterface, context, targetExportMapping, pixelActive, depthBound);
+        }
+    }
+
+    private GraphicsPrograms GetGraphicsProgramsCore(
+        VertexStageRegisters vertex,
+        PixelStageRegisters pixel,
+        ShaderInterfaceRegisters shaderInterface,
+        ContextRegisters context,
+        ReadOnlySpan<ColorComponentMap> targetExportMapping,
+        bool pixelActive,
+        bool depthBound)
     {
         using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.ProgramPreparation);
         var vertexSource = PrepareSource(
@@ -209,44 +232,53 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
             shaderInterface.VertexOutputControl, clipSpace);
     }
 
-    // One past the highest parameter location the pixel program reads, resolved as its translator does.
-    private static uint ReadVertexOutputCount(Gen5ShaderProgram pixelProgram, PixelInputInfo info)
-    {
-        var attributes = pixelProgram.Instructions
+    // The interpolated attributes a pixel program reads, ascending. A program never changes, and this
+    // was rebuilt with LINQ on every draw.
+    private static readonly System.Runtime.CompilerServices.ConditionalWeakTable<Gen5ShaderProgram, uint[]> InterpolatedAttributeCache = new();
+
+    private static uint[] InterpolatedAttributesOf(Gen5ShaderProgram program) =>
+        InterpolatedAttributeCache.GetValue(program, static program => program.Instructions
             .Select(static instruction => instruction.Control)
             .OfType<Gen5InterpolationControl>()
             .Select(static control => control.Attribute)
             .Distinct()
             .Order()
-            .ToArray();
+            .ToArray());
+
+    // One past the highest parameter location the pixel program reads, resolved as its translator does.
+    private static uint ReadVertexOutputCount(Gen5ShaderProgram pixelProgram, PixelInputInfo info)
+    {
+        var attributes = InterpolatedAttributesOf(pixelProgram);
         if (attributes.Length == 0)
         {
             return 0;
         }
 
-        var controls = new uint[32];
+        Span<uint> controls = stackalloc uint[32];
         for (var index = 0u; index < (uint)controls.Length; index++)
         {
-            controls[index] = index < info.InputCount && index < (uint)info.InterpolatorSettings.Length
+            controls[(int)index] = index < info.InputCount && index < (uint)info.InterpolatorSettings.Length
                 ? info.InterpolatorSettings[index]
                 : index;
         }
 
-        return Gen5PixelInputMapping.ResolveLocations(controls, attributes).Max() + 1;
-    }
-
-    private static uint InterpolatedAttributeCount(Gen5ShaderProgram program)
-    {
-        var maxAttribute = -1;
-        foreach (var instruction in program.Instructions)
+        // Attributes are distinct and below 32, so a program never has more than 32 of them.
+        Span<uint> locations = attributes.Length <= 32 ? stackalloc uint[attributes.Length] : new uint[attributes.Length];
+        Gen5PixelInputMapping.ResolveLocations(controls, attributes, locations);
+        var highest = 0u;
+        foreach (var location in locations)
         {
-            if (instruction.Control is Gen5InterpolationControl interpolation)
-            {
-                maxAttribute = Math.Max(maxAttribute, (int)interpolation.Attribute);
-            }
+            highest = Math.Max(highest, location);
         }
 
-        return (uint)(maxAttribute + 1);
+        return highest + 1;
+    }
+
+    // The attribute list is ascending, so its last entry is the highest.
+    private static uint InterpolatedAttributeCount(Gen5ShaderProgram program)
+    {
+        var attributes = InterpolatedAttributesOf(program);
+        return attributes.Length == 0 ? 0 : attributes[^1] + 1;
     }
 
     // The bound colour slots in order; each output mode names the kind the pixel program exports.
@@ -321,6 +353,20 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         uint dimensionX,
         uint dimensionY,
         uint dimensionZ)
+    {
+        lock (_resolveGate)
+        {
+            return GetComputeProgramCore(compute, shaderInterface, dispatchInitiator, dimensionX, dimensionY, dimensionZ);
+        }
+    }
+
+    private ComputeProgram GetComputeProgramCore(
+        ComputeStageRegisters compute,
+        ShaderInterfaceRegisters shaderInterface,
+        uint dispatchInitiator,
+        uint dimensionX,
+        uint dimensionY,
+        uint dimensionZ)
     {
         using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.ProgramPreparation);
         var source = PrepareSource(compute.Address, ShaderStage.Compute, "compute", compute.UserScalars, compute.UserScalarCount, probeWrittenRegisters: false, userDataBase: 0);
@@ -417,6 +463,28 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         ShaderProgram pixelProgram)
     {
         using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.PipelineCreation);
+        // The lookup fills objects this thread reuses for every draw; only a new pipeline gets its own.
+        var lookup = _graphicsLookup ??= new GraphicsPipelineKey
+        {
+            Rendering = new PipelineRenderingState(),
+            VertexInput = new PipelineVertexInputState(),
+            StaticParameters = new PipelineStaticParameters(),
+        };
+        lookup.Rendering.Reset();
+        lookup.VertexInput.Reset();
+        lookup.StaticParameters.Reset();
+        FillGraphicsState(colors, in depth, vertexInput, pixelInput, context, in rendering, topology, primitiveRestartEnabled, disableBlending,
+            vertexProgram, pixelProgram, _host.NoAttachmentSampleCounts, lookup.Rendering, lookup.StaticParameters, lookup.VertexInput);
+        lookup.VertexProgramId = vertexProgram.Id;
+        lookup.PixelProgramId = pixelInput is not null ? pixelProgram.Id : 0;
+        lock (_pipelineGate)
+        {
+            if (_graphicsPipelines.TryGetValue(lookup, out var cached))
+            {
+                return cached;
+            }
+        }
+
         var description = BuildGraphicsDescription(
             colors, in depth, vertexInput, pixelInput, context, in rendering, topology, primitiveRestartEnabled, disableBlending,
             vertexProgram, pixelProgram, _host.NoAttachmentSampleCounts);
@@ -437,12 +505,19 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
             }
 
             var created = _host.CreateGraphicsPipeline(description);
-            _graphicsPipelines.Add(key, created);
+            lock (_pipelineGate)
+            {
+                _graphicsPipelines.Add(key, created);
+            }
+
             ShaderCacheCounters.CountGraphicsPipeline();
             return created;
         }
     }
 
+    [ThreadStatic]
+    private static GraphicsPipelineKey? _graphicsLookup;
+
     // The full static state of one graphics pipeline from the draw's targets, registers and programs.
     internal static GraphicsPipelineDescription BuildGraphicsDescription(
         ReadOnlySpan<ColorTargetState> colors,
@@ -457,6 +532,43 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         ShaderProgram vertexProgram,
         ShaderProgram pixelProgram,
         SampleCountFlags noAttachmentSampleCounts)
+    {
+        var renderingState = new PipelineRenderingState();
+        var parameters = new PipelineStaticParameters();
+        var vertexState = new PipelineVertexInputState();
+        var (vertexStage, pixelStage) = FillGraphicsState(colors, in depth, vertexInput, pixelInput, context, in rendering, topology,
+            primitiveRestartEnabled, disableBlending, vertexProgram, pixelProgram, noAttachmentSampleCounts, renderingState, parameters, vertexState);
+        return new GraphicsPipelineDescription
+        {
+            Rendering = renderingState,
+            VertexInput = vertexState,
+            VertexInfo = vertexInput,
+            VertexProgram = vertexProgram,
+            VertexStage = vertexStage,
+            PixelInfo = pixelInput,
+            PixelProgram = pixelProgram,
+            PixelStage = pixelStage,
+            StaticParameters = parameters,
+        };
+    }
+
+    // Writes the pipeline state into objects in their initial state; returns the stage programs.
+    private static (ShaderProgramInfo VertexStage, ShaderProgramInfo? PixelStage) FillGraphicsState(
+        ReadOnlySpan<ColorTargetState> colors,
+        in DepthAttachmentState depth,
+        VertexInputInfo vertexInput,
+        PixelInputInfo? pixelInput,
+        ContextRegisters context,
+        in RenderingState rendering,
+        PrimitiveTopology topology,
+        bool primitiveRestartEnabled,
+        bool disableBlending,
+        ShaderProgram vertexProgram,
+        ShaderProgram pixelProgram,
+        SampleCountFlags noAttachmentSampleCounts,
+        PipelineRenderingState renderingState,
+        PipelineStaticParameters parameters,
+        PipelineVertexInputState vertexState)
     {
         if (colors.Length > PipelineStaticParameters.ColorAttachmentCount)
         {
@@ -477,8 +589,8 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         var vertexStage = vertexInput.Stage.Program ?? throw SubmissionScheduler.Fatal("The vertex stage has no program.");
         var pixelStage = pixelActive ? pixelInput!.Stage.Program ?? throw SubmissionScheduler.Fatal("The pixel stage has no program.") : null;
         var colorCount = (uint)colors.Length;
-        var parameters = new PipelineStaticParameters { ColorCount = colorCount };
-        var renderingState = new PipelineRenderingState { ColorCount = colorCount };
+        parameters.ColorCount = colorCount;
+        renderingState.ColorCount = colorCount;
         for (var index = 0; index < colors.Length; index++)
         {
             ref readonly var color = ref colors[index];
@@ -564,18 +676,8 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
             parameters.SetBlendBypass(index, ((context.ColorTargets[slot].Info >> 16) & 0x1) != 0);
         }
 
-        return new GraphicsPipelineDescription
-        {
-            Rendering = renderingState,
-            VertexInput = BuildVertexInputState(vertexInput),
-            VertexInfo = vertexInput,
-            VertexProgram = vertexProgram,
-            VertexStage = vertexStage,
-            PixelInfo = pixelInput,
-            PixelProgram = pixelProgram,
-            PixelStage = pixelStage,
-            StaticParameters = parameters,
-        };
+        FillVertexInputState(vertexInput, vertexState);
+        return (vertexStage, pixelStage);
     }
 
     internal static GraphicsPipelineKey KeyOf(GraphicsPipelineDescription description) => new()
@@ -587,18 +689,15 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
         StaticParameters = description.StaticParameters,
     };
 
-    private static PipelineVertexInputState BuildVertexInputState(VertexInputInfo info)
+    private static void FillVertexInputState(VertexInputInfo info, PipelineVertexInputState state)
     {
         if (info.Buffers.Length > VertexInputInfo.MaxBuffers || info.Attributes.Length > VertexInputInfo.MaxBuffers)
         {
             throw SubmissionScheduler.Fatal($"The vertex input is too large: buffers={info.Buffers.Length} attributes={info.Attributes.Length}.");
         }
 
-        var state = new PipelineVertexInputState
-        {
-            BindingCount = (byte)info.Buffers.Length,
-            AttributeCount = (byte)info.Attributes.Length,
-        };
+        state.BindingCount = (byte)info.Buffers.Length;
+        state.AttributeCount = (byte)info.Attributes.Length;
         for (var binding = 0; binding < info.Buffers.Length; binding++)
         {
             var buffer = info.Buffers[binding];
@@ -615,8 +714,6 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
 
             state.Attributes[index] = new PipelineVertexAttribute(attribute.OffsetBytes, (byte)attribute.BufferIndex);
         }
-
-        return state;
     }
 
     public PipelineHandle CreateComputePipeline(ComputeInputInfo input, ShaderProgram program)
@@ -629,6 +726,14 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
 
         var stage = input.Stage.Program ?? throw SubmissionScheduler.Fatal("The compute stage has no program.");
         var key = new ComputePipelineKey(program.Id);
+        lock (_pipelineGate)
+        {
+            if (_computePipelines.TryGetValue(key, out var cached))
+            {
+                return cached;
+            }
+        }
+
         lock (_gate)
         {
             if (_computePipelines.TryGetValue(key, out var cached))
@@ -642,7 +747,11 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
             }
 
             var created = _host.CreateComputePipeline(new ComputePipelineDescription { Input = input, Program = program, Stage = stage });
-            _computePipelines.Add(key, created);
+            lock (_pipelineGate)
+            {
+                _computePipelines.Add(key, created);
+            }
+
             ShaderCacheCounters.CountComputePipeline();
             return created;
         }
@@ -659,6 +768,15 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
 
         var stage = input.Stage.Program ?? throw SubmissionScheduler.Fatal("The compute stage has no program.");
         var key = new ComputePipelineKey(program.Id);
+        lock (_pipelineGate)
+        {
+            if (_computePipelines.TryGetValue(key, out var cached))
+            {
+                handle = cached;
+                return true;
+            }
+        }
+
         lock (_gate)
         {
             if (_computePipelines.TryGetValue(key, out var cached))
@@ -674,7 +792,11 @@ internal sealed partial class ShaderPipelineCache : IShaderPipelineProvider
                 return false;
             }
 
-            _computePipelines.Add(key, created);
+            lock (_pipelineGate)
+            {
+                _computePipelines.Add(key, created);
+            }
+
             ShaderCacheCounters.CountComputePipeline();
             handle = created;
             return true;

src/SharpEmu.Libs/Gpu/Pipelines/ShaderProgramCache.cs

@@ -1,6 +1,7 @@
 // Copyright (C) 2026 SharpEmu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+using System.Runtime.InteropServices;
 using SharpEmu.HLE;
 using SharpEmu.Libs.Gpu.Rendering;
 using SharpEmu.Libs.Gpu.Scheduling;
@@ -68,6 +69,37 @@ public sealed class ProgramKey : IEquatable<ProgramKey>
     public override int GetHashCode() => HashCode.Combine(Stage, Hash, UserDataCount, CodeSize, StaticState.Length);
 }
 
+// A program key as a lookup sees it: the static state words stay in the caller's list, so a draw
+// that finds its program does not build a key object or copy the words.
+internal readonly ref struct ProgramKeyProbe(ShaderStage stage, ulong hash, uint userDataCount, uint codeSize, ReadOnlySpan<uint> staticState)
+{
+    public readonly ShaderStage Stage = stage;
+    public readonly ulong Hash = hash;
+    public readonly uint UserDataCount = userDataCount;
+    public readonly uint CodeSize = codeSize;
+    public readonly ReadOnlySpan<uint> StaticState = staticState;
+}
+
+internal sealed class ProgramKeyComparer : IEqualityComparer<ProgramKey>, IAlternateEqualityComparer<ProgramKeyProbe, ProgramKey>
+{
+    public static readonly ProgramKeyComparer Instance = new();
+
+    public bool Equals(ProgramKey? x, ProgramKey? y) => x is null ? y is null : x.Equals(y);
+
+    public int GetHashCode(ProgramKey key) => key.GetHashCode();
+
+    public bool Equals(ProgramKeyProbe alternate, ProgramKey key) =>
+        alternate.Stage == key.Stage && alternate.Hash == key.Hash && alternate.UserDataCount == key.UserDataCount &&
+        alternate.CodeSize == key.CodeSize && alternate.StaticState.SequenceEqual(key.StaticState);
+
+    // The same value as ProgramKey.GetHashCode.
+    public int GetHashCode(ProgramKeyProbe alternate) =>
+        HashCode.Combine(alternate.Stage, alternate.Hash, alternate.UserDataCount, alternate.CodeSize, alternate.StaticState.Length);
+
+    public ProgramKey Create(ProgramKeyProbe alternate) =>
+        new(alternate.Stage, alternate.Hash, alternate.UserDataCount, alternate.CodeSize, alternate.StaticState.ToArray());
+}
+
 // One compiled module of a program entry for one specialization and push-data start.
 internal sealed class ProgramPermutation
 {
@@ -102,7 +134,7 @@ internal sealed class ShaderProgramCache
     private readonly CpuContext _context;
     private readonly IGuestGpuBackend _compiler;
     private readonly IShaderPipelineHost _host;
-    private readonly Dictionary<ProgramKey, ProgramSourceEntry> _programs = new();
+    private readonly Dictionary<ProgramKey, ProgramSourceEntry> _programs = new(ProgramKeyComparer.Instance);
     private readonly Dictionary<(ulong Hash, uint CodeSize), Gen5ShaderProgram> _decoded = new();
     private readonly Dictionary<(ulong Hash, uint CodeSize), ShaderCodeCapture> _codeCaptures = new();
     private readonly List<uint> _staticState = new(StageStaticKey.MaxWords);
@@ -216,13 +248,14 @@ internal sealed class ShaderProgramCache
 
     private ShaderProgram GetOrCompileCore(ShaderSource source, StageCompileOptions options, ref uint pushDataCursor, out ShaderStageResources stage)
     {
-        ProgramKey key;
+        ProgramKey? key = null;
         ProgramSourceEntry? entry;
         using (RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.ProgramCacheLookup))
         {
             BuildStaticState(source.Stage, options);
-            key = new ProgramKey(source.Stage, source.Hash, (uint)source.UserData.Length, source.CodeSize, _staticState.ToArray());
-            _programs.TryGetValue(key, out entry);
+            var probe = new ProgramKeyProbe(source.Stage, source.Hash, (uint)source.UserData.Length, source.CodeSize,
+                CollectionsMarshal.AsSpan(_staticState));
+            _programs.GetAlternateLookup<ProgramKeyProbe>().TryGetValue(probe, out entry);
         }
         var sourceWasCached = entry is not null;
         if (RenderTrace.Enabled && RenderTrace.Pipeline())
@@ -246,12 +279,14 @@ internal sealed class ShaderProgramCache
         if (entry is null)
         {
             entry = CreateEntry(source, options);
+            key = BuildProgramKey(source);
             _programs.Add(key, entry);
             ShaderCacheCounters.CountProgram();
         }
 
-        var snapshot = new ResourceSnapshot();
-        var specialization = new ResourceSpecialization();
+        // Both are assigned by the materialization that follows; a failed one throws.
+        ResourceSnapshot snapshot = null!;
+        ResourceSpecialization specialization = null!;
         var captureIndirectImageFailure = _spirvDumpEnabled ? ShaderPermutationDump.CreateFailureCapture(source) : null;
         if (Diagnostics.GpuReadTrace.Enabled)
         {
@@ -295,7 +330,7 @@ internal sealed class ShaderProgramCache
             }
         }
 
-        var permutation = CompilePermutation(source, options, entry, specialization, pushDataCursor, key, sourceWasCached);
+        var permutation = CompilePermutation(source, options, entry, specialization, pushDataCursor, key ?? BuildProgramKey(source), sourceWasCached);
         entry.Permutations.Add(permutation);
         ShaderCacheCounters.CountPermutation();
         stage = CreateStageResources(permutation.Program, snapshot, source, options);
@@ -308,6 +343,10 @@ internal sealed class ShaderProgramCache
         return permutation.Handle;
     }
 
+    // The key of the entry the static state was just built for.
+    private ProgramKey BuildProgramKey(ShaderSource source) =>
+        new(source.Stage, source.Hash, (uint)source.UserData.Length, source.CodeSize, _staticState.ToArray());
+
     private static ShaderStageResources CreateStageResources(
         ShaderProgramInfo program, ResourceSnapshot snapshot, ShaderSource source, StageCompileOptions options)
     {
@@ -646,7 +685,7 @@ internal sealed class ShaderProgramCache
 
             default:
                 return BuildComputeRequest(entry.Plan, resources, layout, options.ComputeInfo!, options.ComputeSystemRegisters,
-                    sharedInt64Atomics, _host.ExecGuardElisionEnabled, signedZeroInfNanPreserve);
+                    sharedInt64Atomics, _host.ExecGuardElisionEnabled, signedZeroInfNanPreserve, _host.ComputeWave64SubgroupNative);
         }
     }
 
@@ -663,10 +702,11 @@ internal sealed class ShaderProgramCache
 
     private static ShaderCompileRequest BuildComputeRequest(ShaderResourcePlan plan, SpecializedResourceInfo resources, BindingLayout layout,
         ComputeInputInfo info, Gen5ComputeSystemRegisters? systemRegisters, bool sharedInt64Atomics, bool execGuardElision,
-        bool signedZeroInfNanPreserve) =>
+        bool signedZeroInfNanPreserve, bool hostWave64Supported = false) =>
         new(plan, resources, layout)
         {
             WaveSize = info.WaveSize,
+            HostWave64Supported = info.WaveSize == 64 && hostWave64Supported,
             EnableExecGuardElision = info.WaveSize != 64 || execGuardElision,
             TraceDeviceAddressFaults = SharpEmu.HLE.GpuMemory.GuestGpuMemoryHook.TraceEnabled,
             ScratchDwords = info.ScratchDwords,
@@ -681,7 +721,7 @@ internal sealed class ShaderProgramCache
 
     internal static bool TryCompilePrewarm(ComputePrewarmRecord record, ShaderCodeCapture code, IGuestGpuBackend compiler,
         bool sharedInt64Atomics, bool execGuardElision, bool signedZeroInfNanPreserve, out IGuestCompiledShader? compiled,
-        out BindingLayout? layout, out string error)
+        out BindingLayout? layout, out string error, bool hostWave64Supported = false)
     {
         compiled = null;
         layout = null;
@@ -698,7 +738,7 @@ internal sealed class ShaderProgramCache
             layout = AllocateLayout(program, plan, resources, record.UserDataBase, record.UserDataCount, record.PushDataCursor,
                 record.Info.DispatchThreadDimensions);
             var request = BuildComputeRequest(plan, resources, layout, record.Info, record.SystemRegisters,
-                sharedInt64Atomics, execGuardElision, signedZeroInfNanPreserve);
+                sharedInt64Atomics, execGuardElision, signedZeroInfNanPreserve, hostWave64Supported);
             return compiler.TryCompileProgram(request, out compiled, out error) && compiled is not null;
         }
         catch (Exception exception)

src/SharpEmu.Libs/Gpu/Rendering/ParallelProgramPrefetch.cs

@@ -0,0 +1,282 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Gpu.GpuCommands.Registers;
+using SharpEmu.Libs.Gpu.Scheduling;
+
+namespace SharpEmu.Libs.Gpu.Rendering;
+
+// Resolves the programs of deferred draws on a worker thread, in issue order, while the render
+// thread interprets and records. The render thread takes a result for the draw it runs only when
+// the resolution completed, used the inputs the draw derives and still matches the host's
+// GPU-ownership state; otherwise it resolves the draw itself.
+public sealed class ParallelProgramPrefetch : IProgramPrefetch, IDisposable
+{
+    public sealed class Job(RegisterBanks banks, long gate)
+    {
+        internal readonly RegisterBanks Banks = banks;
+        // The count of stream items other than draws that must have run before the job resolves.
+        internal readonly long Gate = gate;
+        internal GraphicsProgramInputs Inputs;
+        internal GraphicsPrograms? Programs;
+        internal object? Assumptions;
+        // 0 queued, 1 resolved, 2 abandoned.
+        internal volatile int State;
+    }
+
+    private readonly RenderExecutor _executor;
+    private readonly IRenderHost _host;
+    private readonly IConcurrentResolveHost _concurrent;
+    private readonly object _lock = new();
+    private readonly Queue<Job> _queue = new();
+    private readonly Thread _thread;
+    private int _outstanding;
+    private long _finished;
+    private long _executedGate;
+    // The gate the worker sleeps on; long.MaxValue while it does not.
+    private long _sleepingOnGate = long.MaxValue;
+    private bool _paused;
+    private bool _running;
+    private bool _stopping;
+
+    public ParallelProgramPrefetch(RenderExecutor executor, IRenderHost host, IConcurrentResolveHost concurrent)
+    {
+        _executor = executor;
+        _host = host;
+        _concurrent = concurrent;
+        _thread = new Thread(Run) { IsBackground = true, Name = "SharpEmu program resolve" };
+        _thread.Start();
+    }
+
+    // The job of the draw the render thread runs now; set around each deferred draw.
+    public Job? Current { get; set; }
+
+    public long Submitted { get; private set; }
+
+    // A gate above zero holds the job until AdvanceGate reaches it: with the command stream on its own
+    // thread, draws are submitted ahead of the work before them, which may change what they read.
+    public Job Submit(RegisterBanks banks, long gate = 0)
+    {
+        var job = new Job(banks, gate);
+        lock (_lock)
+        {
+            _queue.Enqueue(job);
+            Submitted++;
+            _outstanding++;
+            Monitor.PulseAll(_lock);
+        }
+
+        return job;
+    }
+
+    // Returns once the worker has finished every submitted job: it reads guest memory and cache
+    // state, which the render thread changes only with the worker idle.
+    public void WaitIdle()
+    {
+        if (Volatile.Read(ref _outstanding) == 0)
+        {
+            return;
+        }
+
+        lock (_lock)
+        {
+            while (_outstanding != 0)
+            {
+                Monitor.Wait(_lock);
+            }
+        }
+    }
+
+    // The render thread ran this many non-draw items: jobs gated on them may resolve.
+    public void AdvanceGate(long executed)
+    {
+        Interlocked.Exchange(ref _executedGate, executed);
+        // Only a worker asleep on a gate this reaches needs a wake: one per item costs the render
+        // thread a kernel call each.
+        if (Volatile.Read(ref _sleepingOnGate) <= executed)
+        {
+            lock (_lock)
+            {
+                Monitor.PulseAll(_lock);
+            }
+        }
+    }
+
+    // Returns once the worker finished the first count jobs it was given, so a non-draw item runs
+    // with no resolution of the draws before it still reading.
+    public void WaitFinished(long count)
+    {
+        if (Volatile.Read(ref _finished) >= count)
+        {
+            return;
+        }
+
+        lock (_lock)
+        {
+            while (_finished < count)
+            {
+                Monitor.Wait(_lock);
+            }
+        }
+    }
+
+    // Holds the worker between jobs, for host work that changes what a resolution reads.
+    public void Pause()
+    {
+        lock (_lock)
+        {
+            _paused = true;
+            while (_running)
+            {
+                Monitor.Wait(_lock);
+            }
+        }
+    }
+
+    public void Resume()
+    {
+        lock (_lock)
+        {
+            _paused = false;
+            Monitor.PulseAll(_lock);
+        }
+    }
+
+    public bool TryTakeGraphics(RegisterBanks banks, in GraphicsProgramInputs inputs, out GraphicsPrograms programs)
+    {
+        programs = null!;
+        var job = Current;
+        if (job is null || !ReferenceEquals(job.Banks, banks))
+        {
+            return false;
+        }
+
+        Current = null;
+        WaitFor(job);
+        if (job.State != 1 || !job.Inputs.Equals(inputs) || !_concurrent.IsStillCurrent(job.Assumptions!))
+        {
+            return false;
+        }
+
+        programs = job.Programs!;
+        return true;
+    }
+
+    private void WaitFor(Job job)
+    {
+        var spinner = new SpinWait();
+        for (var spin = 0; spin < 64 && job.State == 0; spin++)
+        {
+            spinner.SpinOnce(sleep1Threshold: -1);
+        }
+
+        if (job.State != 0)
+        {
+            return;
+        }
+
+        lock (_lock)
+        {
+            while (job.State == 0)
+            {
+                Monitor.Wait(_lock);
+            }
+        }
+    }
+
+    private void Run()
+    {
+        SubmissionScheduler.ThrowOnFatal = true;
+        while (true)
+        {
+            Job job;
+            lock (_lock)
+            {
+                while (_queue.Count == 0 && !_stopping)
+                {
+                    Monitor.Wait(_lock);
+                }
+
+                if (_stopping)
+                {
+                    return;
+                }
+
+                job = _queue.Peek();
+                if (job.Gate > Volatile.Read(ref _executedGate))
+                {
+                    // The gate usually opens within a few items: spin before sleeping.
+                    Monitor.Exit(_lock);
+                    try
+                    {
+                        var spinner = new SpinWait();
+                        for (var spin = 0; spin < 50 && job.Gate > Volatile.Read(ref _executedGate); spin++)
+                        {
+                            spinner.SpinOnce(sleep1Threshold: -1);
+                        }
+                    }
+                    finally
+                    {
+                        Monitor.Enter(_lock);
+                    }
+                }
+
+                while ((_paused || job.Gate > Volatile.Read(ref _executedGate)) && !_stopping)
+                {
+                    Interlocked.Exchange(ref _sleepingOnGate, _paused ? long.MaxValue : job.Gate);
+                    if (!_paused && job.Gate <= Volatile.Read(ref _executedGate))
+                    {
+                        break;
+                    }
+
+                    Monitor.Wait(_lock, 50);
+                }
+
+                Volatile.Write(ref _sleepingOnGate, long.MaxValue);
+                if (_stopping)
+                {
+                    return;
+                }
+
+                _queue.Dequeue();
+                _running = true;
+            }
+
+            Resolve(job);
+            lock (_lock)
+            {
+                _outstanding--;
+                _finished++;
+                _running = false;
+                Monitor.PulseAll(_lock);
+            }
+        }
+    }
+
+    private void Resolve(Job job)
+    {
+        _concurrent.BeginConcurrentResolve();
+        try
+        {
+            job.Inputs = RenderExecutor.ProgramInputsOf(job.Banks, _host.FormatSupport, _host.Fatal);
+            job.Programs = _executor.ResolveGraphicsPrograms(job.Banks, in job.Inputs);
+            job.Assumptions = _concurrent.EndConcurrentResolve();
+            job.State = 1;
+        }
+        catch (Exception)
+        {
+            // The render thread resolves the draw itself, and fails there if the error is real.
+            job.Assumptions = _concurrent.EndConcurrentResolve();
+            job.State = 2;
+        }
+    }
+
+    public void Dispose()
+    {
+        lock (_lock)
+        {
+            _stopping = true;
+            Monitor.PulseAll(_lock);
+        }
+    }
+}

src/SharpEmu.Libs/Gpu/Rendering/RenderExecutor.Compute.cs

@@ -714,15 +714,19 @@ public sealed partial class RenderExecutor
         var hash = input.Stage.Program!.Hash;
         if (!_host.TryClearImageFromBuffer(address, clear.Size, clear.PackedClear))
         {
-            // A metadata fill may run before its target is bound; the store keeps it pending and the dispatch runs.
+            // A metadata fill may run before its target is bound; the store keeps it pending. The fill
+            // itself then runs on the host when the range is known DCC metadata: run as a dispatch it
+            // leaves the metadata GPU-written, and every later bind of the surface waits for the GPU to
+            // read it back (Astro Bot fills several surfaces this way each frame).
             var registered = _host.TryAbsorbDccFill(address, clear.Size, clear.PackedClear);
+            var filledOnHost = !registered && _host.TryFillDccMetadata(address, clear.Size, clear.PackedClear);
             if (RenderTrace.Enabled && RenderTrace.MetadataClear())
             {
                 RenderTrace.Write(
-                    $"{(registered ? "Tracked" : "Deferred")} a metadata clear: shader=0x{hash:X16} address=0x{address:X16} size=0x{clear.Size:X16} value=0x{clear.PackedClear:X8}");
+                    $"{(registered ? "Tracked" : filledOnHost ? "Filled on the host" : "Deferred")} a metadata clear: shader=0x{hash:X16} address=0x{address:X16} size=0x{clear.Size:X16} value=0x{clear.PackedClear:X8}");
             }
 
-            return registered;
+            return registered || filledOnHost;
         }
 
         if (RenderTrace.Enabled && RenderTrace.ImageClear())

src/SharpEmu.Libs/Gpu/Rendering/RenderExecutor.Prefetch.cs

@@ -0,0 +1,115 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Gpu.Buffers;
+using SharpEmu.Libs.Gpu.GpuCommands.Registers;
+using SharpEmu.Libs.Gpu.Images;
+
+namespace SharpEmu.Libs.Gpu.Rendering;
+
+// What a draw's program resolution reads besides the stage and context registers: the export
+// mapping of every bound color slot, whether the pixel stage runs and whether a depth target is
+// bound. All three follow from the register banks alone.
+public struct GraphicsProgramInputs : IEquatable<GraphicsProgramInputs>
+{
+    [System.Runtime.CompilerServices.InlineArray(RenderingState.ColorAttachmentCapacity)]
+    public struct ExportMappings
+    {
+        private ColorComponentMap _element0;
+    }
+
+    public ExportMappings Mapping;
+    public bool PixelActive;
+    public bool DepthBound;
+
+    public static GraphicsProgramInputs Create()
+    {
+        var inputs = new GraphicsProgramInputs();
+        ((Span<ColorComponentMap>)inputs.Mapping).Fill(ColorComponentMap.Identity);
+        return inputs;
+    }
+
+    public readonly bool Equals(GraphicsProgramInputs other) =>
+        PixelActive == other.PixelActive && DepthBound == other.DepthBound &&
+        ((ReadOnlySpan<ColorComponentMap>)Mapping).SequenceEqual(other.Mapping);
+
+    public override readonly bool Equals(object? obj) => obj is GraphicsProgramInputs other && Equals(other);
+
+    public override readonly int GetHashCode()
+    {
+        var hash = new HashCode();
+        foreach (var map in (ReadOnlySpan<ColorComponentMap>)Mapping)
+        {
+            hash.Add(map.Packed);
+        }
+
+        hash.Add(PixelActive);
+        hash.Add(DepthBound);
+        return hash.ToHashCode();
+    }
+}
+
+// The host side of resolving programs on a thread other than the render thread. The resolving
+// thread records what it assumed of the host's GPU-ownership state; the render thread checks,
+// when it takes the result, that the assumptions still hold.
+public interface IConcurrentResolveHost
+{
+    // On the resolving thread, around one resolution.
+    void BeginConcurrentResolve();
+
+    object EndConcurrentResolve();
+
+    // On the render thread.
+    bool IsStillCurrent(object assumptions);
+}
+
+// Program resolutions computed ahead of their draws. A draw takes one only when it was resolved
+// from the same register banks with the same inputs; otherwise it resolves its programs itself.
+public interface IProgramPrefetch
+{
+    bool TryTakeGraphics(RegisterBanks banks, in GraphicsProgramInputs inputs, out GraphicsPrograms programs);
+}
+
+public sealed partial class RenderExecutor
+{
+    public IProgramPrefetch? Prefetch { get; set; }
+
+    // The inputs ResolveShaderPrograms derives from the targets TryResolveDrawTargets binds, from the
+    // registers alone: the target requests and the pixel and depth activity, without the images.
+    public static GraphicsProgramInputs ProgramInputsOf(RegisterBanks banks, IImageFormatSupport formats, Func<string, Exception> fatal)
+    {
+        var inputs = GraphicsProgramInputs.Create();
+        var context = banks.Context;
+        Span<ColorComponentMap> mapping = inputs.Mapping;
+        for (var slot = 0u; slot < ContextRegisters.ColorTargetCount; slot++)
+        {
+            if (slot != 0 && (context.RenderTargetMaskForSlot(slot) == 0 || context.ColorTargets[slot].BaseAddress == 0))
+            {
+                continue;
+            }
+
+            if (ColorTargetResolver.Resolve(context, slot, DrawLayerOffset, ignoreTargetMask: false, out var resolvedSlot) is { } resolution)
+            {
+                mapping[(int)resolvedSlot] = resolution.ExportMapping;
+            }
+        }
+
+        inputs.DepthBound = DepthTargetResolver.Resolve(context, formats, fatal) is not null;
+        inputs.PixelActive = HasActivePixelShader(banks);
+        return inputs;
+    }
+
+    // The programs of a draw with these banks and inputs, as ResolveShaderPrograms resolves them.
+    public GraphicsPrograms ResolveGraphicsPrograms(RegisterBanks banks, in GraphicsProgramInputs inputs)
+    {
+        var context = banks.Context;
+        return _pipelines.GetGraphicsPrograms(
+            banks.Shader.Vertex,
+            banks.Shader.Pixel,
+            context.ShaderInterface,
+            context,
+            inputs.Mapping,
+            inputs.PixelActive,
+            inputs.DepthBound);
+    }
+}

src/SharpEmu.Libs/Gpu/Rendering/RenderExecutor.Resources.cs

@@ -35,7 +35,8 @@ public sealed partial class RenderExecutor
             throw _host.Fatal($"The vertex input has too many buffers: count={buffers.Length} max={VertexInputInfo.MaxBuffers}.");
         }
 
-        Span<VertexBufferRange> ranges = stackalloc VertexBufferRange[VertexInputInfo.MaxBuffers];
+        // Sized to this draw's buffers: a MaxBuffers span is over a kilobyte, zeroed on every draw.
+        Span<VertexBufferRange> ranges = stackalloc VertexBufferRange[buffers.Length];
         var rangeCount = 0;
         foreach (ref readonly var vertex in buffers.AsSpan())
         {
@@ -54,7 +55,7 @@ public sealed partial class RenderExecutor
         }
 
         ranges[..rangeCount].Sort(static (left, right) => left.BaseAddress.CompareTo(right.BaseAddress));
-        Span<VertexBufferRange> merged = stackalloc VertexBufferRange[VertexInputInfo.MaxBuffers];
+        Span<VertexBufferRange> merged = stackalloc VertexBufferRange[buffers.Length];
         var mergedCount = 0;
         for (var i = 0; i < rangeCount; i++)
         {

src/SharpEmu.Libs/Gpu/Rendering/RenderExecutor.cs

@@ -667,29 +667,22 @@ public sealed partial class RenderExecutor
     private void ResolveShaderPrograms(RegisterBanks banks, ref DrawState state)
     {
         using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.DrawProgramResolution);
-        var context = banks.Context;
-        Span<ColorComponentMapArray> mappingStorage = stackalloc ColorComponentMapArray[1];
-        Span<ColorComponentMap> targetExportMapping = mappingStorage[0];
-        targetExportMapping.Fill(ColorComponentMap.Identity);
+        var inputs = GraphicsProgramInputs.Create();
+        Span<ColorComponentMap> targetExportMapping = inputs.Mapping;
         foreach (ref readonly var color in BoundColors(ref state))
         {
             targetExportMapping[(int)color.Slot] = color.Resolution.ExportMapping;
         }
 
-        state.Programs = _pipelines.GetGraphicsPrograms(
-            banks.Shader.Vertex,
-            banks.Shader.Pixel,
-            context.ShaderInterface,
-            context,
-            targetExportMapping,
-            state.PixelActive,
-            state.Depth.HasTarget);
-    }
+        inputs.PixelActive = state.PixelActive;
+        inputs.DepthBound = state.Depth.HasTarget;
+        if (Prefetch is { } prefetch && prefetch.TryTakeGraphics(banks, in inputs, out var prefetched))
+        {
+            state.Programs = prefetched;
+            return;
+        }
 
-    [System.Runtime.CompilerServices.InlineArray(RenderingState.ColorAttachmentCapacity)]
-    private struct ColorComponentMapArray
-    {
-        private ColorComponentMap _element0;
+        state.Programs = ResolveGraphicsPrograms(banks, in inputs);
     }
 
     private static void TraceDrawState(ulong submitId, RegisterBanks banks, in DrawCall draw, in DrawState state)

src/SharpEmu.Libs/Gpu/Rendering/RenderingState.cs

@@ -19,7 +19,29 @@ public readonly record struct RenderingAttachment(
     bool HasDepth,
     bool DepthClear,
     bool HasStencil,
-    bool StencilClear);
+    bool StencilClear)
+{
+    // Compared by handle value: the generated comparison goes through the object comparer of ImageView
+    // and boxes both views, which a draw pays for every attachment it re-checks.
+    public bool Equals(RenderingAttachment other) =>
+        View.Handle == other.View.Handle &&
+        Layout == other.Layout &&
+        Format == other.Format &&
+        ClearWord0 == other.ClearWord0 &&
+        ClearWord1 == other.ClearWord1 &&
+        ClearWord2 == other.ClearWord2 &&
+        ClearWord3 == other.ClearWord3 &&
+        IsClear == other.IsClear &&
+        HasDepth == other.HasDepth &&
+        DepthClear == other.DepthClear &&
+        HasStencil == other.HasStencil &&
+        StencilClear == other.StencilClear;
+
+    public override int GetHashCode() => HashCode.Combine(
+        View.Handle,
+        HashCode.Combine(Layout, Format, ClearWord0, ClearWord1, ClearWord2, ClearWord3),
+        HashCode.Combine(IsClear, HasDepth, DepthClear, HasStencil, StencilClear));
+}
 
 [InlineArray(RenderingState.ColorAttachmentCapacity)]
 public struct RenderingColorAttachments

src/SharpEmu.Libs/Gpu/Scheduling/GpuWritebackEpoch.cs

@@ -0,0 +1,16 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.Gpu.Scheduling;
+
+// Counts the stores of GPU data into guest memory: buffer and image downloads, and buffer transfers
+// the caches complete there. Work done ahead of the render thread from guest memory is still
+// current while the count has not moved.
+internal static class GpuWritebackEpoch
+{
+    private static long _value;
+
+    public static long Value => Volatile.Read(ref _value);
+
+    public static void Advance() => Interlocked.Increment(ref _value);
+}

src/SharpEmu.Libs/Gpu/Scheduling/SubmissionScheduler.cs

@@ -88,9 +88,21 @@ public sealed class SubmissionScheduler : IGpuTickScheduler, IDisposable
 
     internal bool PriorityWorkerAlive => _priorityThread.IsAlive;
 
+    // Set on a thread that does work ahead of the render thread: there a fatal condition only
+    // abandons that work, which the render thread then does itself and fails on if it is real.
+    [ThreadStatic]
+    internal static bool ThrowOnFatal;
+
+    public sealed class AbandonedWorkException(string message) : Exception(message);
+
     // Report the fatal error, then return an exception to stop the caller.
     internal static Exception Fatal(string message)
     {
+        if (ThrowOnFatal)
+        {
+            throw new AbandonedWorkException(message);
+        }
+
         Console.Error.WriteLine($"[GPU][FATAL] {message}");
         OnFatal(message);
         return new InvalidOperationException(message);

src/SharpEmu.Libs/Gpu/Vulkan/GraphicsDynamicStateEpoch.cs

@@ -0,0 +1,16 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.Gpu.Vulkan;
+
+// Advanced by every graphics pipeline bind or dynamic state change outside the guest draw path:
+// such a pipeline may fix the state the guest pipelines leave dynamic, so the dynamic state a
+// guest draw last set no longer holds in the command buffer.
+internal static class GraphicsDynamicStateEpoch
+{
+    private static long _value;
+
+    public static long Value => Volatile.Read(ref _value);
+
+    public static void Advance() => Interlocked.Increment(ref _value);
+}

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.CachedImages.cs

@@ -130,6 +130,37 @@ internal static unsafe partial class VulkanVideoPresenter
         // A draw-local copy used when the guest samples a depth aspect while the
         // same image is attached for writes. It retires with the submission.
         public bool IsFeedbackSnapshot;
+        // Taken from the presenter's pool; goes back to it when its submission retires.
+        public bool Pooled;
+        public bool InPool;
+
+        // Back to the state of a new object, so a reused one never carries a previous binding.
+        public void Reset()
+        {
+            Address = 0;
+            ImageIdentifier = default;
+            Request = default;
+            CachedImage = null;
+            MipLevel = 0;
+            Image = default;
+            View = default;
+            Layout = ImageLayout.ShaderReadOnlyOptimal;
+            Sampler = default;
+            SamplerState = default;
+            IsStorage = false;
+            IsHostMovie = false;
+            HostMoviePlane = -1;
+            HostMovieFrameSerial = 0;
+            NeedsUpload = false;
+            StagingBuffer = default;
+            StagingMemory = default;
+            Width = 0;
+            Height = 0;
+            RowLength = 0;
+            DestinationSelect = 0;
+            MipViews = [];
+            IsFeedbackSnapshot = false;
+        }
     }
 
     private sealed partial class Presenter

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.CommandStream.cs

@@ -176,9 +176,6 @@ internal static unsafe partial class VulkanVideoPresenter
         private EndOfPipe _endOfPipe = null!;
         private AgcExports.CommandStreamTranslation _translation = null!;
         private ICpuMemory _guestMemory = null!;
-        private long _lastCommandStreamProgressTicks;
-        private static readonly long _blockedRetryTicks =
-            CommandStreamQueue.AllBlockedRetryMilliseconds * System.Diagnostics.Stopwatch.Frequency / 1000L;
 
         public CommandStreamQueue CommandStream => _commandStream;
 
@@ -188,7 +185,6 @@ internal static unsafe partial class VulkanVideoPresenter
             _guestMemory = guest;
             _endOfPipe = new EndOfPipe(_endOfPipeEvents, this);
             _translation = AgcExports.CreateCommandStreamTranslation(guest, this);
-            _lastCommandStreamProgressTicks = System.Diagnostics.Stopwatch.GetTimestamp();
         }
 
         public void EnqueueCommandStream(uint queue, ulong address, uint dwordCount, ulong submissionId, object? geometrySnapshots)
@@ -204,25 +200,42 @@ internal static unsafe partial class VulkanVideoPresenter
             }
         }
 
-        // Runs slices until the budget ends or nothing is runnable; blocked heads retry every 100 ms.
+        // The command stream thread interprets the submissions; this thread runs what it hands over
+        // until the budget ends or nothing is queued. Null until the first render tick.
+        private CommandThread? _commandThread;
+        private Action? _betweenCommandItems;
+
+        private bool OnCommandThread => _commandThread is { IsProducerThread: true };
+
+        // Items run in batches; completions are collected between them so blocked waits retry.
+        private const int CommandItemBatch = 256;
+
         private void RunCommandStreamSlices(long renderWorkDeadline)
         {
-            var slices = 0;
-            while (slices < _maxGuestWorkPerRender)
+            if (_commandThread is null)
+            {
+                _commandThread = new CommandThread(_commandStream, _guestMemory, RunDeferredWork, _translation.Prefetch);
+                _betweenCommandItems = RunRelayBetweenCommandItems;
+                _translation.Sink = _commandThread;
+                _commandThread.Start();
+            }
+
+            var thread = _commandThread;
+            for (;;)
             {
                 using (RenderPhaseProfile.Measure(RenderPhaseProfile.Phase.Collect))
                 {
                     CollectCompletedGuestSubmissions(waitForOldest: false);
                 }
 
-                SliceResult result;
+                int ran;
                 using (RenderPhaseProfile.Measure(RenderPhaseProfile.Phase.CommandStream))
                 {
                     var aliasAccess = _guestBacking?.TryEnterBackingAliasAccess() == true;
                     _backingAliasAccess = aliasAccess;
                     try
                     {
-                        result = _commandStream.ProcessOne();
+                        ran = thread.RunBatch(renderWorkDeadline, CommandItemBatch, _betweenCommandItems!);
                     }
                     finally
                     {
@@ -234,64 +247,162 @@ internal static unsafe partial class VulkanVideoPresenter
                     }
                 }
 
-                switch (result)
-                {
-                    case SliceResult.NoWork:
-                        return;
-                    case SliceResult.AllBlocked:
-                        RetryBlockedCommandStreamIfDue();
-                        return;
-                    case SliceResult.BlockedWithoutProgress:
-                        if (!_commandStream.HasUnblockedPending)
-                        {
-                            RetryBlockedCommandStreamIfDue();
-                            return;
-                        }
-
-                        break;
-                    default:
-                        _lastCommandStreamProgressTicks = System.Diagnostics.Stopwatch.GetTimestamp();
-                        break;
-                }
-
-                slices++;
-                if (System.Diagnostics.Stopwatch.GetTimestamp() >= renderWorkDeadline)
+                if (ran == 0 || System.Diagnostics.Stopwatch.GetTimestamp() >= renderWorkDeadline)
                 {
                     return;
                 }
             }
         }
 
-        private void RetryBlockedCommandStreamIfDue()
+        private void RunDeferredWork(AgcExports.DeferredWork work)
         {
-            var now = System.Diagnostics.Stopwatch.GetTimestamp();
-            if (now - _lastCommandStreamProgressTicks < _blockedRetryTicks)
+            using var translationScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandDrawTranslation);
+            var started = System.Diagnostics.Stopwatch.GetTimestamp();
+            try
             {
-                return;
+                _translation.RunDeferred(work);
+            }
+            finally
+            {
+                Interlocked.Add(ref _perfDrawTicks, System.Diagnostics.Stopwatch.GetTimestamp() - started);
             }
-
-            _commandStream.RetryBlocked();
-            _lastCommandStreamProgressTicks = now;
         }
 
-        // Milliseconds until the blocked heads are due for a retry; null when none is blocked.
-        private int? BlockedRetryWaitMilliseconds()
+        // Commands other threads post to the GPU worker run between items, with the resolution
+        // worker held: they change the caches it reads.
+        private void RunRelayBetweenCommandItems()
         {
-            if (!_vulkanReady || !_commandStream.HasPending || _commandStream.HasUnblockedPending)
+            if (_relay.HasPendingCommands)
             {
-                return null;
+                RunRelayCommandsWithPrefetchHeld();
             }
-
-            var remaining = _blockedRetryTicks - (System.Diagnostics.Stopwatch.GetTimestamp() - _lastCommandStreamProgressTicks);
-            return remaining <= 0 ? 0 : (int)Math.Max(1, remaining * 1000 / System.Diagnostics.Stopwatch.Frequency);
         }
 
-        ICpuMemory ICommandStreamHost.Memory => _guestMemory;
+        private void RunRelayCommandsWithPrefetchHeld()
+        {
+            var prefetch = _commandThread is not null ? _translation.Prefetch : null;
+            prefetch?.Pause();
+            try
+            {
+                using var relayScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.QueueRelay);
+                _relay.RunPendingCommands();
+            }
+            finally
+            {
+                prefetch?.Resume();
+            }
+        }
+
+        // The interpreter's own stores (labels, atomics, semaphores) follow every deferred draw.
+        private DrainingWriteMemory? _commandMemory;
+
+        ICpuMemory ICommandStreamHost.Memory => _commandMemory ??= new DrainingWriteMemory(_guestMemory, DrainDraws, this);
+
+        // On the command stream thread a store is queued behind the work before it and reads see
+        // the queued stores; elsewhere the deferred draws run first.
+        private sealed class DrainingWriteMemory(ICpuMemory inner, Action drain, Presenter presenter) : ICpuMemory, ICpuMemoryWrapper
+        {
+            public ICpuMemory Inner => inner;
+
+            public bool TryRead(ulong virtualAddress, Span<byte> destination) =>
+                presenter.OnCommandThread
+                    ? presenter._commandThread!.ReadOverlay(virtualAddress, destination)
+                    : inner.TryRead(virtualAddress, destination);
+
+            public bool TryWrite(ulong virtualAddress, ReadOnlySpan<byte> source)
+            {
+                if (presenter.OnCommandThread)
+                {
+                    presenter._commandThread!.PostWrite(virtualAddress, source);
+                    return true;
+                }
+
+                drain();
+                return inner.TryWrite(virtualAddress, source);
+            }
+
+            public bool TryCompare(ulong virtualAddress, ReadOnlySpan<byte> expected, out bool equal)
+            {
+                if (presenter.OnCommandThread)
+                {
+                    var bytes = expected.ToArray();
+                    var compared = false;
+                    var result = presenter._commandThread!.Run(() => TryCompare(virtualAddress, bytes, out compared));
+                    equal = compared;
+                    return result;
+                }
+
+                drain();
+                return inner.TryCompare(virtualAddress, expected, out equal);
+            }
+
+            public bool TryCopy(ulong destinationAddress, ulong sourceAddress, ulong length)
+            {
+                if (presenter.OnCommandThread)
+                {
+                    return presenter._commandThread!.Run(() => TryCopy(destinationAddress, sourceAddress, length));
+                }
+
+                drain();
+                return inner.TryCopy(destinationAddress, sourceAddress, length);
+            }
+
+            public bool TryScanCString(ulong address, byte needle, bool findLast, ulong maxLength, out ulong match) =>
+                inner.TryScanCString(address, needle, findLast, maxLength, out match);
+
+            public bool CanRead(ulong address, ulong size) => inner.CanRead(address, size);
+
+            public string DescribeReadRange(ulong address, ulong size) => inner.DescribeReadRange(address, size);
+        }
+
+        // A read the render thread runs after everything queued before it, for the command thread.
+        private bool ReadOnRenderThread(ulong address, Span<byte> destination, bool operand)
+        {
+            var buffer = new byte[destination.Length];
+            var read = _commandThread!.Run(() => operand ? TryReadGuestOperand(address, buffer) : TryReadGuest(address, buffer));
+            buffer.CopyTo(destination);
+            return read;
+        }
+
+        // Command memory and polled labels: GPU-written pages go through the render thread, the
+        // rest is read here with the stream's queued stores applied.
+        private bool ReadOnCommandThread(ulong address, Span<byte> destination)
+        {
+            _commandThread!.WaitForPendingRange(address, (ulong)destination.Length);
+            return _bufferCache.MayHaveGpuDirtyPages(address, (ulong)destination.Length)
+                ? ReadOnRenderThread(address, destination, operand: false)
+                : _commandThread!.ReadOverlay(address, destination);
+        }
+
+        public bool TryReadGuestWaitOperand(ulong address, Span<byte> destination) =>
+            OnCommandThread ? ReadOnCommandThread(address, destination) : TryReadGuestOperand(address, destination);
+
+        public bool TryReadGuestOperand(ulong address, Span<byte> destination)
+        {
+            if (OnCommandThread)
+            {
+                return ReadOnRenderThread(address, destination, operand: true);
+            }
+
+            DrainDraws();
+            return TryReadGuest(address, destination);
+        }
 
         public bool TryReadGuest(ulong address, Span<byte> destination)
         {
+            if (OnCommandThread)
+            {
+                return ReadOnCommandThread(address, destination);
+            }
+
             using (RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandMemorySync))
             {
+                // GPU-written command memory is downloaded first: deferred draws are recorded before it.
+                if (_translation is { HasPendingDraws: true } && _bufferCache.MayHaveGpuDirtyPages(address, (ulong)destination.Length))
+                {
+                    DrainDraws();
+                }
+
                 if (!_bufferCache.TrySynchronizeCpuRead(address, (ulong)destination.Length,
                     SharpEmu.HLE.GuestMemory.GuestMemoryProfile.ReadbackSource.CommandMemoryRead))
                 {
@@ -305,27 +416,66 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void RunPendingCommands()
         {
+            // With the stream on its own thread the render thread runs them between its items.
+            if (OnCommandThread)
+            {
+                return;
+            }
+
+            if (_relay.HasPendingCommands)
+            {
+                DrainDraws();
+            }
+
             using var relayScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.QueueRelay);
             _relay.RunPendingCommands();
         }
 
         public void BeginSubmission(int queueId, ulong submissionId, object? geometrySnapshots)
+        {
+            if (OnCommandThread)
+            {
+                // The translation state belongs to this thread; the queue context to the render thread.
+                _translation.BeginSubmission(queueId, submissionId, geometrySnapshots, _commandStream.GetInterpreter(queueId));
+                _commandThread!.Post(() => BeginSubmissionContext(queueId, submissionId));
+                return;
+            }
+
+            DrainDraws();
+            BeginSubmissionContext(queueId, submissionId);
+            _translation.BeginSubmission(queueId, submissionId, geometrySnapshots, _commandStream.GetInterpreter(queueId));
+        }
+
+        private void BeginSubmissionContext(int queueId, ulong submissionId)
         {
             using var contextScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.QueueContext);
             _activeGuestQueue = new VulkanGuestQueueIdentity(_commandQueueNames[queueId], submissionId);
             BindSubmissionContext(_activeGuestQueue);
             _ = CurrentRecordingBuffer();
-            _translation.BeginSubmission(queueId, submissionId, geometrySnapshots, _commandStream.GetInterpreter(queueId));
         }
 
         public void Flush()
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.Post(Flush);
+                return;
+            }
+
+            DrainDraws();
             using var flushScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.Flush);
             FlushBatchedGuestCommands();
         }
 
         public void FlushAndWait()
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.Run(FlushAndWait);
+                return;
+            }
+
+            DrainDraws();
             using var waitScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandGpuWait);
             _ = CurrentRecordingBuffer();
             _scheduler.FlushAndWait();
@@ -334,16 +484,40 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void SynchronizeGpu()
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.Run(SynchronizeGpu);
+                return;
+            }
+
+            DrainDraws();
             using var waitScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandGpuWait);
             _ = CurrentRecordingBuffer();
             _scheduler.Finish();
             CollectCompletedGuestSubmissions(waitForOldest: false);
         }
 
-        public void RunGarbageCollector() => RunGuestCacheCollection(endsFrame: false);
+        public void RunGarbageCollector()
+        {
+            if (OnCommandThread)
+            {
+                _commandThread!.Post(RunGarbageCollector);
+                return;
+            }
+
+            DrainDraws();
+            RunGuestCacheCollection(endsFrame: false);
+        }
 
         public void EmitGlobalBarrier()
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.Post(EmitGlobalBarrier);
+                return;
+            }
+
+            DrainDraws();
             // Inside a rendering scope whose draws only wrote attachments, the barrier matters only
             // to what runs after the scope: everything before the scope is ordered by the barrier
             // BeginRendering records, and a draw that stores to memory ends a deferred scope first.
@@ -380,6 +554,13 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void FillBuffer(ulong address, ulong size, uint value, bool isGds)
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.PostRange(() => FillBuffer(address, size, value, isGds), address, isGds ? 0 : size);
+                return;
+            }
+
+            DrainDraws();
             using var transferScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandMemoryTransfer);
             // Transfers the buffer cache completes in guest memory leave the rendering scope open;
             // its GPU paths end the scope before they record.
@@ -389,6 +570,13 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void CopyBuffer(ulong destination, ulong source, ulong size, bool destinationIsGds, bool sourceIsGds)
         {
+            if (OnCommandThread)
+            {
+                _commandThread!.PostRange(() => CopyBuffer(destination, source, size, destinationIsGds, sourceIsGds), destination, destinationIsGds ? 0 : size);
+                return;
+            }
+
+            DrainDraws();
             using var transferScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandMemoryTransfer);
             // Transfers the buffer cache completes in guest memory leave the rendering scope open;
             // its GPU paths end the scope before they record.
@@ -396,11 +584,30 @@ internal static unsafe partial class VulkanVideoPresenter
             _bufferCache.CopyBuffer(destination, source, size, destinationIsGds, sourceIsGds);
         }
 
-        public void ReadGds(Span<uint> destination, uint wordOffset, uint wordCount) =>
+        public void ReadGds(Span<uint> destination, uint wordOffset, uint wordCount)
+        {
+            if (OnCommandThread)
+            {
+                var words = new uint[destination.Length];
+                _commandThread!.Run(() => ReadGds(words, wordOffset, wordCount));
+                words.CopyTo(destination);
+                return;
+            }
+
+            DrainDraws();
             EndOfPipe.ReadGdsWords(_bufferCache.GdsBuffer.Mapped, destination, wordOffset, wordCount);
+        }
 
         public void RecordEndOfPipe(in EndOfPipeWrite write)
         {
+            if (OnCommandThread)
+            {
+                var queued = write;
+                _commandThread!.Post(() => RecordEndOfPipe(in queued));
+                return;
+            }
+
+            DrainDraws();
             using var completionScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandEndOfPipe);
             _ = BeginBatchedGuestCommands();
             var buffer = _scheduler.Current;
@@ -456,10 +663,26 @@ internal static unsafe partial class VulkanVideoPresenter
             }
         }
 
-        public void TriggerInterrupt(int eventId, uint contextId) => _endOfPipeEvents.TriggerInterrupt(eventId, contextId);
+        public void TriggerInterrupt(int eventId, uint contextId)
+        {
+            if (OnCommandThread)
+            {
+                _commandThread!.Post(() => TriggerInterrupt(eventId, contextId));
+                return;
+            }
+
+            DrainDraws();
+            _endOfPipeEvents.TriggerInterrupt(eventId, contextId);
+        }
 
         public bool HasFlipSlot()
         {
+            if (OnCommandThread)
+            {
+                return _commandThread!.Run(HasFlipSlot);
+            }
+
+            DrainDraws();
             lock (_gate)
             {
                 return _pendingGuestImagePresentations.Count < MaxPendingGuestFlipVersions;
@@ -468,17 +691,47 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public ulong PrepareFlip(int handle, int index, int flipMode, long flipArgument)
         {
+            if (OnCommandThread)
+            {
+                return _commandThread!.Run(() => PrepareFlip(handle, index, flipMode, flipArgument));
+            }
+
+            DrainDraws();
             _ = BeginBatchedGuestCommands();
             return _endOfPipe.PrepareVideoOutFlip(_scheduler.Current, handle, index, flipMode, flipArgument);
         }
 
-        public bool IsFlipDone(int handle, int index) => VideoOutExports.IsFlipDone(handle, index);
+        public bool IsFlipDone(int handle, int index)
+        {
+            if (OnCommandThread)
+            {
+                return _commandThread!.Run(() => IsFlipDone(handle, index));
+            }
 
-        public void PrepareCpuFlip(int handle, int index, ulong requestId) =>
+            DrainDraws();
+            return VideoOutExports.IsFlipDone(handle, index);
+        }
+
+        public void PrepareCpuFlip(int handle, int index, ulong requestId)
+        {
+            if (OnCommandThread)
+            {
+                _commandThread!.Run(() => PrepareCpuFlip(handle, index, requestId));
+                return;
+            }
+
+            DrainDraws();
             CaptureFlip(handle, index, requestId, flipMode: 0, flipArg: 0);
+        }
 
         public void DrawIndexed(ulong submitId, in DrawIndexedArguments arguments)
         {
+            if (OnCommandThread)
+            {
+                _translation.DrawIndexed(submitId, in arguments);
+                return;
+            }
+
             using var translationScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandDrawTranslation);
             var started = System.Diagnostics.Stopwatch.GetTimestamp();
             try
@@ -493,6 +746,12 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void DrawAuto(ulong submitId, in DrawAutoArguments arguments)
         {
+            if (OnCommandThread)
+            {
+                _translation.DrawAuto(submitId, in arguments);
+                return;
+            }
+
             using var translationScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandDrawTranslation);
             var started = System.Diagnostics.Stopwatch.GetTimestamp();
             try
@@ -521,6 +780,13 @@ internal static unsafe partial class VulkanVideoPresenter
 
         public void DispatchDirect(ulong submitId, uint groupsX, uint groupsY, uint groupsZ, uint dispatchInitiator, ulong indirectArgumentsAddress = 0)
         {
+            if (OnCommandThread)
+            {
+                _translation.Dispatch(submitId, groupsX, groupsY, groupsZ, dispatchInitiator, indirectArgumentsAddress);
+                return;
+            }
+
+            DrainDraws();
             using var translationScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandDispatchTranslation);
             var started = System.Diagnostics.Stopwatch.GetTimestamp();
             try
@@ -533,13 +799,48 @@ internal static unsafe partial class VulkanVideoPresenter
             }
         }
 
-        public void OnQueueReset(int queueId) => _translation.QueueReset(queueId);
+        public void OnQueueReset(int queueId)
+        {
+            if (OnCommandThread)
+            {
+                _translation.QueueReset(queueId);
+                return;
+            }
+
+            DrainDraws();
+            _translation.QueueReset(queueId);
+        }
+
+        // Deferred draws run before anything else the stream asks of the host: everything but a
+        // draw may record, read or write what they use. Their time counts as draw time. Completion
+        // callbacks reach some of the same entry points from the GPU priority worker; they are not
+        // ordered with the stream and leave the draws to the render thread.
+        private void DrainDraws()
+        {
+            // Hosts built without a command stream have no translation.
+            if (_translation is not { HasPendingDraws: true } || !_relay.IsGpuQueueThread)
+            {
+                return;
+            }
+
+            using var translationScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.CommandDrawTranslation);
+            var started = System.Diagnostics.Stopwatch.GetTimestamp();
+            try
+            {
+                _translation.DrainPendingDraws();
+            }
+            finally
+            {
+                Interlocked.Add(ref _perfDrawTicks, System.Diagnostics.Stopwatch.GetTimestamp() - started);
+            }
+        }
 
         public Exception Fatal(string message) => SubmissionScheduler.Fatal(message);
 
         // The flip packet: translation first, then the request, then the capture in stream order.
         public int SubmitFlipFromGpu(RecordingBuffer buffer, int handle, int index, int flipMode, long flipArg, out ulong requestId)
         {
+            DrainDraws();
             _ = buffer;
             _translation.PrepareFlip(handle, index);
             var result = VideoOutExports.TryReserveFlipRequest(handle, index, flipMode, flipArg, gpuQueued: true, out requestId);
@@ -563,11 +864,16 @@ internal static unsafe partial class VulkanVideoPresenter
         // The interpreter must suspend before capture so this worker can present queued frames.
         public void WaitForSubmitSlot() => throw SubmissionScheduler.Fatal("The GPU worker cannot wait for a flip slot. Suspend the command stream before capture.");
 
-        public void CompleteFlip(ulong requestId) => VideoOutExports.CompleteFlip(requestId);
+        public void CompleteFlip(ulong requestId)
+        {
+            DrainDraws();
+            VideoOutExports.CompleteFlip(requestId);
+        }
 
         // Copies the display surface through the store; presentation shows the copy once its tick retires.
         internal void CaptureFlip(int handle, int index, ulong requestId, int flipMode, long flipArg)
         {
+            DrainDraws();
             // A no-buffer flip has no image to capture; its completion still retires in order.
             if (VideoOutExports.ReleaseNoBufferFlip(index, requestId))
             {
@@ -586,7 +892,7 @@ internal static unsafe partial class VulkanVideoPresenter
 
             FlushBatchedGuestCommands();
             RunGuestCacheCollection(endsFrame: true);
-            EnsureGuestSubmissionCapacity();
+            EnsureGuestSubmissionCapacity(collectNow: true);
             long version;
             lock (_gate)
             {

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.CommandThread.cs

@@ -0,0 +1,511 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.VideoOut;
+
+using System.Diagnostics;
+using SharpEmu.HLE;
+using SharpEmu.Libs.Agc;
+using SharpEmu.Libs.Gpu.GpuCommands;
+using SharpEmu.Libs.Gpu.Rendering;
+using SharpEmu.Libs.Gpu.Scheduling;
+
+internal static partial class VulkanVideoPresenter
+{
+    // The command stream is interpreted on its own thread, which hands its work to the render thread
+    // in stream order. Draws, dispatches, barriers, completions and the interpreter's own stores are
+    // queued; a callback that returns a result waits for the render thread to run everything before it.
+    private sealed class CommandThread : AgcExports.IDeferredWorkSink
+    {
+        // The thread runs this far ahead of the render thread at most: enough to keep it busy, and
+        // no more, since a GPU round trip (a readback, a label wait) waits for everything queued.
+        // Measured in Astro Bot: 4096 items halved the frame rate of a crowd scene, 16 starved the
+        // render thread, 32 to 64 kept it busy.
+        private const int MaxOutstanding = 64;
+
+        private enum Kind : byte
+        {
+            Deferred,
+            Call,
+            Write,
+        }
+
+        private sealed class Item
+        {
+            public Kind Kind;
+            public bool IsDraw;
+            public AgcExports.DeferredWork? Work;
+            public Action? Action;
+            public ulong Address;
+            public byte[]? Data;
+            // The guest range a queued fill or copy writes; zero for other items.
+            public ulong RangeSize;
+            // Resolution jobs submitted before this non-draw item; they finish before it runs.
+            public long JobsBefore;
+            public bool Sync;
+            public volatile bool Done;
+            public Exception? Failure;
+        }
+
+        private readonly CommandStreamQueue _queue;
+        private readonly ICpuMemory _memory;
+        private readonly Action<AgcExports.DeferredWork> _runDeferred;
+        private readonly ParallelProgramPrefetch? _prefetch;
+        private readonly Thread _thread;
+        private readonly object _lock = new();
+        private List<Item> _incoming = new();
+        private List<Item> _draining = new();
+        private int _drainIndex;
+        private int _outstanding;
+        // What the producer sleeps on: room in the queue, or its synchronous item.
+        private volatile bool _waitingForSpace;
+        private volatile bool _waitingForSync;
+        private volatile bool _consumerIdle;
+        private volatile bool _stopping;
+        private volatile Exception? _failure;
+        private long _enqueuedNonDraw;
+        private long _executedNonDraw;
+
+        // The interpreter's stores the render thread has not applied yet, oldest first; reads on
+        // this thread see them over guest memory.
+        private readonly object _writeLock = new();
+        private readonly Queue<Item> _pendingWrites = new();
+        private volatile int _pendingWriteCount;
+
+        // The destinations of queued fills and copies; a read on this thread that overlaps one waits
+        // for the render thread to run it.
+        private readonly List<Item> _pendingRanges = new();
+        private volatile int _pendingRangeCount;
+
+        private static readonly long RetryTicks = CommandStreamQueue.AllBlockedRetryMilliseconds * Stopwatch.Frequency / 1000L;
+
+        public CommandThread(CommandStreamQueue queue, ICpuMemory memory, Action<AgcExports.DeferredWork> runDeferred, ParallelProgramPrefetch? prefetch)
+        {
+            _queue = queue;
+            _memory = memory;
+            _runDeferred = runDeferred;
+            _prefetch = prefetch;
+            _thread = new Thread(Run) { IsBackground = true, Name = "SharpEmu command stream" };
+        }
+
+        public void Start() => _thread.Start();
+
+        public bool IsProducerThread => Environment.CurrentManagedThreadId == _thread.ManagedThreadId;
+
+        public long Gate => _enqueuedNonDraw;
+
+        public bool HasWork => Volatile.Read(ref _outstanding) != 0 || _failure is not null;
+
+        // Read by the render loop before it waits, so a queued item wakes it.
+        public bool ConsumerIdle
+        {
+            set => _consumerIdle = value;
+        }
+
+        public bool IsAlive => _thread.IsAlive;
+
+        // The producer side, on the command stream thread.
+
+        public void EnqueueDeferred(AgcExports.DeferredWork work, bool isDraw) =>
+            Enqueue(new Item { Kind = Kind.Deferred, Work = work, IsDraw = isDraw });
+
+        public void Post(Action action) => Enqueue(new Item { Kind = Kind.Call, Action = action });
+
+        public void Run(Action action)
+        {
+            var item = new Item { Kind = Kind.Call, Action = action, Sync = true };
+            Enqueue(item);
+            WaitFor(item);
+        }
+
+        public T Run<T>(Func<T> function)
+        {
+            T result = default!;
+            Run(() => { result = function(); });
+            return result;
+        }
+
+        public void PostWrite(ulong address, ReadOnlySpan<byte> data)
+        {
+            var item = new Item { Kind = Kind.Write, Address = address, Data = data.ToArray() };
+            lock (_writeLock)
+            {
+                _pendingWrites.Enqueue(item);
+                _pendingWriteCount = _pendingWrites.Count;
+            }
+
+            Enqueue(item);
+        }
+
+        // A fill or copy the render thread runs in order; reads of its destination wait for it.
+        public void PostRange(Action action, ulong address, ulong size)
+        {
+            var item = new Item { Kind = Kind.Call, Action = action, Address = address, RangeSize = size };
+            if (size != 0)
+            {
+                lock (_writeLock)
+                {
+                    _pendingRanges.Add(item);
+                    _pendingRangeCount = _pendingRanges.Count;
+                }
+            }
+
+            Enqueue(item);
+        }
+
+        // Runs every queued item when one of them is a fill or copy into this range.
+        public void WaitForPendingRange(ulong address, ulong size)
+        {
+            if (_pendingRangeCount != 0 && OverlapsPendingRange(address, size))
+            {
+                Run(static () => { });
+            }
+        }
+
+        private bool OverlapsPendingRange(ulong address, ulong size)
+        {
+            lock (_writeLock)
+            {
+                foreach (var pending in _pendingRanges)
+                {
+                    if (address < pending.Address + pending.RangeSize && pending.Address < address + size)
+                    {
+                        return true;
+                    }
+                }
+
+                return false;
+            }
+        }
+
+        // Guest memory as the stream sees it: with its own stores applied, queued or not.
+        public bool ReadOverlay(ulong address, Span<byte> destination)
+        {
+            WaitForPendingRange(address, (ulong)destination.Length);
+            if (_pendingWriteCount == 0)
+            {
+                return _memory.TryRead(address, destination);
+            }
+
+            lock (_writeLock)
+            {
+                if (!_memory.TryRead(address, destination))
+                {
+                    return false;
+                }
+
+                var end = address + (ulong)destination.Length;
+                foreach (var write in _pendingWrites)
+                {
+                    var writeEnd = write.Address + (ulong)write.Data!.Length;
+                    if (write.Address >= end || writeEnd <= address)
+                    {
+                        continue;
+                    }
+
+                    var start = Math.Max(address, write.Address);
+                    var stop = Math.Min(end, writeEnd);
+                    write.Data.AsSpan((int)(start - write.Address), (int)(stop - start))
+                        .CopyTo(destination[(int)(start - address)..]);
+                }
+
+                return true;
+            }
+        }
+
+        private void Enqueue(Item item)
+        {
+            ThrowIfFailed();
+            if (!item.IsDraw)
+            {
+                item.JobsBefore = _prefetch?.Submitted ?? 0;
+                _enqueuedNonDraw++;
+            }
+
+            if (Volatile.Read(ref _outstanding) >= MaxOutstanding)
+            {
+                WaitForSpace();
+            }
+
+            lock (_lock)
+            {
+                _incoming.Add(item);
+            }
+
+            Interlocked.Increment(ref _outstanding);
+            if (_consumerIdle || item.Sync)
+            {
+                Presenter.WakeRenderThread();
+            }
+        }
+
+        private void WaitForSpace()
+        {
+            lock (_lock)
+            {
+                // The flag is set before the check: the render thread reads it after it changes what is checked.
+                for (;;)
+                {
+                    _waitingForSpace = true;
+                    Interlocked.MemoryBarrier();
+                    if (Volatile.Read(ref _outstanding) < MaxOutstanding / 2 || _failure is not null)
+                    {
+                        break;
+                    }
+
+                    Monitor.Wait(_lock, 10);
+                }
+
+                _waitingForSpace = false;
+            }
+
+            ThrowIfFailed();
+        }
+
+        private void WaitFor(Item item)
+        {
+            var spinner = new SpinWait();
+            for (var spin = 0; spin < 100 && !item.Done; spin++)
+            {
+                spinner.SpinOnce(sleep1Threshold: -1);
+            }
+
+            if (!item.Done)
+            {
+                lock (_lock)
+                {
+                    for (;;)
+                    {
+                        _waitingForSync = true;
+                        Interlocked.MemoryBarrier();
+                        if (item.Done || _failure is not null)
+                        {
+                            break;
+                        }
+
+                        Monitor.Wait(_lock, 10);
+                    }
+
+                    _waitingForSync = false;
+                }
+            }
+
+            if (item.Failure is { } failure)
+            {
+                throw new InvalidOperationException("A command stream callback failed on the render thread.", failure);
+            }
+
+            if (!item.Done)
+            {
+                ThrowIfFailed();
+            }
+        }
+
+        // The consumer side, on the render thread. Runs queued items until the deadline or until
+        // none is left; returns how many ran.
+        public int RunBatch(long deadline, int maxItems, Action betweenItems)
+        {
+            if (_drainIndex == _draining.Count)
+            {
+                _draining.Clear();
+                _drainIndex = 0;
+                lock (_lock)
+                {
+                    (_incoming, _draining) = (_draining, _incoming);
+                }
+
+                if (_draining.Count == 0)
+                {
+                    ThrowIfFailed();
+                    return 0;
+                }
+            }
+
+            var ran = 0;
+            while (_drainIndex < _draining.Count)
+            {
+                var item = _draining[_drainIndex];
+                _draining[_drainIndex++] = null!;
+                Execute(item);
+                ran++;
+                var outstanding = Interlocked.Decrement(ref _outstanding);
+                // One pulse per wait: the producer sets its flag again if it keeps waiting.
+                var wakeForSync = item.Sync && _waitingForSync;
+                var wakeForSpace = _waitingForSpace && outstanding < MaxOutstanding / 2;
+                if (wakeForSync || wakeForSpace)
+                {
+                    if (wakeForSync)
+                    {
+                        _waitingForSync = false;
+                    }
+
+                    if (wakeForSpace)
+                    {
+                        _waitingForSpace = false;
+                    }
+
+                    lock (_lock)
+                    {
+                        Monitor.PulseAll(_lock);
+                    }
+                }
+
+                betweenItems();
+                if (ran >= maxItems || ((ran & 31) == 0 && Stopwatch.GetTimestamp() >= deadline))
+                {
+                    break;
+                }
+            }
+
+            return ran;
+        }
+
+        private void Execute(Item item)
+        {
+            if (item.IsDraw)
+            {
+                _runDeferred(item.Work!);
+                return;
+            }
+
+            _prefetch?.WaitFinished(item.JobsBefore);
+            try
+            {
+                switch (item.Kind)
+                {
+                    case Kind.Deferred:
+                        _runDeferred(item.Work!);
+                        break;
+                    case Kind.Call:
+                        item.Action!();
+                        break;
+                    case Kind.Write:
+                        ApplyWrite(item);
+                        break;
+                }
+            }
+            catch (Exception exception) when (item.Sync)
+            {
+                item.Failure = exception;
+            }
+            catch (Exception exception)
+            {
+                _failure ??= exception;
+                throw;
+            }
+            finally
+            {
+                _executedNonDraw++;
+                if (item.RangeSize != 0)
+                {
+                    lock (_writeLock)
+                    {
+                        _pendingRanges.Remove(item);
+                        _pendingRangeCount = _pendingRanges.Count;
+                    }
+                }
+
+                _prefetch?.AdvanceGate(_executedNonDraw);
+                if (item.Sync)
+                {
+                    item.Done = true;
+                }
+            }
+        }
+
+        private void ApplyWrite(Item item)
+        {
+            lock (_writeLock)
+            {
+                var written = _memory.TryWrite(item.Address, item.Data);
+                if (!ReferenceEquals(_pendingWrites.Peek(), item))
+                {
+                    throw SubmissionScheduler.Fatal("The command stream stores ran out of order.");
+                }
+
+                _pendingWrites.Dequeue();
+                _pendingWriteCount = _pendingWrites.Count;
+                if (!written)
+                {
+                    throw SubmissionScheduler.Fatal(
+                        $"The command stream cannot write guest memory: address=0x{item.Address:X16} size={item.Data!.Length}.");
+                }
+            }
+        }
+
+        public void ThrowIfFailed()
+        {
+            if (_failure is { } failure)
+            {
+                throw new InvalidOperationException("The command stream thread failed.", failure);
+            }
+        }
+
+        // Ends the thread after its current slice; the render thread keeps running its items
+        // meanwhile, since the slice may wait for one.
+        public void Stop(Action runItems)
+        {
+            _stopping = true;
+            _queue.Wake();
+            var deadline = Stopwatch.GetTimestamp() + 5 * Stopwatch.Frequency;
+            while (_thread.IsAlive && Stopwatch.GetTimestamp() < deadline)
+            {
+                runItems();
+                _thread.Join(1);
+            }
+
+            runItems();
+        }
+
+        private void Run()
+        {
+            try
+            {
+                var lastProgress = Stopwatch.GetTimestamp();
+                while (!_stopping)
+                {
+                    var result = _queue.ProcessOne();
+                    switch (result)
+                    {
+                        case SliceResult.NoWork:
+                            _ = _queue.WaitForWork(8);
+                            break;
+                        case SliceResult.AllBlocked:
+                            WaitBlocked(ref lastProgress);
+                            break;
+                        case SliceResult.BlockedWithoutProgress:
+                            if (!_queue.HasUnblockedPending)
+                            {
+                                WaitBlocked(ref lastProgress);
+                            }
+
+                            break;
+                        default:
+                            lastProgress = Stopwatch.GetTimestamp();
+                            break;
+                    }
+                }
+            }
+            catch (Exception exception)
+            {
+                _failure ??= exception;
+                Presenter.WakeRenderThread();
+            }
+        }
+
+        // Blocked heads retry when a GPU completion or a presentation unblocks them, or every
+        // retry interval without progress.
+        private void WaitBlocked(ref long lastProgress)
+        {
+            var remaining = RetryTicks - (Stopwatch.GetTimestamp() - lastProgress);
+            if (remaining <= 0)
+            {
+                _queue.RetryBlocked();
+                lastProgress = Stopwatch.GetTimestamp();
+                return;
+            }
+
+            _ = _queue.WaitForRetryInterval((int)Math.Max(1, remaining * 1000 / Stopwatch.Frequency));
+        }
+    }
+}

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.ConcurrentResolve.cs

@@ -0,0 +1,134 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.VideoOut;
+
+using SharpEmu.Libs.Gpu.Rendering;
+using SharpEmu.Libs.Gpu.Scheduling;
+
+internal static unsafe partial class VulkanVideoPresenter
+{
+    private sealed partial class Presenter : IConcurrentResolveHost
+    {
+        // This partial serves the guest readers of program resolution on a thread resolving draws
+        // ahead of the render thread. That thread never synchronizes GPU-owned memory: a word the
+        // GPU owns abandons the resolution, which the render thread then does itself. Every
+        // ownership answer it relied on is recorded and checked again when the result is taken.
+
+        private enum ConcurrentQuery : byte
+        {
+            DirtyBytes,
+            CleanWord,
+            Resident,
+            ResidentClean,
+        }
+
+        private sealed class ConcurrentAssumptions(long bufferVersion, long writebackEpoch)
+        {
+            public readonly long BufferVersion = bufferVersion;
+            public readonly long WritebackEpoch = writebackEpoch;
+            public readonly List<(ConcurrentQuery Query, ulong Address, ulong Size, bool Answer)> Answers = new();
+            public bool ReadsImageOwnership;
+
+            public void Record(ConcurrentQuery query, ulong address, ulong size, bool answer)
+            {
+                Answers.Add((query, address, size, answer));
+                ReadsImageOwnership |= query is ConcurrentQuery.CleanWord or ConcurrentQuery.ResidentClean;
+            }
+        }
+
+        // Set only on the resolving thread, around one resolution.
+        [ThreadStatic]
+        private static ConcurrentAssumptions? _concurrentAssumptions;
+
+        void IConcurrentResolveHost.BeginConcurrentResolve() =>
+            _concurrentAssumptions = new ConcurrentAssumptions(_bufferCache.GpuModifiedVersion, GpuWritebackEpoch.Value);
+
+        object IConcurrentResolveHost.EndConcurrentResolve()
+        {
+            var assumptions = _concurrentAssumptions ?? throw new InvalidOperationException("No concurrent resolution is open.");
+            _concurrentAssumptions = null;
+            return assumptions;
+        }
+
+        // GPU data stored into guest memory since the resolution may have changed what it read;
+        // otherwise the answers stand while the ownership they came from has not changed.
+        bool IConcurrentResolveHost.IsStillCurrent(object assumptionsObject)
+        {
+            var assumptions = (ConcurrentAssumptions)assumptionsObject;
+            if (assumptions.WritebackEpoch != GpuWritebackEpoch.Value)
+            {
+                return false;
+            }
+
+            if (assumptions.BufferVersion == _bufferCache.GpuModifiedVersion && !assumptions.ReadsImageOwnership)
+            {
+                return true;
+            }
+
+            foreach (var (query, address, size, answer) in assumptions.Answers)
+            {
+                var now = query switch
+                {
+                    ConcurrentQuery.DirtyBytes => _bufferCache.HasGpuDirtyBytes(address, size),
+                    ConcurrentQuery.CleanWord => IsGpuOwnedForCleanRead(address, size),
+                    ConcurrentQuery.Resident => IsGpuOwnedForResidentRead(address, size, clean: false),
+                    _ => IsGpuOwnedForResidentRead(address, size, clean: true),
+                };
+                if (now != answer)
+                {
+                    return false;
+                }
+            }
+
+            return true;
+        }
+
+        private bool IsGpuOwnedForCleanRead(ulong address, ulong size) =>
+            (_bufferCache.MayHaveGpuDirtyPages(address, size) && _bufferCache.HasGpuDirtyPages(address, size)) ||
+            _bufferCache.HasGpuDirtyBytes(address, size) ||
+            _imageCache.HasGpuModifiedImageBytes(address, size);
+
+        private bool IsGpuOwnedForResidentRead(ulong address, ulong size, bool clean) =>
+            _bufferCache.HasGpuDirtyBytes(address, size) ||
+            (clean && (_bufferCache.HasGpuDirtyPages(address, size) || _imageCache.HasGpuModifiedImageBytes(address, size)));
+
+        private bool TryReadGuestWordConcurrent(ulong address, out uint word, ConcurrentAssumptions assumptions)
+        {
+            word = 0;
+            if (!_guestMemory.CanRead(address, sizeof(uint)))
+            {
+                return false;
+            }
+
+            var dirty = _bufferCache.HasGpuDirtyBytes(address, sizeof(uint));
+            assumptions.Record(ConcurrentQuery.DirtyBytes, address, sizeof(uint), dirty);
+            if (dirty)
+            {
+                throw new SubmissionScheduler.AbandonedWorkException($"The resolution reads GPU-owned memory: address=0x{address:X16}.");
+            }
+
+            Span<byte> bytes = stackalloc byte[sizeof(uint)];
+            if (!_guestMemory.TryRead(address, bytes))
+            {
+                return false;
+            }
+
+            word = System.Buffers.Binary.BinaryPrimitives.ReadUInt32LittleEndian(bytes);
+            return true;
+        }
+
+        private bool TryReadResidentGuestBytesConcurrent(ulong address, Span<byte> destination, bool clean, ConcurrentAssumptions assumptions)
+        {
+            var size = (ulong)destination.Length;
+            if (!_guestMemory.CanRead(address, size))
+            {
+                return false;
+            }
+
+            var owned = IsGpuOwnedForResidentRead(address, size, clean);
+            assumptions.Record(clean ? ConcurrentQuery.ResidentClean : ConcurrentQuery.Resident, address, size, owned);
+            return !owned && _guestMemory.TryRead(address, destination);
+        }
+    }
+}

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Descriptors.cs

@@ -111,21 +111,81 @@ internal static unsafe partial class VulkanVideoPresenter
             Atomic: image.Atomic);
 
         private static readonly System.Collections.Concurrent.ConcurrentDictionary<string, byte> _tracedTextureBindings = new();
+        // The last lookup of one image of one program. A draw binds the same descriptor as the
+        // previous draw of its program most of the time; while no image the lookup could find has
+        // changed (the cache's lookup generation), the request and the image found are the same.
+        private struct ImageBindingMemo
+        {
+            public uint[]? Words;
+            public ulong Generation;
+            public TextureRequestResolution Resolution;
+            public ResourceSlotIdentifier Found;
+
+            public readonly bool Matches(uint[] words, ulong generation) =>
+                Words is not null && Generation == generation && words.AsSpan().SequenceEqual(Words);
+        }
+
+        private Dictionary<ShaderProgramInfo, ImageBindingMemo[]>? _imageBindingMemos;
+        private ShaderProgramInfo? _lastMemoProgram;
+        private ImageBindingMemo[] _lastMemos = [];
+
+        // A draw's images resolve one after another for one program, so the last program's memos
+        // are kept at hand.
+        private ref ImageBindingMemo ImageBindingMemoOf(ShaderProgramInfo program, int index)
+        {
+            if (!ReferenceEquals(program, _lastMemoProgram) || index >= _lastMemos.Length)
+            {
+                var all = _imageBindingMemos ??= new Dictionary<ShaderProgramInfo, ImageBindingMemo[]>(ReferenceEqualityComparer.Instance);
+                if (!all.TryGetValue(program, out var memos) || index >= memos.Length)
+                {
+                    Array.Resize(ref memos, Math.Max(index + 1, program.Resources?.Info.Images.Count ?? 0));
+                    all[program] = memos;
+                }
+
+                _lastMemoProgram = program;
+                _lastMemos = memos;
+            }
+
+            return ref _lastMemos[index];
+        }
 
         // Render-state discovery for one shader image; the view is acquired later with the draw.
         private TextureResource ResolveImageBinding(ImageResource image, uint[] words, ShaderProgramInfo program, int index)
         {
+            ref var memo = ref ImageBindingMemoOf(program, index);
             if (words.Length < 4)
             {
                 throw SubmissionScheduler.Fatal($"An image descriptor is too short: image={index} words={words.Length} hash=0x{program.Hash:X16}.");
             }
 
             var storage = image.ResourceClass == ShaderCompiler.Resources.ImageResourceClass.Storage;
-            var resolution = ImageRequestBuilders.Texture(words, ShapeOf(image));
             _ = BeginBatchedGuestCommands();
-            var request = resolution.Request;
-            var imageIdentifier = _imageCache.FindImage(ref request, resolution.ExactFormat);
-            resolution = resolution with { Request = request };
+            TextureRequestResolution resolution;
+            ResourceSlotIdentifier imageIdentifier;
+            if (memo.Matches(words, _imageCache.LookupGeneration) && _imageCache.TryTouchRemembered(memo.Found))
+            {
+                resolution = memo.Resolution;
+                imageIdentifier = memo.Found;
+            }
+            else
+            {
+                resolution = ImageRequestBuilders.Texture(words, ShapeOf(image));
+                var request = resolution.Request;
+                imageIdentifier = _imageCache.FindImage(ref request, resolution.ExactFormat);
+                resolution = resolution with { Request = request };
+                // An empty range or DCC metadata needs the full lookup on every draw.
+                memo = !ImageDescription.IsEmptyRange(request.Description.Data) &&
+                    _imageCache.GetImage(imageIdentifier).Description.DccSliceSize == 0
+                    ? new ImageBindingMemo
+                    {
+                        Words = (uint[])words.Clone(),
+                        Generation = _imageCache.LookupGeneration,
+                        Resolution = resolution,
+                        Found = imageIdentifier,
+                    }
+                    : default;
+            }
+
             imageIdentifier = ImageRequestBuilders.ValidateTextureOwner(_imageCache, imageIdentifier, resolution);
             BindImage(imageIdentifier, storage);
             var descriptor = new TextureDescriptorWords(words);
@@ -151,16 +211,47 @@ internal static unsafe partial class VulkanVideoPresenter
                     Console.Error.WriteLine(line);
                 }
             }
-            return new TextureResource
+            var texture = RentTextureResource();
+            texture.Address = descriptor.BaseAddress;
+            texture.ImageIdentifier = imageIdentifier;
+            texture.Request = resolution.Request;
+            texture.IsStorage = storage;
+            texture.DestinationSelect = words[3] & 0xFFFu;
+            texture.Width = descriptor.Width;
+            texture.Height = descriptor.Height;
+            return texture;
+        }
+
+        // Binding descriptions are large (an image request carries its whole mip layout) and one is made
+        // per shader image per draw, so they are reused once the submission that read them has retired.
+        // Only the render thread rents and returns them.
+        // Sized for the bindings of every draw still in flight (a few frames of a thousand draws each);
+        // a smaller cap drops returns during busy scenes and the pool allocates again.
+        private const int MaxPooledTextureResources = 32768;
+        private Stack<TextureResource>? _texturePool;
+
+        private TextureResource RentTextureResource()
+        {
+            if (_texturePool is not null && _texturePool.TryPop(out var texture))
             {
-                Address = descriptor.BaseAddress,
-                ImageIdentifier = imageIdentifier,
-                Request = request,
-                IsStorage = storage,
-                DestinationSelect = words[3] & 0xFFFu,
-                Width = descriptor.Width,
-                Height = descriptor.Height,
-            };
+                texture.InPool = false;
+                return texture;
+            }
+
+            return new TextureResource { Pooled = true };
+        }
+
+        private void ReturnTextureResource(TextureResource texture)
+        {
+            _texturePool ??= new Stack<TextureResource>();
+            if (!texture.Pooled || texture.InPool || _texturePool.Count >= MaxPooledTextureResources)
+            {
+                return;
+            }
+
+            texture.Reset();
+            texture.InPool = true;
+            _texturePool.Push(texture);
         }
 
         // Compare bits stay only on depth-compare samplers; a forced point sampler drops its filters.
@@ -768,8 +859,11 @@ internal static unsafe partial class VulkanVideoPresenter
                     throw SubmissionScheduler.Fatal($"A stage does not belong to the bind point: stage={stage.Program.Stage} bindPoint={bindPoint}.");
                 }
 
-                foreach (var binding in stage.Layout.Descriptors)
+                // Indexed loops over the layout's lists: a foreach on IReadOnlyList allocates an enumerator per draw.
+                var countedBindings = stage.Layout.Descriptors;
+                for (var bindingIndex = 0; bindingIndex < countedBindings.Count; bindingIndex++)
                 {
+                    var binding = countedBindings[bindingIndex];
                     writeCount++;
                     var count = (int)DescriptorWriter.DescriptorCount(binding);
                     if (ImageDescriptorBinding.ResourceClass(binding.Kind) != ShaderCompiler.Resources.ImageResourceClass.None || binding.Kind == DescriptorBindingKind.Samplers)
@@ -862,8 +956,11 @@ internal static unsafe partial class VulkanVideoPresenter
 
                     var occurrences = occurrenceScratch[..descriptors.Images.Length];
                     occurrences.Clear();
-                    foreach (var binding in stage.Layout.Descriptors)
+                    var bindings = stage.Layout.Descriptors;
+                    for (var bindingIndex = 0; bindingIndex < bindings.Count; bindingIndex++)
                     {
+                        var binding = bindings[bindingIndex];
+                        var resources = binding.Resources;
                         var write = new WriteDescriptorSet
                         {
                             SType = StructureType.WriteDescriptorSet,
@@ -875,8 +972,9 @@ internal static unsafe partial class VulkanVideoPresenter
                         var imageStart = imageIndex;
                         if (ImageDescriptorBinding.ResourceClass(binding.Kind) != ShaderCompiler.Resources.ImageResourceClass.None)
                         {
-                            foreach (var resource in binding.Resources)
+                            for (var resourceIndex = 0; resourceIndex < resources.Count; resourceIndex++)
                             {
+                                var resource = resources[resourceIndex];
                                 imageInfos[imageIndex++] = ImageInfo(descriptors.Images[(int)resource], occurrences[(int)resource]++, program, (int)resource);
                             }
                         }
@@ -885,8 +983,9 @@ internal static unsafe partial class VulkanVideoPresenter
                             switch (binding.Kind)
                             {
                                 case DescriptorBindingKind.Buffers:
-                                    foreach (var resource in binding.Resources)
+                                    for (var resourceIndex = 0; resourceIndex < resources.Count; resourceIndex++)
                                     {
+                                        var resource = resources[resourceIndex];
                                         var view = descriptors.Buffers[(int)resource];
                                         if (view.Buffer.Handle == 0)
                                         {
@@ -925,8 +1024,9 @@ internal static unsafe partial class VulkanVideoPresenter
                                 }
 
                                 case DescriptorBindingKind.Samplers:
-                                    foreach (var resource in binding.Resources)
+                                    for (var resourceIndex = 0; resourceIndex < resources.Count; resourceIndex++)
                                     {
+                                        var resource = resources[resourceIndex];
                                         var sampler = descriptors.Samplers[(int)resource];
                                         if (sampler.Handle == 0)
                                         {

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Device.Setup.cs

@@ -613,6 +613,10 @@ internal static unsafe partial class VulkanVideoPresenter
                 subgroupSizeControl.MinSubgroupSize <= RdnaSubgroupSize &&
                 subgroupSizeControl.MaxSubgroupSize >= RdnaSubgroupSize &&
                 (subgroupSizeControl.RequiredSubgroupSizeStages & ShaderStageFlags.ComputeBit) != 0;
+            _canRequireComputeSubgroup64 =
+                subgroupSizeControl.MinSubgroupSize <= Wave64SubgroupSize &&
+                subgroupSizeControl.MaxSubgroupSize >= Wave64SubgroupSize &&
+                (subgroupSizeControl.RequiredSubgroupSizeStages & ShaderStageFlags.ComputeBit) != 0;
             _maxComputeWorkgroupSubgroups = subgroupSizeControl.MaxComputeWorkgroupSubgroups;
             // MoltenVK reports a zero OpArrayLength for storage buffers bound through push
             // descriptors, which turns every bounds-checked load into zero and drops every store.
@@ -945,6 +949,7 @@ internal static unsafe partial class VulkanVideoPresenter
             var supportsNullDescriptor = robustness2Features.NullDescriptor;
             var supportsRobustness2 = supportsRobustImageAccess2 || supportsNullDescriptor;
             _canRequireComputeSubgroup32 &= vulkan13Features.SubgroupSizeControl;
+            _canRequireComputeSubgroup64 &= vulkan13Features.SubgroupSizeControl;
             SetSharedInt64AtomicsCapability(supportsSharedInt64Atomics);
             if (!supportsSharedInt64Atomics)
             {
@@ -1157,7 +1162,7 @@ internal static unsafe partial class VulkanVideoPresenter
                     SType = StructureType.PhysicalDeviceVulkan13Features,
                     DynamicRendering = true,
                     Synchronization2 = true,
-                    SubgroupSizeControl = _canRequireComputeSubgroup32,
+                    SubgroupSizeControl = _canRequireComputeSubgroup32 || _canRequireComputeSubgroup64,
                     PNext = renderingChain,
                 };
                 var features2 = new PhysicalDeviceFeatures2

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Draws.Recording.cs

@@ -275,8 +275,12 @@ internal static unsafe partial class VulkanVideoPresenter
                 {
                     RecycleHostBuffer(texture.StagingBuffer, texture.StagingMemory);
                 }
+
+                ReturnTextureResource(texture);
             }
 
+            resources.Textures = [];
+
             foreach (var snapshot in resources.FeedbackSnapshots)
             {
                 RetireFeedbackSnapshot(snapshot);

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.GuestSubmissions.cs

@@ -108,9 +108,23 @@ internal static unsafe partial class VulkanVideoPresenter
                     retireBuffers));
         }
 
-        private void EnsureGuestSubmissionCapacity()
+        // Collecting completed submissions polls the GPU timeline twice and may wait for the priority
+        // worker. Every draw used to pay that; completed work only frees resources and runs deferred
+        // callbacks, so a collection a few hundred microseconds later changes nothing but how long
+        // those resources stay reserved. A full in-flight queue still collects, and waits, at once.
+        private static readonly long CollectIntervalTicks = System.Diagnostics.Stopwatch.Frequency / 4000;
+        private long _lastCollectTicks;
+
+        private void EnsureGuestSubmissionCapacity(bool collectNow = false)
         {
             using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.SubmissionCapacity);
+            var now = System.Diagnostics.Stopwatch.GetTimestamp();
+            if (!collectNow && _pendingGuestSubmissions.Count < MaxInFlightGuestSubmissions && now - _lastCollectTicks < CollectIntervalTicks)
+            {
+                return;
+            }
+
+            _lastCollectTicks = now;
             CollectCompletedGuestSubmissions(waitForOldest: false);
             if (_pendingGuestSubmissions.Count >= MaxInFlightGuestSubmissions)
             {

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Pipelines.cs

@@ -67,23 +67,43 @@ internal static unsafe partial class VulkanVideoPresenter
         uint IShaderPipelineHost.MaxPushDescriptors => _maxPushDescriptors;
 
         // Wave64 compute runs natively when the device's subgroup is that wide; smaller devices emulate it.
+        bool IShaderPipelineHost.ComputeWave64SubgroupNative =>
+            Volatile.Read(ref _nativeSubgroupSize) >= 64 && (_canRequireComputeSubgroup64 || Volatile.Read(ref _nativeSubgroupSize) == 64);
+
         bool IShaderPipelineHost.ComputeWave64Supported => Volatile.Read(ref _nativeSubgroupSize) >= 64;
 
-        // Only a 64-invocation wave64 workgroup is translated for either host subgroup width. Every
-        // other compute translation maps a guest wave to 32-lane host subgroups, and a 64-lane host
-        // subgroup (AMD's default) left lanes 32..63 inactive: a wave64 8x8x8 group lost rows 4..7.
+        // Wave32 compute maps a guest wave to 32-lane host subgroups, and a 64-lane host subgroup
+        // (AMD's default) left lanes 32..63 inactive, so those pipelines are pinned to 32. Wave64
+        // compute is pinned to 64 instead when the host runs it natively.
         private const uint RdnaSubgroupSize = 32;
         private bool _canRequireComputeSubgroup32;
         private uint _maxComputeWorkgroupSubgroups;
 
+        private const uint Wave64SubgroupSize = 64;
+        private bool _canRequireComputeSubgroup64;
+
+        // A native wave64 translation reads one guest wave as one host subgroup, so the driver must
+        // not pick 32-lane subgroups for it (RDNA drivers may do so for compute).
+        private bool RequiresComputeSubgroup64(ComputeInputInfo input)
+        {
+            return _canRequireComputeSubgroup64 && input.WaveSize == 64 &&
+                   ((IShaderPipelineHost)this).ComputeWave64SubgroupNative;
+        }
+
         private bool RequiresComputeSubgroup32(ComputeInputInfo input)
         {
             var invocations = (ulong)Math.Max(input.ThreadsX, 1) * Math.Max(input.ThreadsY, 1) * Math.Max(input.ThreadsZ, 1);
             return _canRequireComputeSubgroup32 &&
-                   !(input.WaveSize == 64 && invocations == 64) &&
+                   !(input.WaveSize == 64 && ((IShaderPipelineHost)this).ComputeWave64SubgroupNative) &&
                    invocations <= (ulong)_maxComputeWorkgroupSubgroups * RdnaSubgroupSize;
         }
 
+        // The subgroup width a compute pipeline must be pinned to, or 0 to leave it to the driver.
+        private uint ComputeRequiredSubgroupSize(ComputeInputInfo input) =>
+            RequiresComputeSubgroup64(input) ? Wave64SubgroupSize
+            : RequiresComputeSubgroup32(input) ? RdnaSubgroupSize
+            : 0u;
+
         bool IShaderPipelineHost.GraphicsSubgroupOperationsEnabled => GraphicsSubgroupOperationsEnabled;
 
         bool IShaderPipelineHost.SharedInt64AtomicsEnabled => SharedInt64AtomicsEnabled;
@@ -101,6 +121,11 @@ internal static unsafe partial class VulkanVideoPresenter
         // A range the GPU wrote is downloaded first, so the word is what the guest CPU would read.
         public bool TryReadGuestWord(ulong address, out uint word)
         {
+            if (_concurrentAssumptions is { } assumptions)
+            {
+                return TryReadGuestWordConcurrent(address, out word, assumptions);
+            }
+
             using var profile = ResourceMaterializationProfile.Measure(ResourceMaterializationProfile.Phase.GuestRead);
             word = 0;
             var synchronized = false;
@@ -170,9 +195,9 @@ internal static unsafe partial class VulkanVideoPresenter
                 return false;
             }
 
-            if ((_bufferCache.MayHaveGpuDirtyPages(address, sizeof(uint)) && _bufferCache.HasGpuDirtyPages(address, sizeof(uint))) ||
-                _bufferCache.HasGpuDirtyBytes(address, sizeof(uint)) ||
-                _imageCache.HasGpuModifiedImageBytes(address, sizeof(uint)))
+            var owned = IsGpuOwnedForCleanRead(address, sizeof(uint));
+            _concurrentAssumptions?.Record(ConcurrentQuery.CleanWord, address, sizeof(uint), owned);
+            if (owned)
             {
                 return false;
             }
@@ -190,6 +215,11 @@ internal static unsafe partial class VulkanVideoPresenter
         // The same ownership rules as the two word readers, checked once for a whole range.
         public bool TryReadResidentGuestBytes(ulong address, Span<byte> destination, bool clean)
         {
+            if (_concurrentAssumptions is { } assumptions)
+            {
+                return TryReadResidentGuestBytesConcurrent(address, destination, clean, assumptions);
+            }
+
             var size = (ulong)destination.Length;
             if (clean || !IsCleanReadPage(address, size))
             {
@@ -967,15 +997,16 @@ internal static unsafe partial class VulkanVideoPresenter
             Pipeline pipeline;
             try
             {
+                var pin = ComputeRequiredSubgroupSize(description.Input);
                 var requiredSubgroupSize = new PipelineShaderStageRequiredSubgroupSizeCreateInfo
                 {
                     SType = StructureType.PipelineShaderStageRequiredSubgroupSizeCreateInfo,
-                    RequiredSubgroupSize = RdnaSubgroupSize,
+                    RequiredSubgroupSize = pin,
                 };
                 var stageInfo = new PipelineShaderStageCreateInfo
                 {
                     SType = StructureType.PipelineShaderStageCreateInfo,
-                    PNext = RequiresComputeSubgroup32(description.Input) ? &requiredSubgroupSize : null,
+                    PNext = pin != 0 ? &requiredSubgroupSize : null,
                     Stage = ShaderStageFlags.ComputeBit,
                     Module = computeModule,
                     PName = entryPoint,
@@ -1099,6 +1130,7 @@ internal static unsafe partial class VulkanVideoPresenter
             var device = _device;
             var cacheSource = GetGuestPipelineCacheSource(ComputeCacheKey(description.Stage.Hash, computeModule.Handle));
             var vk = _vk;
+            var requiredSubgroupSize = ComputeRequiredSubgroupSize(description.Input);
             var started = new PendingComputePipeline
             {
                 SetLayout = setLayout,
@@ -1122,7 +1154,7 @@ internal static unsafe partial class VulkanVideoPresenter
                         // Importing a MoltenVK cache compiles its MSL libraries.
                         // Keep that work inside the same bounded compiler slot.
                         var cache = ResolveGuestPipelineCache(cacheSource);
-                        return CompileComputePipeline(vk, device, cache, computeModule, layout);
+                        return CompileComputePipeline(vk, device, cache, computeModule, layout, requiredSubgroupSize);
                     }
                     finally
                     {
@@ -1146,14 +1178,21 @@ internal static unsafe partial class VulkanVideoPresenter
             Device device,
             PipelineCache cache,
             ShaderModule module,
-            PipelineLayout layout)
+            PipelineLayout layout,
+            uint requiredSubgroupSize)
         {
             var entryPoint = (byte*)SilkMarshal.StringToPtr("main");
             try
             {
+                var subgroupInfo = new PipelineShaderStageRequiredSubgroupSizeCreateInfo
+                {
+                    SType = StructureType.PipelineShaderStageRequiredSubgroupSizeCreateInfo,
+                    RequiredSubgroupSize = requiredSubgroupSize,
+                };
                 var stageInfo = new PipelineShaderStageCreateInfo
                 {
                     SType = StructureType.PipelineShaderStageCreateInfo,
+                    PNext = requiredSubgroupSize != 0 ? &subgroupInfo : null,
                     Stage = ShaderStageFlags.ComputeBit,
                     Module = module,
                     PName = entryPoint,

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Present.Hdr.cs

@@ -357,6 +357,7 @@ internal static unsafe partial class VulkanVideoPresenter
                 _commandBuffer,
                 PipelineBindPoint.Graphics,
                 isHdr ? _hdrPqPipeline : _hdrPipeline);
+            SharpEmu.Libs.Gpu.Vulkan.GraphicsDynamicStateEpoch.Advance();
             var descriptorSet = isHdr
                 ? _hdrPqDescriptorSets[imageIndex]
                 : _hdrDescriptorSets[imageIndex];

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.RenderLoop.cs

@@ -46,38 +46,50 @@ internal static unsafe partial class VulkanVideoPresenter
             using var profileScope = RenderPhaseProfile.Measure(RenderPhaseProfile.Phase.Idle);
             var gpuWorkInFlight = _pendingGuestSubmissions.Count > 0 ||
                 Array.Exists(_frameInFlight, static pending => pending);
-            var blockedRetryWait = BlockedRetryWaitMilliseconds();
-            lock (_gate)
+            // The command stream thread retries its blocked heads itself and wakes this thread for its
+            // items; before it starts, queued submissions start it on the next tick.
+            var commandThread = _commandThread;
+            if (commandThread is not null)
             {
-                if (_closed ||
-                    Volatile.Read(ref _presenterCloseRequested) ||
-                    _relay.HasPendingCommands ||
-                    (_vulkanReady && _commandStream.HasUnblockedPending) ||
-                    blockedRetryWait == 0 ||
-                    HasReadyPresentationLocked())
-                {
-                    return;
-                }
+                commandThread.ConsumerIdle = true;
+            }
 
-                var waitMilliseconds = gpuWorkInFlight ? 1 : 8;
-                if (blockedRetryWait is { } retryWait)
+            try
+            {
+                lock (_gate)
                 {
-                    waitMilliseconds = Math.Min(waitMilliseconds, retryWait);
-                }
+                    if (_closed ||
+                        Volatile.Read(ref _presenterCloseRequested) ||
+                        _relay.HasPendingCommands ||
+                        (commandThread is not null ? commandThread.HasWork : _vulkanReady && _commandStream.HasPending) ||
+                        HasReadyPresentationLocked())
+                    {
+                        return;
+                    }
 
-                var waitPhase = _commandStream.HasPending
-                    ? RenderPhaseProfile.Phase.IdleBlockedCommands
-                    : _pendingGuestImagePresentations.Count > 0 || _pendingVideoPresentations.Count > 0 ||
-                        (_latestPresentation is { } latest && latest.Sequence != _presentedSequence)
-                        ? RenderPhaseProfile.Phase.IdlePendingPresentation
-                        : RenderPhaseProfile.Phase.IdleNoQueuedWork;
-                using var waitProfile = RenderPhaseProfile.MeasureDetail(waitPhase);
-                SubmissionFlowProfile.Record(SubmissionFlowProfile.EventKind.WaitStarted,
-                    detail: waitMilliseconds);
-                var signaled = System.Threading.Monitor.Wait(_gate, waitMilliseconds);
-                // The result describes the monitor wait, not the arrival of runnable guest work.
-                SubmissionFlowProfile.Record(signaled ? SubmissionFlowProfile.EventKind.WaitSignaled
-                    : SubmissionFlowProfile.EventKind.WaitTimedOut, detail: (int)waitPhase);
+                    var waitMilliseconds = gpuWorkInFlight ? 1 : 8;
+
+                    var waitPhase = _commandStream.HasPending
+                        ? RenderPhaseProfile.Phase.IdleBlockedCommands
+                        : _pendingGuestImagePresentations.Count > 0 || _pendingVideoPresentations.Count > 0 ||
+                            (_latestPresentation is { } latest && latest.Sequence != _presentedSequence)
+                            ? RenderPhaseProfile.Phase.IdlePendingPresentation
+                            : RenderPhaseProfile.Phase.IdleNoQueuedWork;
+                    using var waitProfile = RenderPhaseProfile.MeasureDetail(waitPhase);
+                    SubmissionFlowProfile.Record(SubmissionFlowProfile.EventKind.WaitStarted,
+                        detail: waitMilliseconds);
+                    var signaled = System.Threading.Monitor.Wait(_gate, waitMilliseconds);
+                    // The result describes the monitor wait, not the arrival of runnable guest work.
+                    SubmissionFlowProfile.Record(signaled ? SubmissionFlowProfile.EventKind.WaitSignaled
+                        : SubmissionFlowProfile.EventKind.WaitTimedOut, detail: (int)waitPhase);
+                }
+            }
+            finally
+            {
+                if (commandThread is not null)
+                {
+                    commandThread.ConsumerIdle = false;
+                }
             }
         }
 
@@ -117,9 +129,16 @@ internal static unsafe partial class VulkanVideoPresenter
                 return;
             }
 
-            using (RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.QueueRelay))
+            if (_commandThread is not null)
             {
-                _relay.RunPendingCommands();
+                RunRelayCommandsWithPrefetchHeld();
+            }
+            else
+            {
+                using (RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.QueueRelay))
+                {
+                    _relay.RunPendingCommands();
+                }
             }
             if (_deviceLost)
             {
@@ -388,6 +407,7 @@ internal static unsafe partial class VulkanVideoPresenter
                     _commandBuffer,
                     PipelineBindPoint.Graphics,
                     _barycentricPipeline);
+                SharpEmu.Libs.Gpu.Vulkan.GraphicsDynamicStateEpoch.Advance();
                 _vk.CmdDraw(_commandBuffer, 3, 1, 0, 0);
                 _vk.CmdEndRenderPass(_commandBuffer);
                 waitStage = PipelineStageFlags.ColorAttachmentOutputBit;

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Rendering.cs

@@ -12,6 +12,7 @@ using SharpEmu.Libs.Gpu.GpuCommands.Registers;
 using SharpEmu.Libs.Gpu.Images;
 using SharpEmu.Libs.Gpu.Pipelines;
 using SharpEmu.Libs.Gpu.Rendering;
+using SharpEmu.Libs.Gpu.Vulkan;
 using SharpEmu.Libs.Gpu.Scheduling;
 using SharpEmu.Libs.Gpu.Vulkan;
 using Silk.NET.Core;
@@ -280,14 +281,21 @@ internal static unsafe partial class VulkanVideoPresenter
             _boundGraphicsPipeline = null;
         }
 
+        // The acquisition below reads the attachment back into its result and keeps no reference to it,
+        // so one render-thread instance serves every draw.
+        private ColorAttachment? _colorAttachmentScratch;
+
         ColorAttachmentAcquisition IRenderHost.AcquireColorAttachment(in ColorTargetState target)
         {
-            var attachment = new ColorAttachment
-            {
-                ImageIdentifier = target.Image,
-                Request = target.Resolution.Request,
-                Resolution = target.Resolution,
-            };
+            var attachment = _colorAttachmentScratch ??= new ColorAttachment();
+            attachment.ImageIdentifier = target.Image;
+            attachment.Request = target.Resolution.Request;
+            attachment.Resolution = target.Resolution;
+            attachment.Image = null!;
+            attachment.View = default;
+            attachment.Layout = default;
+            attachment.Clear = false;
+            attachment.ClearValue = default;
             AcquireColorAttachment(attachment);
             return new ColorAttachmentAcquisition(
                 attachment.ImageIdentifier,
@@ -644,29 +652,80 @@ internal static unsafe partial class VulkanVideoPresenter
             _vk.CmdClearDepthStencilImage(command, image.Backing.Handle, ImageLayout.TransferDstOptimal, &value, 1, &vkRange);
         }
 
+        // The dynamic state the last guest draw set, and where it holds: the command buffer and its
+        // tick (a new command buffer starts without state), and no other graphics pipeline since.
+        private DynamicDrawState _lastDynamicState;
+        private CommandBuffer _lastDynamicStateBuffer;
+        private ulong _lastDynamicStateTick;
+        private long _lastDynamicStateEpoch = -1;
+
         public void SetDynamicState(in DynamicDrawState state)
         {
             using var profileScope = RenderPhaseProfile.MeasureDetail(RenderPhaseProfile.Phase.DrawDynamicStateRecording);
             var command = BeginBatchedGuestCommands();
-            var viewport = new Viewport(state.ViewportX, state.ViewportY, state.ViewportWidth, state.ViewportHeight, state.ViewportMinDepth, state.ViewportMaxDepth);
-            _vk.CmdSetViewport(command, 0, 1, &viewport);
-            var scissor = new Rect2D(
-                new Offset2D(state.Scissor.Left, state.Scissor.Top),
-                new Extent2D((uint)(state.Scissor.Right - state.Scissor.Left), (uint)(state.Scissor.Bottom - state.Scissor.Top)));
-            _vk.CmdSetScissor(command, 0, 1, &scissor);
-            _vk.CmdSetLineWidth(command, state.LineWidth);
-            var blendConstants = stackalloc float[4] { state.BlendRed, state.BlendGreen, state.BlendBlue, state.BlendAlpha };
-            _vk.CmdSetBlendConstants(command, blendConstants);
-            _vk.CmdSetDepthTestEnable(command, state.DepthTestEnabled);
-            _vk.CmdSetDepthWriteEnable(command, state.DepthWriteEnabled);
-            _vk.CmdSetDepthCompareOp(command, state.DepthCompare);
-            _vk.CmdSetDepthBiasEnable(command, state.DepthBiasEnabled);
-            if (state.DepthBiasEnabled)
+            // Every guest pipeline keeps these states dynamic, so they persist across its binds;
+            // only what changed since the last draw in this command buffer is set again.
+            var known = command.Handle == _lastDynamicStateBuffer.Handle && _scheduler.CurrentTick == _lastDynamicStateTick &&
+                GraphicsDynamicStateEpoch.Value == _lastDynamicStateEpoch;
+            ref readonly var last = ref _lastDynamicState;
+            if (!known || state.ViewportX != last.ViewportX || state.ViewportY != last.ViewportY ||
+                state.ViewportWidth != last.ViewportWidth || state.ViewportHeight != last.ViewportHeight ||
+                state.ViewportMinDepth != last.ViewportMinDepth || state.ViewportMaxDepth != last.ViewportMaxDepth)
+            {
+                var viewport = new Viewport(state.ViewportX, state.ViewportY, state.ViewportWidth, state.ViewportHeight, state.ViewportMinDepth, state.ViewportMaxDepth);
+                _vk.CmdSetViewport(command, 0, 1, &viewport);
+            }
+
+            if (!known || state.Scissor != last.Scissor)
+            {
+                var scissor = new Rect2D(
+                    new Offset2D(state.Scissor.Left, state.Scissor.Top),
+                    new Extent2D((uint)(state.Scissor.Right - state.Scissor.Left), (uint)(state.Scissor.Bottom - state.Scissor.Top)));
+                _vk.CmdSetScissor(command, 0, 1, &scissor);
+            }
+
+            if (!known || state.LineWidth != last.LineWidth)
+            {
+                _vk.CmdSetLineWidth(command, state.LineWidth);
+            }
+
+            if (!known || state.BlendRed != last.BlendRed || state.BlendGreen != last.BlendGreen ||
+                state.BlendBlue != last.BlendBlue || state.BlendAlpha != last.BlendAlpha)
+            {
+                var blendConstants = stackalloc float[4] { state.BlendRed, state.BlendGreen, state.BlendBlue, state.BlendAlpha };
+                _vk.CmdSetBlendConstants(command, blendConstants);
+            }
+
+            if (!known || state.DepthTestEnabled != last.DepthTestEnabled)
+            {
+                _vk.CmdSetDepthTestEnable(command, state.DepthTestEnabled);
+            }
+
+            if (!known || state.DepthWriteEnabled != last.DepthWriteEnabled)
+            {
+                _vk.CmdSetDepthWriteEnable(command, state.DepthWriteEnabled);
+            }
+
+            if (!known || state.DepthCompare != last.DepthCompare)
+            {
+                _vk.CmdSetDepthCompareOp(command, state.DepthCompare);
+            }
+
+            if (!known || state.DepthBiasEnabled != last.DepthBiasEnabled)
+            {
+                _vk.CmdSetDepthBiasEnable(command, state.DepthBiasEnabled);
+            }
+
+            // The factors only matter while the bias is on; they are set when it turns on or changes.
+            if (state.DepthBiasEnabled && (!known || !last.DepthBiasEnabled ||
+                state.DepthBiasConstantFactor != last.DepthBiasConstantFactor || state.DepthBiasClamp != last.DepthBiasClamp ||
+                state.DepthBiasSlopeFactor != last.DepthBiasSlopeFactor))
             {
                 _vk.CmdSetDepthBias(command, state.DepthBiasConstantFactor, _supportsDepthBiasClamp ? state.DepthBiasClamp : 0f, state.DepthBiasSlopeFactor);
             }
 
-            if (state.StencilTestEnabled)
+            if (state.StencilTestEnabled && (!known || !last.StencilTestEnabled ||
+                state.FrontStencil != last.FrontStencil || state.BackStencil != last.BackStencil))
             {
                 _vk.CmdSetStencilCompareMask(command, StencilFaceFlags.FaceFrontBit, state.FrontStencil.CompareMask);
                 _vk.CmdSetStencilCompareMask(command, StencilFaceFlags.FaceBackBit, state.BackStencil.CompareMask);
@@ -686,6 +745,11 @@ internal static unsafe partial class VulkanVideoPresenter
 
                 colorWriteEnable.CmdSetColorWriteEnable(command, state.ColorWriteCount, enables);
             }
+
+            _lastDynamicState = state;
+            _lastDynamicStateBuffer = command;
+            _lastDynamicStateTick = _scheduler.CurrentTick;
+            _lastDynamicStateEpoch = GraphicsDynamicStateEpoch.Value;
         }
 
 

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.Scheduler.cs

@@ -120,6 +120,8 @@ internal static unsafe partial class VulkanVideoPresenter
                 // Drain accepted submissions before closing the relay.
                 // Cancel blocked submissions if a full retry cycle makes no progress.
                 _commandStream.StopAccepting();
+                // The command stream thread ends after its slice; the rest drains here.
+                _commandThread?.Stop(() => RunCommandStreamSlices(long.MaxValue));
                 var outcome = _commandStream.DrainForShutdown(cancelBlockedOnNoProgress: true);
                 FlushBatchedGuestCommands();
                 Console.Error.WriteLine(

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.ShaderPrewarm.cs

@@ -163,7 +163,8 @@ internal static unsafe partial class VulkanVideoPresenter
         {
             if (!ShaderProgramCache.TryCompilePrewarm(
                     record, code, compiler, sharedInt64Atomics, execGuardElision,
-                    _supportsShaderSignedZeroInfNanPreserveFloat32, out var compiled, out var layout, out var error))
+                    _supportsShaderSignedZeroInfNanPreserveFloat32, out var compiled, out var layout, out var error,
+                    ((IShaderPipelineHost)this).ComputeWave64SubgroupNative))
             {
                 NoteShaderPrewarmFailure(record, error);
                 return;
@@ -181,7 +182,7 @@ internal static unsafe partial class VulkanVideoPresenter
                 setLayout = CreateDescriptorSetLayout(bindings, out _, out _);
                 pipelineLayout = CreatePipelineLayout(setLayout, ShaderStageFlags.ComputeBit);
                 module = CreateShaderModule(payload);
-                var pipeline = CompileComputePipeline(_vk, _device, _pipelineCache, module, pipelineLayout);
+                var pipeline = CompileComputePipeline(_vk, _device, _pipelineCache, module, pipelineLayout, ComputeRequiredSubgroupSize(record.Info));
                 _vk.DestroyPipeline(_device, pipeline, null);
                 Interlocked.Increment(ref _shaderPrewarmCompiled);
             }

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.Alu.cs

@@ -52,39 +52,7 @@ public static partial class Gen5SpirvTranslator
                 }
 
                 var value = GetRawSource(instruction, 0);
-                if (_subgroupInvocationIdInput != 0)
-                {
-                    if (_emulateWave64)
-                    {
-                        value = BroadcastFirstWave64Active(value);
-                    }
-                    else
-                    {
-                        // SPIR-V's BroadcastFirst uses the first host-active
-                        // invocation. Guest EXEC is modeled as data, so obtain the
-                        // guest-active mask explicitly and broadcast from its first
-                        // set lane instead. This also updates the private SGPR copy
-                        // for lanes that are currently disabled and may be restored
-                        // by a later saveexec sequence.
-                        var activeLanes = _module.AddInstruction(
-                            SpirvOp.GroupNonUniformBallot,
-                            _uvec4Type,
-                            UInt(3),
-                            Load(_boolType, _exec));
-                        var activeLow = _module.AddInstruction(
-                            SpirvOp.CompositeExtract,
-                            _uintType,
-                            activeLanes,
-                            0);
-                        var firstActiveLane = Ext(73, _uintType, activeLow);
-                        value = _module.AddInstruction(
-                            SpirvOp.GroupNonUniformBroadcast,
-                            _uintType,
-                            UInt(3),
-                            value,
-                            firstActiveLane);
-                    }
-                }
+                value = BroadcastFirstWaveActive(value);
 
                 StoreS(instruction.Destinations[0].Value, value);
                 return true;
@@ -3422,9 +3390,7 @@ public static partial class Gen5SpirvTranslator
             var left = GetRawSource64(instruction, 0);
             if (instruction.Opcode.EndsWith("SaveexecB64", StringComparison.Ordinal))
             {
-                var oldExec = _emulateWave64 && _subgroupInvocationIdInput != 0
-                    ? LoadS64(126)
-                    : BooleanToWaveMask(Load(_boolType, _exec));
+                var oldExec = CurrentExecMask();
                 var notLeft = _module.AddInstruction(SpirvOp.Not, _ulongType, left);
                 var newExec = instruction.Opcode switch
                 {
@@ -3493,12 +3459,7 @@ public static partial class Gen5SpirvTranslator
                     return false;
                 }
 
-                if (!_emulateWave64 && _waveLaneCount != 64)
-                {
-                    newExec = BitwiseAnd64(
-                        newExec,
-                        _module.Constant64(_ulongType, 0xFFFF_FFFFUL));
-                }
+                newExec = NormalizeExecMask(newExec);
 
                 StoreS64(destination, oldExec);
                 StoreS64(126, newExec);
@@ -4958,52 +4919,6 @@ public static partial class Gen5SpirvTranslator
         private uint GuestLaneSelect(uint selector) =>
             BitwiseAnd(selector, UInt(_waveLaneCount == 64 ? 63u : RdnaWaveLaneCount - 1));
 
-        // An emulated wave64 spans two host subgroups, so a subgroup broadcast cannot reach
-        // the other half. The selected guest lane publishes through workgroup scratch, like
-        // BroadcastFirstWave64Active.
-        private uint BroadcastWave64Lane(uint value, uint lane) =>
-            ExchangeWave64Value(_module.AddInstruction(SpirvOp.IEqual, _boolType, GuestWaveLane(), lane), value);
-
-        private uint BroadcastFirstWave64Active(uint value)
-        {
-            var lane = GuestWaveLane();
-            var upperHalf = ShiftRightLogical(lane, UInt(5));
-            var activeInHalf = OwnHalfBallot(Load(_boolType, _exec));
-            var halfHasActive = IsNotZero(activeInHalf);
-            var firstInHalf = _module.AddInstruction(
-                SpirvOp.Select,
-                _uintType,
-                halfHasActive,
-                Ext(73, _uintType, activeInHalf),
-                UInt(0));
-            var halfBase = BitwiseAnd(Load(_uintType, _subgroupInvocationIdInput), UInt(~31u));
-            var halfValue = _module.AddInstruction(
-                SpirvOp.GroupNonUniformBroadcast,
-                _uintType,
-                UInt(3),
-                value,
-                IAdd(halfBase, firstInHalf));
-            var exchange = BeginWave64Exchange();
-            EmitConditional(IsHalfWaveLeader(lane), () =>
-            {
-                Store(Wave64ExchangePointer(exchange, upperHalf), halfValue);
-                Store(
-                    Wave64ExchangePointer(exchange, IAdd(UInt(2), upperHalf)),
-                    _module.AddInstruction(SpirvOp.Select, _uintType, halfHasActive, UInt(1), UInt(0)));
-            });
-            EmitWave64Barrier();
-            var lowerValue = Load(_uintType, Wave64ExchangePointer(exchange, UInt(0)));
-            var upperValue = Load(_uintType, Wave64ExchangePointer(exchange, UInt(1)));
-            var lowerActive = IsNotZero(Load(_uintType, Wave64ExchangePointer(exchange, UInt(2))));
-            var upperActive = IsNotZero(Load(_uintType, Wave64ExchangePointer(exchange, UInt(3))));
-            return _module.AddInstruction(
-                SpirvOp.Select,
-                _uintType,
-                _module.AddInstruction(SpirvOp.LogicalAnd, _boolType, LogicalNot(lowerActive), upperActive),
-                upperValue,
-                lowerValue);
-        }
-
         private void StoreCarryOut(
             Gen5ShaderInstruction instruction,
             uint carry)
@@ -5038,20 +4953,15 @@ public static partial class Gen5SpirvTranslator
             var sourceValue = GetRawSource(instruction, 0);
             var selectedLane = BitwiseAnd(GetRawSource(instruction, 1), UInt(LaneSelectMask));
 
-            if (_emulateWave64)
+            if (_stage == Gen5SpirvStage.Compute && _waveLowering is EmulatedWave64Lowering)
             {
                 // The selected guest lane can belong to another host subgroup.
                 // Read it even when the guest execution mask disables that lane.
-                StoreS(destination, BroadcastWave64Lane(sourceValue, selectedLane));
+                StoreS(destination, BroadcastWaveLane(sourceValue, selectedLane));
             }
             else if (_subgroupInvocationIdInput != 0 && _stage == Gen5SpirvStage.Compute)
             {
-                var broadcast = _module.AddInstruction(
-                    SpirvOp.GroupNonUniformBroadcast,
-                    _uintType,
-                    UInt(3),
-                    sourceValue,
-                    selectedLane);
+                var broadcast = BroadcastWaveLane(sourceValue, selectedLane);
                 StoreS(destination, ReadLaneSpillSlot(instruction, selectedLane, broadcast));
             }
             else

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.CommonWave.cs

@@ -0,0 +1,27 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.ShaderCompiler.Vulkan;
+
+public static partial class Gen5SpirvTranslator
+{
+    private sealed partial class CompilationContext
+    {
+        private uint BroadcastFirstWaveActive(uint value) =>
+            _waveLowering.BroadcastFirstActive(value);
+
+        private uint BroadcastWaveLane(uint value, uint lane) =>
+            _waveLowering.BroadcastLane(value, lane);
+
+        private uint BroadcastWaveValue(uint writer, uint value, uint lane) =>
+            _waveLowering.BroadcastValue(writer, value, lane);
+
+        private uint CurrentExecMask() =>
+            _waveLowering.CurrentExecMask();
+
+        private uint NormalizeExecMask(uint mask) =>
+            _waveLowering.NormalizeExecMask(mask);
+    }
+}
+
+

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.EmulatedWave64Lowering.cs

@@ -0,0 +1,578 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.ShaderCompiler.Resources;
+
+namespace SharpEmu.ShaderCompiler.Vulkan;
+
+public static partial class Gen5SpirvTranslator
+{
+    private sealed partial class CompilationContext
+    {
+        private sealed class EmulatedWave64Lowering : IWaveLowering
+        {
+            private const uint ExchangeSlotCount = 4;
+            private const uint ExchangeDwordCount = ExchangeSlotCount * 2;
+            private const uint GuestWaveLaneCount = 64;
+            private static readonly bool HalfWaveMasks = !string.Equals(
+                Environment.GetEnvironmentVariable("SHARPEMU_WAVE64_HALF_MASKS"),
+                "0",
+                StringComparison.Ordinal);
+
+            private readonly CompilationContext _context;
+            private uint _exchange;
+            private uint _exchangeElementPointer;
+            private uint _exchangeOffset;
+            private uint _exchangeParity;
+            // Several guest waves share one workgroup. Each gets its own exchange area plus an
+            // arrival counter, and its two halves meet at that counter instead of at a workgroup
+            // barrier, which other waves (possibly in other blocks) would never reach.
+            private readonly uint _waveCount;
+            private readonly bool _multiWave;
+            private readonly uint _waveStride;
+            private uint _rendezvousEpoch;
+            private Ir.Gen5Wave64HalfMaskPlan? _halfMaskPlan;
+            private bool _halfMaskPlanBuilt;
+            private bool _synchronizeSharedMemory;
+            private SharedMemoryPhase _sharedMemoryPhase;
+            private uint? _emittingPc;
+
+            private enum SharedMemoryPhase { None, Read, Write }
+
+            public EmulatedWave64Lowering(CompilationContext context)
+            {
+                _context = context;
+                var threads = (ulong)context._localSizeX * context._localSizeY * context._localSizeZ;
+                _waveCount = (uint)Math.Max((threads + GuestWaveLaneCount - 1) / GuestWaveLaneCount, 1);
+                _multiWave = _waveCount > 1;
+                _waveStride = _multiWave ? ExchangeDwordCount + 1 : ExchangeDwordCount;
+            }
+
+            private uint ExchangeTotalDwords => _waveCount * _waveStride;
+
+            public bool RequiresCrossSubgroupSynchronization => true;
+
+            private bool UsesExchange => _context.UsesSubgroupOperations();
+
+            public uint ReserveLdsDwords(uint guestDwordCount) =>
+                UsesExchange && guestDwordCount < LdsDwordCount
+                    ? guestDwordCount + ExchangeTotalDwords
+                    : guestDwordCount;
+
+            public void DeclareResources()
+            {
+                var context = _context;
+                if (!UsesExchange)
+                {
+                    return;
+                }
+
+                if (context._lds != 0)
+                {
+                    // Metal exposes 32 KiB of threadgroup memory on the Apple
+                    // GPUs we target. Reuse the tail of guest LDS so the
+                    // cross-subgroup bridge stays within that limit.
+                    _exchange = context._lds;
+                    _exchangeElementPointer = context._ldsElementPointer;
+                    _exchangeOffset = context._ldsDwordMask + 1 < LdsDwordCount
+                        ? context._ldsDwordMask + 1
+                        : LdsDwordCount - ExchangeTotalDwords;
+                    return;
+                }
+
+                var exchangeArrayType = context._module.TypeArray(context._uintType, ExchangeTotalDwords);
+                _exchangeElementPointer = context._module.TypePointer(SpirvStorageClass.Workgroup, context._uintType);
+                _exchange = context._module.AddGlobalVariable(
+                    context._module.TypePointer(SpirvStorageClass.Workgroup, exchangeArrayType),
+                    SpirvStorageClass.Workgroup);
+                _exchangeOffset = 0;
+                context._module.AddName(_exchange, "wave64Exchange");
+                context._interfaces.Add(_exchange);
+            }
+
+            public void DeclareFunctionResources()
+            {
+                if (!UsesExchange)
+                {
+                    return;
+                }
+
+                _exchangeParity = _context.DeclareFunctionUint(
+                    _context._module.Constant(_context._uintType, 0));
+                _context._module.AddName(_exchangeParity, "wave64ExchangeParity");
+                if (_multiWave)
+                {
+                    _rendezvousEpoch = _context.DeclareFunctionUint(
+                        _context._module.Constant(_context._uintType, 0));
+                    _context._module.AddName(_rendezvousEpoch, "wave64RendezvousEpoch");
+                }
+            }
+
+            public void InitializeMasks()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput != 0)
+                {
+                    context.StoreS64(106, BothHalvesOfOwnBallot(context._module.ConstantBool(false)));
+                    context.StoreS64(126, BothHalvesOfOwnBallot(context._module.ConstantBool(true)));
+                    InitializeRendezvous();
+                    return;
+                }
+
+                context.StoreS64(106, context._module.Constant64(context._ulongType, 0));
+                context.StoreS64(126, context._module.Constant64(context._ulongType, 1));
+            }
+
+            public void BeginBlock()
+            {
+                _synchronizeSharedMemory = HalfMaskPlan() is null;
+                _sharedMemoryPhase = SharedMemoryPhase.None;
+            }
+
+            public void BeforeInstruction(Gen5ShaderInstruction instruction)
+            {
+                var context = _context;
+                _emittingPc = instruction.Pc;
+                if (HalfMaskPlan() is { } plan)
+                {
+                    if (plan.ExactPairsBefore.TryGetValue(instruction.Pc, out var exactPairs))
+                    {
+                        EmitExactWaveMasks(exactPairs);
+                    }
+
+                    if (plan.SharedMemoryBarriersBefore.Contains(instruction.Pc))
+                    {
+                        Barrier();
+                    }
+                }
+
+                if (!_synchronizeSharedMemory)
+                {
+                    return;
+                }
+
+                var nextPhase = instruction.Control is Gen5DataShareControl { Gds: false }
+                    ? instruction.Opcode.StartsWith("DsRead", StringComparison.Ordinal) ? SharedMemoryPhase.Read
+                    : instruction.Opcode.StartsWith("DsWrite", StringComparison.Ordinal) ? SharedMemoryPhase.Write
+                    : SharedMemoryPhase.None
+                    : SharedMemoryPhase.None;
+                if (instruction.Opcode == "SBarrier")
+                {
+                    _sharedMemoryPhase = SharedMemoryPhase.None;
+                }
+                if (nextPhase != SharedMemoryPhase.None)
+                {
+                    if (_sharedMemoryPhase != SharedMemoryPhase.None && _sharedMemoryPhase != nextPhase)
+                    {
+                        Barrier();
+                    }
+                    _sharedMemoryPhase = nextPhase;
+                }
+            }
+
+            public void EndBlock()
+            {
+                if (_synchronizeSharedMemory && _sharedMemoryPhase != SharedMemoryPhase.None)
+                {
+                    Barrier();
+                }
+            }
+
+            public uint SubgroupAny(uint condition) =>
+                _context._subgroupInvocationIdInput == 0
+                    ? condition
+                    : _context.IsNotZero64(BooleanToWaveMask(condition));
+
+            public uint CurrentLaneBit()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context._module.Constant64(context._ulongType, 1);
+                }
+
+                return context.ShiftLeftLogical64(
+                    context._module.Constant64(context._ulongType, 1),
+                    context._module.AddInstruction(
+                        SpirvOp.UConvert,
+                        context._ulongType,
+                        context.GuestWaveLane()));
+            }
+
+            public uint BooleanToWaveMask(uint condition)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context.BooleanToLaneMask(condition);
+                }
+
+                var lane = context.GuestWaveLane();
+                var owned = OwnHalfBallot(condition);
+                var exchange = BeginExchange();
+                context.EmitConditional(IsHalfWaveLeader(lane), () =>
+                    context.Store(ExchangePointer(exchange, context.ShiftRightLogical(lane, context.UInt(5))), owned));
+                Barrier();
+                return context.Pair64(
+                    context.Load(context._uintType, ExchangePointer(exchange, context.UInt(0))),
+                    context.Load(context._uintType, ExchangePointer(exchange, context.UInt(1))));
+            }
+
+            public uint BroadcastFirstActive(uint value)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return value;
+                }
+
+                var lane = context.GuestWaveLane();
+                var upperHalf = context.ShiftRightLogical(lane, context.UInt(5));
+                var activeInHalf = OwnHalfBallot(context.Load(context._boolType, context._exec));
+                var halfHasActive = context.IsNotZero(activeInHalf);
+                var firstInHalf = context._module.AddInstruction(
+                    SpirvOp.Select,
+                    context._uintType,
+                    halfHasActive,
+                    context.Ext(73, context._uintType, activeInHalf),
+                    context.UInt(0));
+                var halfBase = context.BitwiseAnd(
+                    context.Load(context._uintType, context._subgroupInvocationIdInput),
+                    context.UInt(~31u));
+                var halfValue = context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBroadcast,
+                    context._uintType,
+                    context.UInt(3),
+                    value,
+                    context.IAdd(halfBase, firstInHalf));
+                var exchange = BeginExchange();
+                context.EmitConditional(IsHalfWaveLeader(lane), () =>
+                {
+                    context.Store(ExchangePointer(exchange, upperHalf), halfValue);
+                    context.Store(
+                        ExchangePointer(exchange, context.IAdd(context.UInt(2), upperHalf)),
+                        context._module.AddInstruction(
+                            SpirvOp.Select,
+                            context._uintType,
+                            halfHasActive,
+                            context.UInt(1),
+                            context.UInt(0)));
+                });
+                Barrier();
+                var lowerValue = context.Load(context._uintType, ExchangePointer(exchange, context.UInt(0)));
+                var upperValue = context.Load(context._uintType, ExchangePointer(exchange, context.UInt(1)));
+                var lowerActive = context.IsNotZero(context.Load(context._uintType, ExchangePointer(exchange, context.UInt(2))));
+                var upperActive = context.IsNotZero(context.Load(context._uintType, ExchangePointer(exchange, context.UInt(3))));
+                return context._module.AddInstruction(
+                    SpirvOp.Select,
+                    context._uintType,
+                    context._module.AddInstruction(
+                        SpirvOp.LogicalAnd,
+                        context._boolType,
+                        context.LogicalNot(lowerActive),
+                        upperActive),
+                    upperValue,
+                    lowerValue);
+            }
+
+            public uint BroadcastLane(uint value, uint lane) =>
+                ExchangeValue(
+                    _context._module.AddInstruction(
+                        SpirvOp.IEqual,
+                        _context._boolType,
+                        _context.GuestWaveLane(),
+                        lane),
+                    value);
+
+            public uint BroadcastValue(uint writer, uint value, uint lane) => ExchangeValue(writer, value);
+
+            public uint WaveMaskAny(uint register, uint laneActive) =>
+                _context._subgroupInvocationIdInput != 0
+                    ? _context.IsNotZero64(_context.LoadS64(register))
+                    : SubgroupAny(_context.Load(_context._boolType, laneActive));
+
+            public uint CurrentExecMask() =>
+                _context._subgroupInvocationIdInput != 0
+                    ? _context.LoadS64(126)
+                    : BooleanToWaveMask(_context.Load(_context._boolType, _context._exec));
+
+            public uint NormalizeExecMask(uint mask) => mask;
+
+            public void StoreWaveMask(uint register, uint condition)
+            {
+                var context = _context;
+                if (context._waveLaneCount == 32)
+                {
+                    context.StoreS(register, context.Narrow(BooleanToWaveMask(condition)));
+                    return;
+                }
+
+                if (context._subgroupInvocationIdInput != 0 &&
+                    _emittingPc is { } pc &&
+                    HalfMaskPlan() is { } plan &&
+                    plan.HalfMaskWrites.TryGetValue(pc, out var halfRegister) &&
+                    halfRegister == register)
+                {
+                    context.StoreS64(register, BooleanToHalfWaveMask(condition));
+                    return;
+                }
+
+                context.StoreS64(register, BooleanToWaveMask(condition));
+            }
+
+            private uint OwnHalfBallot(uint condition)
+            {
+                var context = _context;
+                var ballot = context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBallot,
+                    context._uvec4Type,
+                    context.UInt(3),
+                    condition);
+                return context._module.AddInstruction(
+                    SpirvOp.VectorExtractDynamic,
+                    context._uintType,
+                    ballot,
+                    context.ShiftRightLogical(
+                        context.Load(context._uintType, context._subgroupInvocationIdInput),
+                        context.UInt(5)));
+            }
+
+            private uint BooleanToHalfWaveMask(uint condition)
+            {
+                var context = _context;
+                var owned = context._module.AddInstruction(SpirvOp.UConvert, context._ulongType, OwnHalfBallot(condition));
+                return context._module.AddInstruction(
+                    SpirvOp.Select,
+                    context._ulongType,
+                    context._module.AddInstruction(
+                        SpirvOp.UGreaterThanEqual,
+                        context._boolType,
+                        context.GuestWaveLane(),
+                        context.UInt(32)),
+                    context.ShiftLeftLogical64(owned, context._module.Constant64(context._ulongType, 32)),
+                    owned);
+            }
+
+            private uint BothHalvesOfOwnBallot(uint condition)
+            {
+                var context = _context;
+                var owned = context._module.AddInstruction(SpirvOp.UConvert, context._ulongType, OwnHalfBallot(condition));
+                return context.BitwiseOr64(
+                    owned,
+                    context.ShiftLeftLogical64(owned, context._module.Constant64(context._ulongType, 32)));
+            }
+
+            private uint IsHalfWaveLeader(uint lane) =>
+                _context._module.AddInstruction(
+                    SpirvOp.IEqual,
+                    _context._boolType,
+                    _context.BitwiseAnd(lane, _context.UInt(31)),
+                    _context.UInt(0));
+
+            private uint BeginExchange()
+            {
+                var context = _context;
+                var parity = context.Load(context._uintType, _exchangeParity);
+                context.Store(_exchangeParity, context.BitwiseXor(parity, context.UInt(1)));
+                return context.IAdd(
+                    WaveExchangeBase(),
+                    context._module.AddInstruction(
+                        SpirvOp.IMul,
+                        context._uintType,
+                        parity,
+                        context.UInt(ExchangeSlotCount)));
+            }
+
+            private uint WaveExchangeBase()
+            {
+                var context = _context;
+                if (!_multiWave)
+                {
+                    return context.UInt(_exchangeOffset);
+                }
+
+                var wave = context.ShiftRightLogical(
+                    context.Load(context._uintType, context._localInvocationIndexInput),
+                    context.UInt(6));
+                return context.IAdd(
+                    context.UInt(_exchangeOffset),
+                    context._module.AddInstruction(SpirvOp.IMul, context._uintType, wave, context.UInt(_waveStride)));
+            }
+
+            // Zeroes each wave's arrival counter. Runs once at entry, where the whole workgroup is
+            // converged, so the workgroup barrier is safe here.
+            private void InitializeRendezvous()
+            {
+                if (!_multiWave || !UsesExchange)
+                {
+                    return;
+                }
+
+                var context = _context;
+                context.EmitConditional(
+                    context._module.AddInstruction(
+                        SpirvOp.IEqual,
+                        context._boolType,
+                        context.GuestWaveLane(),
+                        context.UInt(0)),
+                    () => context.Store(
+                        ExchangePointer(WaveExchangeBase(), context.UInt(ExchangeDwordCount)),
+                        context.UInt(0)));
+                WorkgroupBarrier();
+            }
+
+            // Both halves of the wave arrive, then wait until the counter shows two arrivals per
+            // rendezvous so far. The halves of one wave are co-resident in the workgroup, so the
+            // wait always ends.
+            private void WaveRendezvous()
+            {
+                var context = _context;
+                var counter = ExchangePointer(WaveExchangeBase(), context.UInt(ExchangeDwordCount));
+                context.EmitConditional(
+                    IsHalfWaveLeader(context.GuestWaveLane()),
+                    () => context._module.AddInstruction(
+                        SpirvOp.AtomicIAdd,
+                        context._uintType,
+                        counter,
+                        context.UInt(2),
+                        context.UInt(0x108),
+                        context.UInt(1)));
+                var epoch = context.IAdd(context.Load(context._uintType, _rendezvousEpoch), context.UInt(1));
+                context.Store(_rendezvousEpoch, epoch);
+                var target = context._module.AddInstruction(SpirvOp.IMul, context._uintType, epoch, context.UInt(2));
+
+                var header = context._module.AllocateId();
+                var continueLabel = context._module.AllocateId();
+                var mergeLabel = context._module.AllocateId();
+                context._module.AddStatement(SpirvOp.Branch, header);
+                context._module.AddLabel(header);
+                var arrived = context._module.AddInstruction(
+                    SpirvOp.AtomicLoad,
+                    context._uintType,
+                    counter,
+                    context.UInt(2),
+                    context.UInt(0x102));
+                var done = context._module.AddInstruction(
+                    SpirvOp.UGreaterThanEqual,
+                    context._boolType,
+                    arrived,
+                    target);
+                context._module.AddStatement(SpirvOp.LoopMerge, mergeLabel, continueLabel, 0);
+                context._module.AddStatement(SpirvOp.BranchConditional, done, mergeLabel, continueLabel);
+                context._module.AddLabel(continueLabel);
+                context._module.AddStatement(SpirvOp.Branch, header);
+                context._module.AddLabel(mergeLabel);
+            }
+
+            private uint ExchangePointer(uint exchange, uint slot) =>
+                _context._module.AddInstruction(
+                    SpirvOp.AccessChain,
+                    _exchangeElementPointer,
+                    _exchange,
+                    _context.IAdd(exchange, slot));
+
+            private uint ExchangeValue(uint writer, uint value)
+            {
+                var context = _context;
+                var exchange = BeginExchange();
+                context.EmitConditional(writer, () => context.Store(ExchangePointer(exchange, context.UInt(0)), value));
+                Barrier();
+                return context.Load(context._uintType, ExchangePointer(exchange, context.UInt(0)));
+            }
+
+            private void EmitExactWaveMasks(IReadOnlyList<uint> pairs)
+            {
+                var context = _context;
+                var lane = context.GuestWaveLane();
+                var upperHalf = context.ShiftRightLogical(lane, context.UInt(5));
+                var lowerHalf = context._module.AddInstruction(SpirvOp.IEqual, context._boolType, upperHalf, context.UInt(0));
+                const int pairsPerExchange = (int)ExchangeSlotCount / 2;
+                for (var start = 0; start < pairs.Count; start += pairsPerExchange)
+                {
+                    var count = Math.Min(pairsPerExchange, pairs.Count - start);
+                    var owned = new uint[count];
+                    for (var index = 0; index < count; index++)
+                    {
+                        var pair = pairs[start + index];
+                        owned[index] = context._module.AddInstruction(
+                            SpirvOp.Select,
+                            context._uintType,
+                            lowerHalf,
+                            context.LoadS(pair),
+                            context.LoadS(pair + 1));
+                    }
+
+                    var exchange = BeginExchange();
+                    context.EmitConditional(IsHalfWaveLeader(lane), () =>
+                    {
+                        for (var index = 0; index < count; index++)
+                        {
+                            context.Store(
+                                ExchangePointer(exchange, context.IAdd(context.UInt((uint)index * 2), upperHalf)),
+                                owned[index]);
+                        }
+                    });
+                    Barrier();
+                    for (var index = 0; index < count; index++)
+                    {
+                        var pair = pairs[start + index];
+                        context.StoreS(pair, context.Load(context._uintType, ExchangePointer(exchange, context.UInt((uint)index * 2))));
+                        context.StoreS(pair + 1, context.Load(context._uintType, ExchangePointer(exchange, context.UInt((uint)index * 2 + 1))));
+                    }
+                }
+            }
+
+            private Ir.Gen5Wave64HalfMaskPlan? HalfMaskPlan()
+            {
+                if (_halfMaskPlanBuilt)
+                {
+                    return _halfMaskPlan;
+                }
+
+                _halfMaskPlanBuilt = true;
+                var context = _context;
+                if (HalfWaveMasks && UsesExchange && context._subgroupInvocationIdInput != 0)
+                {
+                    var sharedFlatPcs = context._request.Memory.Entries
+                        .Where(static memory => memory.AddressSpace is FlatAddressSpace.Shared or FlatAddressSpace.SharedOrPrivate)
+                        .Select(static memory => memory.Pc)
+                        .ToHashSet();
+                    _halfMaskPlan = Ir.Gen5Wave64HalfMaskAnalysis.Analyze(context._request.Program, sharedFlatPcs);
+                }
+
+                return _halfMaskPlan;
+            }
+
+            private void Barrier()
+            {
+                if (_multiWave)
+                {
+                    // The rendezvous needs the exchange area, declared only with subgroup operations.
+                    if (UsesExchange)
+                    {
+                        WaveRendezvous();
+                    }
+
+                    return;
+                }
+
+                WorkgroupBarrier();
+            }
+
+            private void WorkgroupBarrier()
+            {
+                var workgroup = _context.UInt(2);
+                _context._module.AddStatement(
+                    SpirvOp.ControlBarrier,
+                    workgroup,
+                    workgroup,
+                    _context.UInt(0x108));
+            }
+        }
+    }
+}
+
+
+
+
+

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.NativeWave32Lowering.cs

@@ -0,0 +1,156 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.ShaderCompiler.Vulkan;
+
+public static partial class Gen5SpirvTranslator
+{
+    private sealed partial class CompilationContext
+    {
+        // A guest Wave32 maps directly to one 32-lane host subgroup. This is the
+        // former implicit/default lowering, made explicit so every guest wave size
+        // has a concrete lowering strategy.
+        private sealed class NativeWave32Lowering : IWaveLowering
+        {
+            private readonly CompilationContext _context;
+
+            public NativeWave32Lowering(CompilationContext context)
+            {
+                _context = context;
+            }
+
+            public bool RequiresCrossSubgroupSynchronization => false;
+
+            public uint ReserveLdsDwords(uint guestDwordCount) => guestDwordCount;
+
+            public void DeclareResources() { }
+            public void DeclareFunctionResources() { }
+            public void BeginBlock() { }
+            public void BeforeInstruction(Gen5ShaderInstruction instruction) { }
+            public void EndBlock() { }
+
+            public void InitializeMasks()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput != 0)
+                {
+                    StoreWaveMask(106, context._module.ConstantBool(false));
+                    StoreWaveMask(126, context._module.ConstantBool(true));
+                    return;
+                }
+
+                context.StoreS64(106, context._module.Constant64(context._ulongType, 0));
+                context.StoreS64(126, context._module.Constant64(context._ulongType, 1));
+            }
+
+            public uint BooleanToWaveMask(uint condition)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context._module.AddInstruction(
+                        SpirvOp.Select,
+                        context._ulongType,
+                        condition,
+                        CurrentLaneBit(),
+                        context._module.Constant64(context._ulongType, 0));
+                }
+
+                var ballot = context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBallot,
+                    context._uvec4Type,
+                    context.UInt(3),
+                    condition);
+                var low = context._module.AddInstruction(
+                    SpirvOp.CompositeExtract,
+                    context._uintType,
+                    ballot,
+                    0);
+                return context._module.AddInstruction(SpirvOp.UConvert, context._ulongType, low);
+            }
+
+            public uint SubgroupAny(uint condition) =>
+                _context._subgroupInvocationIdInput == 0
+                    ? condition
+                    : _context._module.AddInstruction(
+                        SpirvOp.GroupNonUniformAny,
+                        _context._boolType,
+                        _context.UInt(3),
+                        condition);
+
+            public uint CurrentLaneBit()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context._module.Constant64(context._ulongType, 1);
+                }
+
+                var shifted = context.ShiftLeftLogical64(
+                    context._module.Constant64(context._ulongType, 1),
+                    context._module.AddInstruction(
+                        SpirvOp.UConvert,
+                        context._ulongType,
+                        context.GuestWaveLane()));
+                return context._module.AddInstruction(
+                    SpirvOp.Select,
+                    context._ulongType,
+                    context.IsCurrentLaneInRdnaWave(),
+                    shifted,
+                    context._module.Constant64(context._ulongType, 0));
+            }
+
+            public uint BroadcastFirstActive(uint value)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return value;
+                }
+
+                var activeLanes = context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBallot,
+                    context._uvec4Type,
+                    context.UInt(3),
+                    context.Load(context._boolType, context._exec));
+                var activeLow = context._module.AddInstruction(
+                    SpirvOp.CompositeExtract,
+                    context._uintType,
+                    activeLanes,
+                    0);
+                var firstActiveLane = context.Ext(73, context._uintType, activeLow);
+                return context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBroadcast,
+                    context._uintType,
+                    context.UInt(3),
+                    value,
+                    firstActiveLane);
+            }
+
+            public uint BroadcastLane(uint value, uint lane) =>
+                _context._subgroupInvocationIdInput == 0
+                    ? value
+                    : _context._module.AddInstruction(
+                        SpirvOp.GroupNonUniformBroadcast,
+                        _context._uintType,
+                        _context.UInt(3),
+                        value,
+                        lane);
+
+            public uint BroadcastValue(uint writer, uint value, uint lane) =>
+                _context.ShuffleLane(value, lane);
+
+            public uint WaveMaskAny(uint register, uint laneActive) =>
+                SubgroupAny(_context.Load(_context._boolType, laneActive));
+
+            public uint CurrentExecMask() =>
+                BooleanToWaveMask(_context.Load(_context._boolType, _context._exec));
+
+            public uint NormalizeExecMask(uint mask) =>
+                _context.BitwiseAnd64(mask, _context._module.Constant64(_context._ulongType, 0xFFFF_FFFFUL));
+
+            public void StoreWaveMask(uint register, uint condition) =>
+                _context.StoreS(register, _context.Narrow(BooleanToWaveMask(condition)));
+        }
+    }
+}

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.NativeWave64Lowering.cs

@@ -0,0 +1,160 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.ShaderCompiler.Vulkan;
+
+public static partial class Gen5SpirvTranslator
+{
+    private sealed partial class CompilationContext
+    {
+        // One guest wave is one 64-lane host subgroup (ballot.x = lanes 0..31, ballot.y = lanes
+        // 32..63). It has no workgroup exchange, LDS reservation, half-mask plan or
+        // cross-subgroup barrier state.
+        private sealed class NativeWave64Lowering : IWaveLowering
+        {
+            private readonly CompilationContext _context;
+
+            public NativeWave64Lowering(CompilationContext context)
+            {
+                _context = context;
+            }
+
+            public bool RequiresCrossSubgroupSynchronization => false;
+
+            public uint ReserveLdsDwords(uint guestDwordCount) => guestDwordCount;
+
+            public void DeclareResources() { }
+            public void DeclareFunctionResources() { }
+            public void BeginBlock() { }
+            public void BeforeInstruction(Gen5ShaderInstruction instruction) { }
+            public void EndBlock() { }
+
+            public void InitializeMasks()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput != 0)
+                {
+                    StoreWaveMask(106, context._module.ConstantBool(false));
+                    StoreWaveMask(126, context._module.ConstantBool(true));
+                }
+                else
+                {
+                    context.StoreS64(106, context._module.Constant64(context._ulongType, 0));
+                    context.StoreS64(126, context._module.Constant64(context._ulongType, 1));
+                }
+            }
+
+            public uint BooleanToWaveMask(uint condition)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context._module.AddInstruction(
+                        SpirvOp.Select,
+                        context._ulongType,
+                        condition,
+                        CurrentLaneBit(),
+                        context._module.Constant64(context._ulongType, 0));
+                }
+
+                return BallotToMask(condition);
+            }
+
+            // The host subgroup is at least 64 wide: lanes 0..31 are ballot.x and lanes 32..63 ballot.y.
+            private uint BallotToMask(uint condition)
+            {
+                var context = _context;
+                var ballot = context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBallot,
+                    context._uvec4Type,
+                    context.UInt(3),
+                    condition);
+                var low = context._module.AddInstruction(SpirvOp.CompositeExtract, context._uintType, ballot, 0);
+                var high = context._module.AddInstruction(SpirvOp.CompositeExtract, context._uintType, ballot, 1);
+                return context.BitwiseOr64(
+                    context._module.AddInstruction(SpirvOp.UConvert, context._ulongType, low),
+                    context.ShiftLeftLogical64(
+                        context._module.AddInstruction(SpirvOp.UConvert, context._ulongType, high),
+                        context._module.Constant64(context._ulongType, 32)));
+            }
+
+            public uint SubgroupAny(uint condition) =>
+                _context._subgroupInvocationIdInput == 0
+                    ? condition
+                    : _context._module.AddInstruction(
+                        SpirvOp.GroupNonUniformAny,
+                        _context._boolType,
+                        _context.UInt(3),
+                        condition);
+
+            public uint CurrentLaneBit()
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return context._module.Constant64(context._ulongType, 1);
+                }
+
+                var shifted = context.ShiftLeftLogical64(
+                    context._module.Constant64(context._ulongType, 1),
+                    context._module.AddInstruction(
+                        SpirvOp.UConvert,
+                        context._ulongType,
+                        context.GuestWaveLane()));
+                return shifted;
+            }
+
+            public uint BroadcastFirstActive(uint value)
+            {
+                var context = _context;
+                if (context._subgroupInvocationIdInput == 0)
+                {
+                    return value;
+                }
+
+                var activeLanes = BallotToMask(context.Load(context._boolType, context._exec));
+                var activeLow = context._module.AddInstruction(SpirvOp.UConvert, context._uintType, activeLanes);
+                var activeHigh = context._module.AddInstruction(
+                    SpirvOp.UConvert,
+                    context._uintType,
+                    context.ShiftRightLogical64(activeLanes, context._module.Constant64(context._ulongType, 32)));
+                var firstActiveLane = context._module.AddInstruction(
+                    SpirvOp.Select,
+                    context._uintType,
+                    context._module.AddInstruction(SpirvOp.IEqual, context._boolType, activeLow, context.UInt(0)),
+                    context.IAdd(context.Ext(73, context._uintType, activeHigh), context.UInt(32)),
+                    context.Ext(73, context._uintType, activeLow));
+                return context._module.AddInstruction(
+                    SpirvOp.GroupNonUniformBroadcast,
+                    context._uintType,
+                    context.UInt(3),
+                    value,
+                    firstActiveLane);
+            }
+
+            public uint BroadcastLane(uint value, uint lane) =>
+                _context._subgroupInvocationIdInput == 0
+                    ? value
+                    : _context._module.AddInstruction(
+                        SpirvOp.GroupNonUniformBroadcast,
+                        _context._uintType,
+                        _context.UInt(3),
+                        value,
+                        lane);
+
+            public uint BroadcastValue(uint writer, uint value, uint lane) =>
+                _context.ShuffleLane(value, lane);
+
+            public uint WaveMaskAny(uint register, uint laneActive) =>
+                SubgroupAny(_context.Load(_context._boolType, laneActive));
+
+            public uint CurrentExecMask() =>
+                BooleanToWaveMask(_context.Load(_context._boolType, _context._exec));
+
+            public uint NormalizeExecMask(uint mask) => mask;
+
+            public void StoreWaveMask(uint register, uint condition) =>
+                _context.StoreS64(register, BooleanToWaveMask(condition));
+        }
+    }
+}

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.Resources.cs

@@ -116,7 +116,23 @@ public static partial class Gen5SpirvTranslator
             _localSizeY = Math.Max(request.LocalSizeY, 1);
             _localSizeZ = Math.Max(request.LocalSizeZ, 1);
             _physicalAxisOfLogical = ComputeWorkgroupAxisOrder(_localSizeX, _localSizeY, _localSizeZ);
-            _emulateWave64 = _stage == Gen5SpirvStage.Compute && _waveLaneCount == 64 && (ulong)_localSizeX * _localSizeY * _localSizeZ == 64;
+            // A wave64 runs as one 64-lane host subgroup when the host pins it to one
+            // (HostWave64Supported). Hosts without 64-lane subgroups bridge the two 32-lane
+            // halves through workgroup memory.
+            if (_waveLaneCount == 32)
+            {
+                _waveLowering = new NativeWave32Lowering(this);
+            }
+            else if (request.HostWave64Supported)
+            {
+                _waveLowering = new NativeWave64Lowering(this);
+                _guestLaneFromSubgroup = true;
+            }
+            else
+            {
+                _waveLowering = new EmulatedWave64Lowering(this);
+            }
+
             _requiredVertexOutputCount = request.RequiredVertexOutputCount;
             _pixelInputEnable = request.PixelInputEnable;
             _pixelInputAddress = request.PixelInputAddress;
@@ -1909,21 +1925,7 @@ public static partial class Gen5SpirvTranslator
             });
 
             var firstValue = LoadV(destination);
-            uint broadcast;
-            if (_emulateWave64)
-            {
-                broadcast = ExchangeWave64Value(isFirstActive, firstValue);
-            }
-            else if (_stage == Gen5SpirvStage.Compute || _enableGraphicsSubgroupOperations)
-            {
-                // Every active lane receives the first lane's old counter, in graphics stages too.
-                broadcast = _module.AddInstruction(SpirvOp.GroupNonUniformShuffle, _uintType, UInt(3), firstValue, firstLane);
-            }
-            else
-            {
-                // Without subgroup operations a graphics invocation is its own wave of one lane.
-                broadcast = firstValue;
-            }
+            var broadcast = BroadcastWaveValue(isFirstActive, firstValue, firstLane);
 
             var validResult = LogicalAnd(inBounds, IsNotZero64(activeMask));
             StoreV(destination, _module.AddInstruction(SpirvOp.Select, _uintType, validResult, broadcast, UInt(0)));

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.WaveLowering.cs

@@ -0,0 +1,34 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.ShaderCompiler.Vulkan;
+
+public static partial class Gen5SpirvTranslator
+{
+    private sealed partial class CompilationContext
+    {
+        private interface IWaveLowering
+        {
+            bool RequiresCrossSubgroupSynchronization { get; }
+
+            uint ReserveLdsDwords(uint guestDwordCount);
+            void DeclareResources();
+            void DeclareFunctionResources();
+            void InitializeMasks();
+            void BeginBlock();
+            void BeforeInstruction(Gen5ShaderInstruction instruction);
+            void EndBlock();
+            uint BooleanToWaveMask(uint condition);
+            uint SubgroupAny(uint condition);
+            uint CurrentLaneBit();
+            uint BroadcastFirstActive(uint value);
+            uint BroadcastLane(uint value, uint lane);
+            uint BroadcastValue(uint writer, uint value, uint lane);
+            uint WaveMaskAny(uint register, uint laneActive);
+            uint CurrentExecMask();
+            uint NormalizeExecMask(uint mask);
+            void StoreWaveMask(uint register, uint condition);
+        }
+    }
+}
+

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.cs

@@ -17,8 +17,6 @@ public static partial class Gen5SpirvTranslator
     // so a smaller array is safe.
     private const uint PrivateLdsDwordCount = 2048;
     private const uint RdnaWaveLaneCount = 32;
-    private const uint Wave64ExchangeSlotCount = 4;
-    private const uint Wave64ExchangeDwordCount = Wave64ExchangeSlotCount * 2;
 
     internal static SpirvImageFormat DecodeStorageImageFormat(
         uint dataFormat,
@@ -71,7 +69,8 @@ public static partial class Gen5SpirvTranslator
         private readonly bool _usesPixelValidMask;
         private readonly bool _enableGraphicsSubgroupOperations;
         private readonly uint _waveLaneCount;
-        private readonly bool _emulateWave64;
+        private readonly IWaveLowering _waveLowering;
+        private bool _guestLaneFromSubgroup;
 
         // Safety valve for the PC-dispatcher loop. Each iteration executes one
         // GCN basic block; a correctly-translated shader always reaches its
@@ -195,12 +194,6 @@ public static partial class Gen5SpirvTranslator
         private uint _workGroupIdInput;
         private uint _pushConstantUintPointer;
         private uint _subgroupInvocationIdInput;
-        private uint _wave64Exchange;
-        private uint _wave64ExchangeElementPointer;
-        private uint _wave64ExchangeOffset;
-        private uint _wave64ExchangeParity;
-        private Ir.Gen5Wave64HalfMaskPlan? _halfMaskPlan;
-        private uint? _emittingPc;
         private uint _glsl;
 
         private enum ImageComponentKind
@@ -421,8 +414,8 @@ public static partial class Gen5SpirvTranslator
                 var active = Load(_boolType, _programActive);
                 if (_maxDispatcherSteps > 0)
                 {
-                    var steps = IAdd(Load(_uintType, _iterationGuard), UInt(1));
-                    Store(_iterationGuard, steps);
+                    var steps = IAdd(Load(_uintType, IterationGuard()), UInt(1));
+                    Store(IterationGuard(), steps);
                     var withinLimit = _module.AddInstruction(
                         SpirvOp.ULessThan,
                         _boolType,
@@ -625,7 +618,13 @@ public static partial class Gen5SpirvTranslator
             _scc = Flag(false);
             _vcc = Flag(false);
             _exec = Flag(true);
-            _reachedPixelExport = Flag(false);
+            // Only the title-shader trace reads it; elsewhere it would be a write-only Private
+            // global, and AMD's driver crashes linking stages that carry an unused one.
+            if (TraceTitleShaderState)
+            {
+                _reachedPixelExport = Flag(false);
+            }
+
             if (_usesPixelValidMask)
             {
                 _pixelValidMaskActive = Flag(true);
@@ -635,22 +634,7 @@ public static partial class Gen5SpirvTranslator
                 SpirvStorageClass.Private,
                 _module.Constant(_uintType, 0));
             _programActive = Flag(true);
-            if (UsesWave64Exchange())
-            {
-                _wave64ExchangeParity = _module.AddGlobalVariable(_privateUintPointer, SpirvStorageClass.Private, UInt(0));
-                _interfaces.Add(_wave64ExchangeParity);
-                _module.AddName(_wave64ExchangeParity, "wave64ExchangeParity");
-            }
-            if (_maxDispatcherSteps > 0)
-            {
-                _iterationGuard = _module.AddGlobalVariable(
-                    _privateUintPointer,
-                    SpirvStorageClass.Private,
-                    _module.Constant(_uintType, 0));
-                _interfaces.Add(_iterationGuard);
-                _module.AddName(_iterationGuard, "pcGuard");
-            }
-
+            _waveLowering.DeclareFunctionResources();
             foreach (var slot in FindLaneSpillSlots())
             {
                 var variable = _module.AddGlobalVariable(_privateUintPointer, SpirvStorageClass.Private, _module.Constant(_uintType, 0));
@@ -668,7 +652,11 @@ public static partial class Gen5SpirvTranslator
             _interfaces.Add(_scc);
             _interfaces.Add(_vcc);
             _interfaces.Add(_exec);
-            _interfaces.Add(_reachedPixelExport);
+            if (_reachedPixelExport != 0)
+            {
+                _interfaces.Add(_reachedPixelExport);
+            }
+
             if (_pixelValidMaskActive != 0)
             {
                 _interfaces.Add(_pixelValidMaskActive);
@@ -681,7 +669,7 @@ public static partial class Gen5SpirvTranslator
                 DeclareLayoutBindings();
                 DeclareScratch();
                 DeclareLds();
-                DeclareWave64Scratch();
+                _waveLowering.DeclareResources();
                 DeclareStageInterface();
                 return;
             }
@@ -709,43 +697,6 @@ public static partial class Gen5SpirvTranslator
             _module.AddName(_scratch, "scratch");
         }
 
-        private void DeclareWave64Scratch()
-        {
-            if (!UsesWave64Exchange())
-            {
-                return;
-            }
-
-            // Metal exposes 32 KiB of threadgroup memory on the Apple GPUs we
-            // target. Some PS5 compute shaders legitimately request all of it,
-            // so allocating another workgroup variable for the wave64 bridge
-            // makes pipeline creation fail. Reuse the final dwords of the
-            // existing LDS allocation in that case. The translator already
-            // bounds guest LDS accesses to this fixed allocation; keeping the
-            // bridge inside it preserves the host limit and still provides the
-            // cross-subgroup rendezvous needed to model one 64-lane guest wave.
-            if (_lds != 0)
-            {
-                _wave64Exchange = _lds;
-                _wave64ExchangeElementPointer = _ldsElementPointer;
-                _wave64ExchangeOffset = _ldsDwordMask + 1 < LdsDwordCount
-                    ? _ldsDwordMask + 1
-                    : LdsDwordCount - Wave64ExchangeDwordCount;
-                return;
-            }
-
-            var exchangeArrayType = _module.TypeArray(_uintType, Wave64ExchangeDwordCount);
-            _wave64ExchangeElementPointer = _module.TypePointer(SpirvStorageClass.Workgroup, _uintType);
-            _wave64Exchange = _module.AddGlobalVariable(
-                _module.TypePointer(SpirvStorageClass.Workgroup, exchangeArrayType),
-                SpirvStorageClass.Workgroup);
-            _wave64ExchangeOffset = 0;
-            _module.AddName(_wave64Exchange, "wave64Exchange");
-            _interfaces.Add(_wave64Exchange);
-        }
-
-        private bool UsesWave64Exchange() => _emulateWave64 && UsesSubgroupOperations();
-
         private uint ComputeLdsGuestDwordCount()
         {
             var declared = _request.LocalDataShareDwords;
@@ -776,9 +727,7 @@ public static partial class Gen5SpirvTranslator
                 ? ComputeLdsGuestDwordCount()
                 : PrivateLdsDwordCount;
             _ldsDwordMask = dwordCount - 1;
-            var arrayDwordCount = UsesWave64Exchange() && dwordCount < LdsDwordCount
-                ? dwordCount + Wave64ExchangeDwordCount
-                : dwordCount;
+            var arrayDwordCount = _waveLowering.ReserveLdsDwords(dwordCount);
 
             var ldsArrayType = _module.TypeArray(_uintType, arrayDwordCount);
             var ldsPointer = _module.TypePointer(storageClass, ldsArrayType);
@@ -1176,32 +1125,16 @@ public static partial class Gen5SpirvTranslator
             }
 
             Store(_scc, _module.ConstantBool(false));
-            Store(_reachedPixelExport, _module.ConstantBool(false));
+            if (_reachedPixelExport != 0)
+            {
+                Store(_reachedPixelExport, _module.ConstantBool(false));
+            }
+
             if (_pixelValidMaskActive != 0)
             {
                 Store(_pixelValidMaskActive, _module.ConstantBool(true));
             }
-            if (_subgroupInvocationIdInput != 0 && _emulateWave64)
-            {
-                StoreS64(106, BothHalvesOfOwnBallot(_module.ConstantBool(false)));
-                StoreS64(126, BothHalvesOfOwnBallot(_module.ConstantBool(true)));
-            }
-            else if (_subgroupInvocationIdInput != 0)
-            {
-                StoreWaveMask(106, _module.ConstantBool(false));
-                StoreWaveMask(126, _module.ConstantBool(true));
-            }
-            else
-            {
-                // Graphics stages emulate one logical wave lane. Keep the
-                // guest-visible VCC/EXEC scalar pairs synchronized with the
-                // internal booleans: shaders commonly save EXEC from s126:s127
-                // and restore it after divergent work. Initializing only _exec
-                // left those registers at zero, so the first restore disabled
-                // every fragment before its color export.
-                StoreS64(106, _module.Constant64(_ulongType, 0));
-                StoreS64(126, _module.Constant64(_ulongType, 1));
-            }
+            _waveLowering.InitializeMasks();
 
             // Primitive (NGG) vertex programs rebuild EXEC from the merged wave info in s3:
             // vertex count in bits 0-7, primitive count in bits 8-15. Left at zero, EXEC
@@ -1440,8 +1373,6 @@ public static partial class Gen5SpirvTranslator
             StoreS(register.Value, value);
         }
 
-        private enum SharedMemoryPhase { None, Read, Write }
-
         private bool TryEmitBlock(
             IReadOnlyList<ShaderBlock> blocks,
             int blockIndex,
@@ -1463,51 +1394,18 @@ public static partial class Gen5SpirvTranslator
         {
             error = string.Empty;
             var block = blocks[blockIndex];
-            var halfMaskPlan = HalfMaskPlan();
-            // One guest wave can span two host subgroups. Keep its shared-memory phases ordered.
-            // The half-mask plan inserts its own barriers; other wave64 programs also
-            // need LDS phase ordering when reads and writes span basic blocks.
-            var synchronizeSharedMemory = _emulateWave64 && halfMaskPlan is null;
-            var sharedMemoryPhase = SharedMemoryPhase.None;
+            _waveLowering.BeginBlock();
             for (var index = block.StartIndex; index < block.EndIndex; index++)
             {
                 var instruction = _request.Program.Instructions[index];
-                if (halfMaskPlan is not null)
-                {
-                    if (halfMaskPlan.ExactPairsBefore.TryGetValue(instruction.Pc, out var exactPairs))
-                    {
-                        EmitExactWaveMasks(exactPairs);
-                    }
-
-                    if (halfMaskPlan.SharedMemoryBarriersBefore.Contains(instruction.Pc))
-                    {
-                        EmitWave64Barrier();
-                    }
-                }
-
+                _waveLowering.BeforeInstruction(instruction);
                 if (IsBranch(instruction.Opcode) || instruction.Opcode == "SEndpgm")
                 {
                     continue;
                 }
 
-                if (synchronizeSharedMemory)
-                {
-                    var nextPhase = instruction.Control is Gen5DataShareControl { Gds: false }
-                        ? instruction.Opcode.StartsWith("DsRead", StringComparison.Ordinal) ? SharedMemoryPhase.Read
-                        : instruction.Opcode.StartsWith("DsWrite", StringComparison.Ordinal) ? SharedMemoryPhase.Write : SharedMemoryPhase.None
-                        : SharedMemoryPhase.None;
-                    if (instruction.Opcode == "SBarrier") sharedMemoryPhase = SharedMemoryPhase.None;
-                    if (nextPhase != SharedMemoryPhase.None)
-                    {
-                        if (sharedMemoryPhase != SharedMemoryPhase.None && sharedMemoryPhase != nextPhase) EmitWave64Barrier();
-                        sharedMemoryPhase = nextPhase;
-                    }
-                }
-
                 _execKnownFull = IsExecKnownFull(instruction.Pc);
-                _emittingPc = instruction.Pc;
                 var emitted = TryEmitInstruction(instruction, out error);
-                _emittingPc = null;
                 _execKnownFull = false;
                 if (!emitted)
                 {
@@ -1521,7 +1419,7 @@ public static partial class Gen5SpirvTranslator
                 CapturePixelExec(instruction);
             }
 
-            if (synchronizeSharedMemory && sharedMemoryPhase != SharedMemoryPhase.None) EmitWave64Barrier();
+            _waveLowering.EndBlock();
             return true;
         }
 
@@ -1908,8 +1806,8 @@ public static partial class Gen5SpirvTranslator
 
             if (_maxDispatcherSteps > 0)
             {
-                var steps = IAdd(Load(_uintType, _iterationGuard), UInt(1));
-                Store(_iterationGuard, steps);
+                var steps = IAdd(Load(_uintType, IterationGuard()), UInt(1));
+                Store(IterationGuard(), steps);
                 var withinLimit = _module.AddInstruction(SpirvOp.ULessThan, _boolType, steps, UInt((uint)_maxDispatcherSteps));
                 again = _module.AddInstruction(SpirvOp.LogicalAnd, _boolType, again, withinLimit);
             }
@@ -2063,8 +1961,8 @@ public static partial class Gen5SpirvTranslator
             {
                 // The dispatcher's safety valve, counted per loop iteration: a mistranslated
                 // exit ends the invocation instead of wedging the GPU queue.
-                var steps = IAdd(Load(_uintType, _iterationGuard), UInt(1));
-                Store(_iterationGuard, steps);
+                var steps = IAdd(Load(_uintType, IterationGuard()), UInt(1));
+                Store(IterationGuard(), steps);
                 var withinLimit = _module.AddInstruction(SpirvOp.ULessThan, _boolType, steps, UInt((uint)_maxDispatcherSteps));
                 again = _module.AddInstruction(SpirvOp.LogicalAnd, _boolType, again, withinLimit);
             }
@@ -2079,7 +1977,7 @@ public static partial class Gen5SpirvTranslator
                 // the block was skipped. Preserve that diagnostic safety budget.
                 if (_maxDispatcherSteps > 0)
                 {
-                    Store(_iterationGuard, IAdd(Load(_uintType, _iterationGuard), UInt(1)));
+                    Store(IterationGuard(), IAdd(Load(_uintType, IterationGuard()), UInt(1)));
                 }
                 _module.AddStatement(SpirvOp.Branch, entryMerge);
                 _module.AddLabel(entryMerge);
@@ -2657,15 +2555,7 @@ public static partial class Gen5SpirvTranslator
             });
 
             var firstValue = LoadV(destination);
-            uint broadcast;
-            if (_emulateWave64)
-            {
-                broadcast = ExchangeWave64Value(isFirstActive, firstValue);
-            }
-            else
-            {
-                broadcast = ShuffleLane(firstValue, firstLane);
-            }
+            var broadcast = BroadcastWaveValue(isFirstActive, firstValue, firstLane);
 
             var validResult = _module.AddInstruction(
                 SpirvOp.LogicalAnd,
@@ -6347,7 +6237,10 @@ public static partial class Gen5SpirvTranslator
                     return true;
                 }
 
-                Store(_reachedPixelExport, _module.ConstantBool(true));
+                if (_reachedPixelExport != 0)
+                {
+                    Store(_reachedPixelExport, _module.ConstantBool(true));
+                }
 
                 var values = new uint[4];
                 for (var component = 0; component < 4; component++)
@@ -7469,6 +7362,13 @@ public static partial class Gen5SpirvTranslator
                 UInt(0),
                 dwordAddress);
 
+        private uint DeclareFunctionUint(uint initializer)
+        {
+            var variable = _module.AddGlobalVariable(_privateUintPointer, SpirvStorageClass.Private, initializer);
+            _interfaces.Add(variable);
+            return variable;
+        }
+
         // Without subgroup operations each invocation is a one-lane wave, so a V_WRITELANE to
         // lane N > 0 used to be dropped and V_READLANE returned lane 0. Compilers spill SGPRs
         // (EXEC included) to VGPR lanes like this; restoring the wrong EXEC made waterfall
@@ -7476,7 +7376,7 @@ public static partial class Gen5SpirvTranslator
         private List<(uint Register, uint Lane)> FindLaneSpillSlots()
         {
             var slots = new List<(uint Register, uint Lane)>();
-            if (_emulateWave64)
+            if (_stage == Gen5SpirvStage.Compute && _waveLowering is EmulatedWave64Lowering)
             {
                 return slots;
             }
@@ -7686,6 +7586,24 @@ public static partial class Gen5SpirvTranslator
         // Declared on first use when Private. AMD's compiler keeps an unused 4 KiB private
         // array as a named .bss global: two stages then fail to link, and the driver copies
         // the NOBITS section as file data and reads past the end of the ELF.
+        private uint IterationGuard()
+        {
+            if (_iterationGuard == 0)
+            {
+                _iterationGuard = _module.AddGlobalVariable(
+                    _privateUintPointer,
+                    SpirvStorageClass.Private,
+                    _module.Constant(_uintType, 0));
+                _interfaces.Add(_iterationGuard);
+                _module.AddName(_iterationGuard, "pcGuard");
+            }
+
+            return _iterationGuard;
+        }
+
+        private static readonly bool TraceTitleShaderState =
+            Environment.GetEnvironmentVariable("SHARPEMU_TRACE_TITLE_SHADER_STATE") == "1";
+
         private uint PackedHalfRegisters()
         {
             if (_packedHalfRegisters != 0)
@@ -7956,18 +7874,19 @@ public static partial class Gen5SpirvTranslator
             _module.AddInstruction(SpirvOp.LogicalNot, _boolType, value);
 
         private uint SubgroupAny(uint condition) =>
-            _subgroupInvocationIdInput == 0
-                ? condition
-                : _emulateWave64
-                    ? IsNotZero64(BooleanToWaveMask(condition))
-                : _module.AddInstruction(
-                    SpirvOp.GroupNonUniformAny,
-                    _boolType,
-                    UInt(3),
-                    condition);
+            _waveLowering.SubgroupAny(condition);
 
         private uint GuestWaveLane()
         {
+            // A native wave64 is one host subgroup: its lane is the subgroup lane, the same one
+            // ballot and shuffle address, whatever order the driver gave the invocations.
+            if (_waveLaneCount == 64 && _guestLaneFromSubgroup && _subgroupInvocationIdInput != 0)
+            {
+                return BitwiseAnd(
+                    Load(_uintType, _subgroupInvocationIdInput),
+                    UInt(63));
+            }
+
             if (_waveLaneCount == 64 && _localInvocationIndexInput != 0)
             {
                 return BitwiseAnd(
@@ -8017,29 +7936,8 @@ public static partial class Gen5SpirvTranslator
                 ? value
                 : _module.AddInstruction(SpirvOp.GroupNonUniformShuffle, _uintType, UInt(3), value, lane);
 
-        private uint CurrentLaneBit()
-        {
-            if (_subgroupInvocationIdInput == 0)
-            {
-                return _module.Constant64(_ulongType, 1);
-            }
-
-            var maskedLane = GuestWaveLane();
-            var shifted = ShiftLeftLogical64(
-                _module.Constant64(_ulongType, 1),
-                _module.AddInstruction(
-                    SpirvOp.UConvert,
-                    _ulongType,
-                    maskedLane));
-            return _emulateWave64
-                ? shifted
-                : _module.AddInstruction(
-                    SpirvOp.Select,
-                    _ulongType,
-                    IsCurrentLaneInRdnaWave(),
-                    shifted,
-                    _module.Constant64(_ulongType, 0));
-        }
+        private uint CurrentLaneBit() =>
+            _waveLowering.CurrentLaneBit();
 
         private uint IsCurrentLaneInRdnaWave() =>
             _module.AddInstruction(
@@ -8056,193 +7954,11 @@ public static partial class Gen5SpirvTranslator
                 CurrentLaneBit(),
                 _module.Constant64(_ulongType, 0));
 
-        private uint BooleanToWaveMask(uint condition)
-        {
-            if (_subgroupInvocationIdInput == 0)
-            {
-                return BooleanToLaneMask(condition);
-            }
-
-            if (_emulateWave64)
-            {
-                var lane = GuestWaveLane();
-                var owned = OwnHalfBallot(condition);
-                var exchange = BeginWave64Exchange();
-                EmitConditional(IsHalfWaveLeader(lane), () =>
-                    Store(Wave64ExchangePointer(exchange, ShiftRightLogical(lane, UInt(5))), owned));
-                EmitWave64Barrier();
-                return Pair64(
-                    Load(_uintType, Wave64ExchangePointer(exchange, UInt(0))),
-                    Load(_uintType, Wave64ExchangePointer(exchange, UInt(1))));
-            }
-
-            var ballot = _module.AddInstruction(
-                SpirvOp.GroupNonUniformBallot,
-                _uvec4Type,
-                UInt(3),
-                condition);
-            var low = _module.AddInstruction(
-                SpirvOp.CompositeExtract,
-                _uintType,
-                ballot,
-                0);
-            var widened = _module.AddInstruction(SpirvOp.UConvert, _ulongType, low);
-            if (_waveLaneCount != 64)
-            {
-                return widened;
-            }
-
-            return _module.AddInstruction(
-                SpirvOp.Select,
-                _ulongType,
-                _module.AddInstruction(
-                    SpirvOp.UGreaterThanEqual,
-                    _boolType,
-                    GuestWaveLane(),
-                    UInt(32)),
-                ShiftLeftLogical64(
-                    widened,
-                    _module.Constant64(_ulongType, 32)),
-                widened);
-        }
-
-        private uint OwnHalfBallot(uint condition)
-        {
-            var ballot = _module.AddInstruction(
-                SpirvOp.GroupNonUniformBallot,
-                _uvec4Type,
-                UInt(3),
-                condition);
-            var component = ShiftRightLogical(Load(_uintType, _subgroupInvocationIdInput), UInt(5));
-            // Avoid dynamic extraction of subgroup ballot words: on the NVIDIA
-            // device tests it produced zero masks. Select the same word explicitly.
-            var result = _module.AddInstruction(SpirvOp.CompositeExtract, _uintType, ballot, 0);
-            for (uint index = 1; index < 4; index++)
-            {
-                var word = _module.AddInstruction(SpirvOp.CompositeExtract, _uintType, ballot, index);
-                result = _module.AddInstruction(SpirvOp.Select, _uintType,
-                    _module.AddInstruction(SpirvOp.IEqual, _boolType, component, UInt(index)), word, result);
-            }
-            return result;
-        }
-
-        private uint BooleanToHalfWaveMask(uint condition)
-        {
-            var owned = _module.AddInstruction(SpirvOp.UConvert, _ulongType, OwnHalfBallot(condition));
-            return _module.AddInstruction(
-                SpirvOp.Select,
-                _ulongType,
-                _module.AddInstruction(SpirvOp.UGreaterThanEqual, _boolType, GuestWaveLane(), UInt(32)),
-                ShiftLeftLogical64(owned, _module.Constant64(_ulongType, 32)),
-                owned);
-        }
-
-        private uint BothHalvesOfOwnBallot(uint condition)
-        {
-            var owned = _module.AddInstruction(SpirvOp.UConvert, _ulongType, OwnHalfBallot(condition));
-            return BitwiseOr64(owned, ShiftLeftLogical64(owned, _module.Constant64(_ulongType, 32)));
-        }
-
-        private uint IsHalfWaveLeader(uint lane) =>
-            _module.AddInstruction(SpirvOp.IEqual, _boolType, BitwiseAnd(lane, UInt(31)), UInt(0));
-
-        private uint BeginWave64Exchange()
-        {
-            var parity = Load(_uintType, _wave64ExchangeParity);
-            Store(_wave64ExchangeParity, BitwiseXor(parity, UInt(1)));
-            return IAdd(
-                UInt(_wave64ExchangeOffset),
-                _module.AddInstruction(SpirvOp.IMul, _uintType, parity, UInt(Wave64ExchangeSlotCount)));
-        }
-
-        private uint Wave64ExchangePointer(uint exchange, uint slot) =>
-            _module.AddInstruction(
-                SpirvOp.AccessChain,
-                _wave64ExchangeElementPointer,
-                _wave64Exchange,
-                IAdd(exchange, slot));
-
-        private uint ExchangeWave64Value(uint writer, uint value)
-        {
-            var exchange = BeginWave64Exchange();
-            EmitConditional(writer, () => Store(Wave64ExchangePointer(exchange, UInt(0)), value));
-            EmitWave64Barrier();
-            return Load(_uintType, Wave64ExchangePointer(exchange, UInt(0)));
-        }
-
-        private void EmitExactWaveMasks(IReadOnlyList<uint> pairs)
-        {
-            var lane = GuestWaveLane();
-            var upperHalf = ShiftRightLogical(lane, UInt(5));
-            var lowerHalf = _module.AddInstruction(SpirvOp.IEqual, _boolType, upperHalf, UInt(0));
-            const int pairsPerExchange = (int)Wave64ExchangeSlotCount / 2;
-            for (var start = 0; start < pairs.Count; start += pairsPerExchange)
-            {
-                var count = Math.Min(pairsPerExchange, pairs.Count - start);
-                var owned = new uint[count];
-                for (var index = 0; index < count; index++)
-                {
-                    var pair = pairs[start + index];
-                    owned[index] = _module.AddInstruction(SpirvOp.Select, _uintType, lowerHalf, LoadS(pair), LoadS(pair + 1));
-                }
-
-                var exchange = BeginWave64Exchange();
-                EmitConditional(IsHalfWaveLeader(lane), () =>
-                {
-                    for (var index = 0; index < count; index++)
-                    {
-                        Store(Wave64ExchangePointer(exchange, IAdd(UInt((uint)index * 2), upperHalf)), owned[index]);
-                    }
-                });
-                EmitWave64Barrier();
-                for (var index = 0; index < count; index++)
-                {
-                    var pair = pairs[start + index];
-                    StoreS(pair, Load(_uintType, Wave64ExchangePointer(exchange, UInt((uint)index * 2))));
-                    StoreS(pair + 1, Load(_uintType, Wave64ExchangePointer(exchange, UInt((uint)index * 2 + 1))));
-                }
-            }
-        }
+        private uint BooleanToWaveMask(uint condition) =>
+            _waveLowering.BooleanToWaveMask(condition);
 
         private uint WaveMaskAny(uint register, uint laneActive) =>
-            _emulateWave64 && _subgroupInvocationIdInput != 0
-                ? IsNotZero64(LoadS64(register))
-                : SubgroupAny(Load(_boolType, laneActive));
-
-        private static readonly bool HalfWaveMasks = !string.Equals(
-            Environment.GetEnvironmentVariable("SHARPEMU_WAVE64_HALF_MASKS"), "0", StringComparison.Ordinal);
-
-        private bool _halfMaskPlanBuilt;
-
-        private Ir.Gen5Wave64HalfMaskPlan? HalfMaskPlan()
-        {
-            if (_halfMaskPlanBuilt)
-            {
-                return _halfMaskPlan;
-            }
-
-            _halfMaskPlanBuilt = true;
-            if (HalfWaveMasks && UsesWave64Exchange() && _subgroupInvocationIdInput != 0)
-            {
-                var sharedFlatPcs = _request.Memory.Entries
-                    .Where(static memory => memory.AddressSpace is FlatAddressSpace.Shared or FlatAddressSpace.SharedOrPrivate)
-                    .Select(static memory => memory.Pc)
-                    .ToHashSet();
-                _halfMaskPlan = Ir.Gen5Wave64HalfMaskAnalysis.Analyze(_request.Program, sharedFlatPcs);
-            }
-
-            return _halfMaskPlan;
-        }
-
-        private void EmitWave64Barrier()
-        {
-            var workgroup = UInt(2);
-            _module.AddStatement(
-                SpirvOp.ControlBarrier,
-                workgroup,
-                workgroup,
-                UInt(0x108));
-        }
+            _waveLowering.WaveMaskAny(register, laneActive);
 
         // A wave-mask SGPR (VCC/EXEC) consumed as a per-lane predicate — the
         // condition of VCndmask, a VCC/EXEC branch, or the derived _vcc/_exec
@@ -8269,27 +7985,8 @@ public static partial class Gen5SpirvTranslator
 
         // In wave32 a lane mask fills its register alone and the next one keeps its value.
         // That includes VCC and EXEC: compilers use VCC_HI (s107) as an ordinary SGPR.
-        private void StoreWaveMask(uint register, uint condition)
-        {
-            if (_waveLaneCount == 32)
-            {
-                StoreS(register, Narrow(BooleanToWaveMask(condition)));
-                return;
-            }
-
-            if (_emulateWave64 &&
-                _subgroupInvocationIdInput != 0 &&
-                _emittingPc is { } pc &&
-                HalfMaskPlan() is { } plan &&
-                plan.HalfMaskWrites.TryGetValue(pc, out var halfRegister) &&
-                halfRegister == register)
-            {
-                StoreS64(register, BooleanToHalfWaveMask(condition));
-                return;
-            }
-
-            StoreS64(register, BooleanToWaveMask(condition));
-        }
+        private void StoreWaveMask(uint register, uint condition) =>
+            _waveLowering.StoreWaveMask(register, condition);
 
         private void EmitExecConditional(Action emit)
         {

src/SharpEmu.ShaderCompiler/Gen5PixelInputMapping.cs

@@ -9,8 +9,18 @@ public static class Gen5PixelInputMapping
         ReadOnlySpan<uint> controls,
         ReadOnlySpan<uint> activeInputs)
     {
-        Span<bool> usedLocations = stackalloc bool[32];
         var locations = new uint[activeInputs.Length];
+        ResolveLocations(controls, activeInputs, locations);
+        return locations;
+    }
+
+    // Fills `locations` (one slot per active input) without allocating.
+    public static void ResolveLocations(
+        ReadOnlySpan<uint> controls,
+        ReadOnlySpan<uint> activeInputs,
+        Span<uint> locations)
+    {
+        Span<bool> usedLocations = stackalloc bool[32];
 
         for (var index = 0; index < activeInputs.Length; index++)
         {
@@ -39,7 +49,5 @@ public static class Gen5PixelInputMapping
                 usedLocations[(int)location] = true;
             }
         }
-
-        return locations;
     }
 }

src/SharpEmu.ShaderCompiler/Resources/BindingLayout.cs

@@ -359,7 +359,20 @@ public sealed class BindingLayout : IEquatable<BindingLayout>
     public static uint NativeBindingIndex(ShaderStage stage, DescriptorBindingKind kind) =>
         (uint)kind + (stage == ShaderStage.Pixel ? (uint)DescriptorBindingKind.Count : 0);
 
-    public DescriptorBinding? Find(DescriptorBindingKind kind) => Descriptors.FirstOrDefault(binding => binding.Kind == kind);
+    // Called for every bound stage of every draw: an indexed loop allocates no closure or enumerator.
+    public DescriptorBinding? Find(DescriptorBindingKind kind)
+    {
+        var descriptors = Descriptors;
+        for (var index = 0; index < descriptors.Count; index++)
+        {
+            if (descriptors[index].Kind == kind)
+            {
+                return descriptors[index];
+            }
+        }
+
+        return null;
+    }
 
     // The user-data registers live at program entry: those some path reads before it
     // writes them, found by liveness over the control-flow graph, in ascending order.

src/SharpEmu.ShaderCompiler/ShaderCompileRequest.cs

@@ -162,6 +162,8 @@ public sealed class ShaderCompileRequest
     public IReadOnlyDictionary<int, uint> WrittenRangeSlotByMemoryIndex { get; }
 
     public uint WaveSize { get; init; } = 32;
+    // True when the pipeline runs on a 64-lane host subgroup, so one subgroup is one guest Wave64.
+    public bool HostWave64Supported { get; init; }
     public bool EnableExecGuardElision { get; init; } = true;
     public uint ScratchDwords { get; init; }
     public bool EnableGraphicsSubgroupOperations { get; init; } = true;

tests/SharpEmu.Libs.Tests/Gpu/GpuCommands/Registers/RegisterBanksSnapshotTests.cs

@@ -0,0 +1,236 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Gpu.GpuCommands;
+using SharpEmu.Libs.Gpu.GpuCommands.Registers;
+using Xunit;
+using static SharpEmu.Libs.Gpu.GpuCommands.Registers.ContextRegisterOffset;
+using static SharpEmu.Libs.Gpu.GpuCommands.Registers.ShaderRegisterOffset;
+using static SharpEmu.Libs.Gpu.GpuCommands.Registers.UserConfigRegisterOffset;
+
+namespace SharpEmu.Libs.Tests.Gpu.GpuCommands.Registers;
+
+// A snapshot shares the live banks' parts until the live banks write them; it never changes.
+public sealed class RegisterBanksSnapshotTests
+{
+    private const ulong PacketAddress = 0x1_0000_1000;
+
+    private static RegisterBanks NewBanks() => new(static message => new InvalidOperationException(message));
+
+    private static PacketContext Packet(uint opcode, int valueCount) =>
+        new(PacketHeader.Make((uint)valueCount + 2, opcode), PacketAddress, 0, 100, 100);
+
+    private static void WriteContext(RegisterBanks banks, uint offset, params uint[] values) =>
+        RegisterWriteTable.WriteContextPacket(banks, Packet(PacketOpcode.SetContextRegister, values.Length), offset, values);
+
+    private static void WriteShader(RegisterBanks banks, uint offset, params uint[] values) =>
+        RegisterWriteTable.WriteShaderPacket(banks, Packet(PacketOpcode.SetShaderRegister, values.Length), offset, values);
+
+    private static void WriteUserConfig(RegisterBanks banks, uint offset, params uint[] values) =>
+        RegisterWriteTable.WriteUserConfigPacket(banks, Packet(PacketOpcode.SetUserConfigRegister, values.Length), offset, values);
+
+    [Fact]
+    public void WritesAfterASnapshotLeaveItUnchanged()
+    {
+        var banks = NewBanks();
+        WriteContext(banks, DbDepthControl, 0x0000_0006);
+        WriteShader(banks, SpiShaderPgmLoPs, 0x1000);
+        WriteShader(banks, SpiShaderUserDataPs0, 11, 12);
+        WriteShader(banks, SpiShaderPgmLoEs, 0x2000);
+        WriteShader(banks, ComputePgmLo, 0x3000);
+        WriteUserConfig(banks, VgtPrimitiveType, 4);
+        var snapshot = banks.Snapshot();
+
+        WriteContext(banks, DbDepthControl, 0x0000_0000);
+        WriteShader(banks, SpiShaderPgmLoPs, 0x5000);
+        WriteShader(banks, SpiShaderUserDataPs0, 21, 22);
+        WriteShader(banks, SpiShaderPgmLoEs, 0x6000);
+        WriteShader(banks, ComputePgmLo, 0x7000);
+        WriteUserConfig(banks, VgtPrimitiveType, 6);
+
+        Assert.True(snapshot.Context.DepthTarget.DepthTestEnabled);
+        Assert.False(banks.Context.DepthTarget.DepthTestEnabled);
+        Assert.Equal(0x1000ul << 8, snapshot.Shader.Pixel.Address);
+        Assert.Equal(0x5000ul << 8, banks.Shader.Pixel.Address);
+        Assert.Equal([11u, 12u], snapshot.Shader.Pixel.UserScalars.Values[..2]);
+        Assert.Equal([21u, 22u], banks.Shader.Pixel.UserScalars.Values[..2]);
+        Assert.Equal(0x2000ul << 8, snapshot.Shader.Vertex.ExportAddress);
+        Assert.Equal(0x6000ul << 8, banks.Shader.Vertex.ExportAddress);
+        Assert.Equal(0x3000ul << 8, snapshot.Shader.Compute.Address);
+        Assert.Equal(0x7000ul << 8, banks.Shader.Compute.Address);
+        Assert.Equal(4u, snapshot.UserConfig.PrimitiveType);
+        Assert.Equal(6u, banks.UserConfig.PrimitiveType);
+    }
+
+    [Fact]
+    public void AWriteCopiesOnlyThePartItStoresInto()
+    {
+        var banks = NewBanks();
+        var snapshot = banks.Snapshot();
+
+        WriteShader(banks, SpiShaderUserDataPs0, 7);
+
+        Assert.NotSame(snapshot.Shader.Pixel, banks.Shader.Pixel);
+        Assert.Same(snapshot.Shader.Vertex, banks.Shader.Vertex);
+        Assert.Same(snapshot.Shader.Compute, banks.Shader.Compute);
+        Assert.Same(snapshot.Context, banks.Context);
+        Assert.Same(snapshot.UserConfig, banks.UserConfig);
+    }
+
+    [Fact]
+    public void AUserScalarWriteCopiesOnlyItsSet()
+    {
+        var banks = NewBanks();
+        WriteShader(banks, SpiShaderUserDataGs0, 1);
+        WriteShader(banks, SpiShaderUserDataHs0, 2);
+        var snapshot = banks.Snapshot();
+
+        WriteShader(banks, SpiShaderUserDataGs0, 3);
+        WriteShader(banks, SpiShaderPgmLoPs, 0x4000);
+
+        Assert.NotSame(snapshot.Shader.Vertex, banks.Shader.Vertex);
+        Assert.NotSame(snapshot.Shader.Vertex.GeometryUserScalars, banks.Shader.Vertex.GeometryUserScalars);
+        Assert.Same(snapshot.Shader.Vertex.HullUserScalars, banks.Shader.Vertex.HullUserScalars);
+        Assert.Equal(1u, snapshot.Shader.Vertex.GeometryUserScalars.Values[0]);
+        Assert.Equal(3u, banks.Shader.Vertex.GeometryUserScalars.Values[0]);
+        Assert.NotSame(snapshot.Shader.Pixel, banks.Shader.Pixel);
+        Assert.Same(snapshot.Shader.Pixel.UserScalars, banks.Shader.Pixel.UserScalars);
+
+        // The set copied after the snapshot is the live banks' own: a later write does not copy it again.
+        var geometry = banks.Shader.Vertex.GeometryUserScalars;
+        WriteShader(banks, SpiShaderUserDataGs0, 4);
+        Assert.Same(geometry, banks.Shader.Vertex.GeometryUserScalars);
+        Assert.Equal(1u, snapshot.Shader.Vertex.GeometryUserScalars.Values[0]);
+    }
+
+    [Fact]
+    public void ContextStateOperationsAfterASnapshotLeaveItUnchanged()
+    {
+        var banks = NewBanks();
+        WriteContext(banks, DbDepthControl, 0x0000_0006);
+        var snapshot = banks.Snapshot();
+
+        banks.ApplyContextState(ContextStateOperation.PushClear);
+        WriteContext(banks, DbDepthControl, 0x0000_0000);
+        banks.ApplyContextState(ContextStateOperation.Pop);
+        WriteContext(banks, DbDepthControl, 0x0000_0000);
+
+        Assert.True(snapshot.Context.DepthTarget.DepthTestEnabled);
+        Assert.False(banks.Context.DepthTarget.DepthTestEnabled);
+    }
+
+    // Every context register, through the table path and the packet path, against a snapshot
+    // taken after a first write: the context's nested parts are copied only by their writers.
+    [Fact]
+    public void EveryContextRegisterWriteLeavesASnapshotUnchanged()
+    {
+        var changed = 0;
+        for (var offset = 0u; offset < RegisterBankLayout.ContextRegisterCount; offset++)
+        {
+            foreach (var packet in new[] { false, true })
+            {
+                var banks = NewBanks();
+                WriteContext(banks, DbDepthControl, 0x0000_0006);
+                var snapshot = banks.Snapshot();
+                var before = Describe(snapshot.Context);
+                foreach (var value in new[] { 0xFFFF_FFFFu, 0x1234_5678u })
+                {
+                    try
+                    {
+                        if (packet)
+                            WriteContext(banks, offset, value);
+                        else
+                            RegisterWriteTable.WriteContextEntry(banks, offset, value, PacketAddress);
+                    }
+                    catch (InvalidOperationException)
+                    {
+                        // A register whose writer refuses a single value.
+                    }
+                }
+
+                Assert.True(before == Describe(snapshot.Context), $"offset=0x{offset:X} packet={packet}");
+                if (Describe(banks.Context) != before)
+                    changed++;
+            }
+        }
+
+        // The description sees the writes: most registers store something.
+        Assert.True(changed > 400, $"changed={changed}");
+    }
+
+    private static string Describe(object? value)
+    {
+        var text = new System.Text.StringBuilder();
+        Describe(value, text);
+        return text.ToString();
+    }
+
+    private static void Describe(object? value, System.Text.StringBuilder text)
+    {
+        switch (value)
+        {
+            case null:
+                text.Append("null;");
+                return;
+            case System.Collections.IDictionary dictionary:
+                foreach (var key in dictionary.Keys.Cast<object>().OrderBy(key => key))
+                    text.Append(key).Append('=').Append(dictionary[key]).Append(',');
+                text.Append(';');
+                return;
+            case Array array:
+                text.Append('[');
+                foreach (var element in array)
+                    Describe(element, text);
+                text.Append(']');
+                return;
+        }
+
+        var type = value.GetType();
+        if (type.IsPrimitive || type.IsEnum || value is string)
+        {
+            text.Append(value).Append(';');
+            return;
+        }
+
+        text.Append('{');
+        foreach (var field in type.GetFields(System.Reflection.BindingFlags.Instance |
+            System.Reflection.BindingFlags.Public | System.Reflection.BindingFlags.NonPublic))
+        {
+            text.Append(field.Name).Append(':');
+            Describe(field.GetValue(value), text);
+        }
+
+        text.Append('}');
+    }
+
+    [Fact]
+    public void AContextWriteCopiesOnlyTheNestedPartsItStoresInto()
+    {
+        var banks = NewBanks();
+        var snapshot = banks.Snapshot();
+
+        WriteContext(banks, DbDepthControl, 0x0000_0006);
+        WriteContext(banks, PaClVteCntl, 0x43F);
+
+        Assert.NotSame(snapshot.Context, banks.Context);
+        Assert.NotSame(snapshot.Context.ScreenViewport, banks.Context.ScreenViewport);
+        Assert.Same(snapshot.Context.ScreenViewport.Viewports, banks.Context.ScreenViewport.Viewports);
+        Assert.Same(snapshot.Context.ShaderInterface, banks.Context.ShaderInterface);
+        Assert.Same(snapshot.Context.ColorTargets, banks.Context.ColorTargets);
+        Assert.Same(snapshot.Context.UnmodeledTableRegisters, banks.Context.UnmodeledTableRegisters);
+    }
+
+    [Fact]
+    public void ASecondSnapshotSharesAgainUntilTheNextWrite()
+    {
+        var banks = NewBanks();
+        var first = banks.Snapshot();
+        WriteContext(banks, DbDepthControl, 0x0000_0006);
+        var second = banks.Snapshot();
+        WriteContext(banks, DbDepthControl, 0x0000_0000);
+
+        Assert.False(first.Context.DepthTarget.DepthTestEnabled);
+        Assert.True(second.Context.DepthTarget.DepthTestEnabled);
+        Assert.False(banks.Context.DepthTarget.DepthTestEnabled);
+    }
+}

tests/SharpEmu.Libs.Tests/Gpu/Images/GuestImageCacheTests.Metadata.cs

@@ -99,6 +99,24 @@ public sealed partial class GuestImageCacheTests
         harness.Shutdown();
     }
 
+    // A fill seen before its target leaves its range known as DCC metadata, so the next fill of it
+    // runs on the host and the range stays CPU-readable.
+    [Fact]
+    public void APendingDccFill_MarksItsRangeAsDccMetadata()
+    {
+        if (!GatePrerequisites.Ready(_vulkan)) return;
+        using var fatal = new FatalScope();
+        using var harness = new CacheHarness(_vulkan);
+        var address = harness.MapBacked(0x20000, ReadWrite);
+        var metadata = address + 0x10000;
+        Assert.False(harness.Images.OverlapsDccMetadata(metadata, 0x100));
+
+        Assert.False(harness.Images.TryAbsorbDccFill(metadata, 0x100, 0x00000000u));
+
+        Assert.True(harness.Images.OverlapsDccMetadata(metadata + 0x80, 0x10));
+        Assert.False(harness.Images.OverlapsDccMetadata(metadata + 0x100, 0x10));
+    }
+
     [Fact]
     public void DccFills_RemainGuestProducersUntilDescriptorDiscovery()
     {
@@ -451,6 +469,35 @@ public sealed partial class GuestImageCacheTests
         harness.Shutdown();
     }
 
+    // The pages inside a filled range are clean after it; a page it shares with another GPU-written
+    // range keeps that range's state.
+    [Fact]
+    public void DirtyDccFill_KeepsASharedBoundaryPageDirty()
+    {
+        if (!GatePrerequisites.Ready(_vulkan)) return;
+        using var harness = new CacheHarness(_vulkan);
+        const ulong Page = SharpEmu.HLE.GpuMemory.TrackerLayout.PageBytes;
+        var memory = harness.MapBacked(8 * Page, ReadWrite);
+        var metadata = memory + Page + Page / 2;
+        var size = 3 * Page;
+        var neighbour = metadata - 0x100;
+        harness.Images.RegisterDccMetadataForTest(metadata, size);
+        harness.Worker.Run(() =>
+        {
+            _ = harness.Cache.ObtainBuffer(neighbour, 0x100, isWritten: true);
+            _ = harness.Cache.ObtainBuffer(metadata, size, isWritten: true);
+        });
+
+        harness.Worker.Run(() => harness.Cache.FillBuffer(metadata, size, 0x00000000, isGds: false));
+
+        Assert.False(harness.Cache.HasGpuDirtyBytes(metadata, size));
+        Assert.False(harness.Cache.HasGpuDirtyPages(memory + 2 * Page, 2 * Page));
+        Assert.False(harness.Cache.HasGpuDirtyPages(memory + 4 * Page, Page));
+        Assert.True(harness.Cache.HasGpuDirtyPages(memory + Page, Page));
+        Assert.True(harness.Cache.HasGpuDirtyBytes(neighbour, 0x100));
+        harness.Shutdown();
+    }
+
     [Fact]
     public void CleanDccFill_WritesBothCopiesWithoutMarkingTheGpuCopyNewer()
     {

tests/SharpEmu.Libs.Tests/Gpu/Pipelines/ShaderPipelineCacheTests.cs

@@ -266,6 +266,9 @@ public sealed class ShaderPipelineCacheTests : IDisposable
         executor.DrawAuto(1, Banks(), Auto(3));
         executor.DrawAuto(2, Banks(), Auto(3));
         executor.DrawAuto(3, Banks(PrimitiveTriangleStrip), Auto(4));
+        // The lookup key is reused across draws: after another pipeline, the first is found again.
+        executor.DrawAuto(4, Banks(), Auto(3));
+        executor.DrawAuto(5, Banks(PrimitiveTriangleStrip), Auto(4));
 
         Assert.Equal(2, guest.Host.GraphicsPipelines.Count);
         Assert.Equal(2, cache.GraphicsPipelineCount);

tests/SharpEmu.Libs.Tests/Gpu/Rendering/RenderExecutorComputeTests.cs

@@ -279,9 +279,25 @@ public sealed class RenderExecutorComputeTests : IDisposable
 
         Assert.Contains($"clear_image {MetadataAddress:X} 1000 {ClearValue:X8}", _host.Calls);
         Assert.Contains($"absorb_dcc {MetadataAddress:X} 1000 {ClearValue:X8}", _host.Calls);
+        Assert.Contains($"fill_dcc {MetadataAddress:X} 1000 {ClearValue:X8}", _host.Calls);
         AssertDispatched(4, 1, 1);
     }
 
+    // A pending surface's metadata the host knows is filled there: as a dispatch it would be GPU-written
+    // and read back at every later bind.
+    [Fact]
+    public void PendingMetadataTheHostKnows_IsFilledOnTheHost()
+    {
+        _host.HostFillableDcc.Add(MetadataAddress);
+        ConfigureClearKernel(256);
+        _executor.Dispatch(1, Banks(), 256, 1, 1, ClearInitiator);
+
+        Assert.Contains($"absorb_dcc {MetadataAddress:X} 1000 {ClearValue:X8}", _host.Calls);
+        Assert.Contains($"fill_dcc {MetadataAddress:X} 1000 {ClearValue:X8}", _host.Calls);
+        AssertNotDispatched();
+        Assert.Contains("reset_bindings", _host.Calls);
+    }
+
     [Fact]
     public void RegisteredDccAddress_ConsumesTheFillAndRecordsNothing()
     {

tests/SharpEmu.Libs.Tests/Gpu/Rendering/RenderExecutorDrawTests.cs

@@ -22,9 +22,21 @@ public sealed class RenderExecutorDrawTests : IDisposable
     private readonly RenderExecutor _executor;
     private readonly FatalScope _fatal = new();
 
-    public RenderExecutorDrawTests() => _executor = new RenderExecutor(_host, _pipelines);
+    private readonly InputsCheckingPrefetch _prefetch;
 
-    public void Dispose() => _fatal.Dispose();
+    public RenderExecutorDrawTests()
+    {
+        _executor = new RenderExecutor(_host, _pipelines);
+        _prefetch = new InputsCheckingPrefetch(_host);
+        _executor.Prefetch = _prefetch;
+    }
+
+    // Every draw of these tests also checks that its program inputs follow from its registers.
+    public void Dispose()
+    {
+        _fatal.Dispose();
+        Assert.Empty(_prefetch.Mismatches);
+    }
 
     private void AssertOrder(params string[] prefixes)
     {

tests/SharpEmu.Libs.Tests/Gpu/Rendering/RenderExecutorPrefetchTests.cs

@@ -0,0 +1,87 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Gpu.GpuCommands.Registers;
+using SharpEmu.Libs.Gpu.Rendering;
+using SharpEmu.Libs.Tests.Gpu.Scheduling;
+using Xunit;
+using static SharpEmu.Libs.Tests.Gpu.Rendering.RenderExecutorFixtures;
+
+namespace SharpEmu.Libs.Tests.Gpu.Rendering;
+
+// A draw takes the programs a prefetch resolved from its banks, or resolves them itself.
+[Collection(SchedulingStateCollection.Name)]
+public sealed class RenderExecutorPrefetchTests : IDisposable
+{
+    private readonly RecordingRenderHost _host = new();
+    private readonly FakePipelineProvider _pipelines = new();
+    private readonly RenderExecutor _executor;
+    private readonly FatalScope _fatal = new();
+
+    public RenderExecutorPrefetchTests() => _executor = new RenderExecutor(_host, _pipelines);
+
+    public void Dispose() => _fatal.Dispose();
+
+    // Resolves ahead with the executor, as a worker would, and hands the result back once.
+    private sealed class AheadPrefetch(RenderExecutor executor, IRenderHost host) : IProgramPrefetch
+    {
+        private RegisterBanks? _banks;
+        private GraphicsProgramInputs _inputs;
+        private GraphicsPrograms? _programs;
+
+        public int Taken { get; private set; }
+
+        public int Declined { get; private set; }
+
+        public void Resolve(RegisterBanks banks)
+        {
+            _banks = banks;
+            _inputs = RenderExecutor.ProgramInputsOf(banks, host.FormatSupport, host.Fatal);
+            _programs = executor.ResolveGraphicsPrograms(banks, in _inputs);
+        }
+
+        public bool TryTakeGraphics(RegisterBanks banks, in GraphicsProgramInputs inputs, out GraphicsPrograms programs)
+        {
+            if (_programs is not null && ReferenceEquals(banks, _banks) && inputs.Equals(_inputs))
+            {
+                programs = _programs;
+                _programs = null;
+                Taken++;
+                return true;
+            }
+
+            programs = null!;
+            Declined++;
+            return false;
+        }
+    }
+
+    [Fact]
+    public void ADrawTakesTheProgramsResolvedAheadFromItsBanks()
+    {
+        var prefetch = new AheadPrefetch(_executor, _host);
+        _executor.Prefetch = prefetch;
+        var banks = Banks();
+        prefetch.Resolve(banks);
+        var resolutions = _pipelines.Calls.Count(call => call.StartsWith("get_graphics_programs", StringComparison.Ordinal));
+
+        _executor.DrawAuto(1, banks, Auto(3));
+
+        Assert.Equal(1, prefetch.Taken);
+        Assert.Equal(resolutions, _pipelines.Calls.Count(call => call.StartsWith("get_graphics_programs", StringComparison.Ordinal)));
+        Assert.Contains(_host.Calls, call => call.StartsWith("draw ", StringComparison.Ordinal));
+    }
+
+    [Fact]
+    public void ADrawWithOtherBanksResolvesItsOwnPrograms()
+    {
+        var prefetch = new AheadPrefetch(_executor, _host);
+        _executor.Prefetch = prefetch;
+        prefetch.Resolve(Banks());
+
+        _executor.DrawAuto(1, Banks(), Auto(3));
+
+        Assert.Equal((0, 1), (prefetch.Taken, prefetch.Declined));
+        Assert.Equal(2, _pipelines.Calls.Count(call => call.StartsWith("get_graphics_programs", StringComparison.Ordinal)));
+    }
+}

tests/SharpEmu.Libs.Tests/Gpu/Rendering/RenderExecutorTestSupport.cs

@@ -108,6 +108,9 @@ internal sealed class RecordingRenderHost : IRenderHost
 
     public HashSet<ulong> RegisteredDcc { get; } = new();
 
+    // Metadata the host knows and fills itself although its surface is not registered yet.
+    public HashSet<ulong> HostFillableDcc { get; } = new();
+
     public HashSet<ulong> ClearableImages { get; } = new();
 
     public Func<ulong, ulong, ulong>? ClampOverride { get; set; }
@@ -391,7 +394,7 @@ internal sealed class RecordingRenderHost : IRenderHost
     public bool TryFillDccMetadata(ulong address, ulong size, uint fillValue)
     {
         Calls.Add($"fill_dcc {address:X} {size:X} {fillValue:X8}");
-        return RegisteredDcc.Contains(address);
+        return RegisteredDcc.Contains(address) || HostFillableDcc.Contains(address);
     }
 
     public bool HostCopyAccepted { get; set; } = true;
@@ -405,6 +408,33 @@ internal sealed class RecordingRenderHost : IRenderHost
     public Exception Fatal(string message) => new RenderExecutorFatalException(message);
 }
 
+// Takes nothing, but checks that the inputs a draw resolves its programs with are the ones its
+// register banks alone give: a prefetch computed ahead of the draw would use those.
+internal sealed class InputsCheckingPrefetch(IRenderHost host) : IProgramPrefetch
+{
+    public List<string> Mismatches { get; } = new();
+
+    public int Checks { get; private set; }
+
+    public bool TryTakeGraphics(RegisterBanks banks, in GraphicsProgramInputs inputs, out GraphicsPrograms programs)
+    {
+        Checks++;
+        var expected = RenderExecutor.ProgramInputsOf(banks, host.FormatSupport, host.Fatal);
+        if (!expected.Equals(inputs))
+        {
+            Mismatches.Add(
+                $"draw pixelActive={inputs.PixelActive}/{expected.PixelActive} depthBound={inputs.DepthBound}/{expected.DepthBound} " +
+                $"mapping=[{Describe(inputs.Mapping)}]/[{Describe(expected.Mapping)}]");
+        }
+
+        programs = null!;
+        return false;
+    }
+
+    private static string Describe(ReadOnlySpan<ColorComponentMap> mapping) =>
+        string.Join(",", mapping.ToArray().Select(map => map.Packed));
+}
+
 // Hands back the configured programs and records every pipeline request.
 internal sealed class FakePipelineProvider : IShaderPipelineProvider
 {

tests/SharpEmu.Libs.Tests/VideoOut/PresenterImageBindingTests.cs

@@ -92,6 +92,41 @@ public sealed class PresenterImageBindingTests : IClassFixture<HeadlessVulkanFix
         harness.Shutdown();
     }
 
+    [Fact]
+    public void RepeatedBinding_ReusesTheLookupUntilTheImageIsRemoved()
+    {
+        if (!GatePrerequisites.Ready(_vulkan)) return;
+        using var fatal = new FatalScope();
+        using var presenter = new PresenterUnderTest(_vulkan);
+        var harness = presenter.Harness;
+        var address = harness.MapBacked(0x10000, ReadWrite);
+        harness.Write(address, new byte[0x10000]);
+        presenter.Run(() =>
+        {
+            var resource = new ImageResource
+            {
+                ResourceClass = PlanImageResourceClass.Sampled,
+                NumericClass = ImageNumericClass.Float,
+                Dimension = ImageDimension.Dim2D,
+                Read = true,
+            };
+            var program = new ShaderProgramInfo();
+            var words = RegisterWords.Texture(address, GuestPixelFormat.Bits32Float, 48, 24);
+            var first = (ResourceSlotIdentifier)GetFieldValue(presenter.InvokeMethod("ResolveImageBinding", resource, words, program, 0)!, "ImageIdentifier");
+            var second = (ResourceSlotIdentifier)GetFieldValue(presenter.InvokeMethod("ResolveImageBinding", resource, words, program, 0)!, "ImageIdentifier");
+            Assert.Equal(first, second);
+
+            // A removed image is not served again: the binding looks the descriptor up afresh.
+            var generation = harness.Images.LookupGeneration;
+            typeof(GuestImageCache).GetMethod("DeleteImage", InstanceMembers)!.Invoke(harness.Images, [first]);
+            Assert.NotEqual(generation, harness.Images.LookupGeneration);
+            var third = (ResourceSlotIdentifier)GetFieldValue(presenter.InvokeMethod("ResolveImageBinding", resource, words, program, 0)!, "ImageIdentifier");
+            Assert.True(harness.Images.TryGetImage(third, out var image) && image.Registered);
+            presenter.RenderHost.ResetBindings();
+        });
+        harness.Shutdown();
+    }
+
     [Theory]
     [InlineData(true, 1u, false)]
     [InlineData(true, 2u, true)]

tests/SharpEmu.ShaderCompiler.Tests/Gen5WaveLoweringTests.cs

@@ -0,0 +1,250 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using System.Text;
+using SharpEmu.ShaderCompiler.Vulkan;
+using Xunit;
+using static SharpEmu.ShaderCompiler.Tests.Resources.ResourceTestProgram;
+
+namespace SharpEmu.ShaderCompiler.Tests;
+
+// Checks the SPIR-V each wave lowering (NativeWave32, NativeWave64, EmulatedWave64) produces for the
+// same guest programs: which subgroup instructions appear, how a ballot is split into a 64-bit mask,
+// and whether the workgroup exchange / rendezvous machinery exists.
+public sealed class Gen5WaveLoweringTests
+{
+    public enum Lowering { Wave32, NativeWave64, EmulatedWave64 }
+
+    private sealed record Module(
+        IReadOnlyList<(SpirvOp Op, uint[] Operands)> Instructions,
+        IReadOnlyList<string> Names)
+    {
+        public int Count(SpirvOp op) => Instructions.Count(instruction => instruction.Op == op);
+
+        // CompositeExtract of a ballot result: its literal index is the last operand.
+        public IReadOnlyList<uint> BallotExtractIndices()
+        {
+            var ballots = Instructions.Where(i => i.Op == SpirvOp.GroupNonUniformBallot).Select(i => i.Operands[1]).ToHashSet();
+            return Instructions
+                .Where(i => i.Op == SpirvOp.CompositeExtract && ballots.Contains(i.Operands[2]))
+                .Select(i => i.Operands[^1])
+                .ToList();
+        }
+    }
+
+    private static Module Parse(byte[] spirv)
+    {
+        var instructions = new List<(SpirvOp, uint[])>();
+        var names = new List<string>();
+        for (var offset = 20; offset < spirv.Length;)
+        {
+            var word = BinaryPrimitives.ReadUInt32LittleEndian(spirv.AsSpan(offset));
+            var length = (int)(word >> 16);
+            var op = (SpirvOp)(word & 0xFFFF);
+            var operands = new uint[length - 1];
+            for (var index = 0; index < operands.Length; index++)
+            {
+                operands[index] = BinaryPrimitives.ReadUInt32LittleEndian(spirv.AsSpan(offset + 4 + index * 4));
+            }
+
+            instructions.Add((op, operands));
+            if (op == SpirvOp.Name)
+            {
+                var text = new List<byte>();
+                for (var index = 1; index < operands.Length; index++)
+                {
+                    text.AddRange(BitConverter.GetBytes(operands[index]));
+                }
+
+                names.Add(Encoding.UTF8.GetString(text.ToArray()).TrimEnd('\0'));
+            }
+
+            offset += length * 4;
+        }
+
+        return new Module(instructions, names);
+    }
+
+    private static Module Compile(Gen5ShaderProgram program, Lowering lowering, uint threads = 64)
+    {
+        var (plan, resources, layout) = Prepare(program);
+        var configured = new ShaderCompileRequest(plan, resources, layout)
+        {
+            WaveSize = lowering == Lowering.Wave32 ? 32u : 64u,
+            HostWave64Supported = lowering == Lowering.NativeWave64,
+            LocalSizeX = threads,
+        };
+        Assert.True(Gen5SpirvTranslator.TryCompileProgram(configured, out var shader, out var error), error);
+        return Parse(shader.Spirv);
+    }
+
+    // v_cmp_lt_u32 vcc, v0, v1 : a per-lane compare whose result becomes a wave mask.
+    private static Gen5ShaderProgram CompareProgram() => Program(
+        Vopc(0, "VCmpLtU32", Operand(5), 1),
+        EndProgram(4));
+
+    private static Gen5ShaderProgram ReadFirstLaneProgram() => Program(
+        ReadFirstLane(0, 4, 0),
+        EndProgram(4));
+
+    private static Gen5ShaderProgram ReadLaneProgram(uint lane) => Program(
+        ReadLane(0, 4, 0, lane),
+        EndProgram(4));
+
+    private static Gen5ShaderProgram DataShareProgram(bool barrierBetween) =>
+        Program(
+            new[]
+            {
+                DataShare(0, "DsWriteB32", false, [Gen5Operand.Vector(0), Gen5Operand.Vector(1)], [], 0, 0),
+            }
+            .Concat(barrierBetween
+                ? new[] { new Gen5ShaderInstruction(4, Gen5ShaderEncoding.Sopp, "SBarrier", [0u], [], [], null) }
+                : [])
+            .Concat(new[]
+            {
+                DataShare(8, "DsReadB32", false, [Gen5Operand.Vector(0)], [2u], 0, 0),
+                EndProgram(16),
+            })
+            .ToArray());
+
+    [Theory]
+    [InlineData(Lowering.Wave32)]
+    [InlineData(Lowering.NativeWave64)]
+    [InlineData(Lowering.EmulatedWave64)]
+    public void EveryLoweringCompilesTheBasicWavePrograms(Lowering lowering)
+    {
+        foreach (var program in new[] { CompareProgram(), ReadFirstLaneProgram(), ReadLaneProgram(0), ReadLaneProgram(33), DataShareProgram(false), DataShareProgram(true) })
+        {
+            Compile(program, lowering);
+        }
+    }
+
+    [Fact]
+    public void Wave32_BallotUsesOnlyTheLowWord()
+    {
+        var module = Compile(CompareProgram(), Lowering.Wave32);
+        Assert.True(module.Count(SpirvOp.GroupNonUniformBallot) > 0);
+        Assert.All(module.BallotExtractIndices(), index => Assert.Equal(0u, index));
+    }
+
+    [Fact]
+    public void NativeWave64_BallotCombinesBothWords()
+    {
+        var module = Compile(CompareProgram(), Lowering.NativeWave64);
+        var indices = module.BallotExtractIndices();
+        Assert.Contains(0u, indices);
+        Assert.Contains(1u, indices);
+    }
+
+    [Fact]
+    public void OnlyTheEmulatedLoweringDeclaresTheExchange()
+    {
+        foreach (var lowering in new[] { Lowering.Wave32, Lowering.NativeWave64 })
+        {
+            var module = Compile(ReadFirstLaneProgram(), lowering);
+            Assert.DoesNotContain("wave64Exchange", module.Names);
+            Assert.DoesNotContain("wave64ExchangeParity", module.Names);
+        }
+
+        var emulated = Compile(ReadFirstLaneProgram(), Lowering.EmulatedWave64);
+        Assert.Contains("wave64ExchangeParity", emulated.Names);
+        Assert.Contains("wave64Exchange", emulated.Names);
+    }
+
+    [Fact]
+    public void EmulatedWave64_NeedsNoExchangeWithoutSubgroupOperations()
+    {
+        var module = Compile(Program(Vop2(0, "VAddI32", 1, Operand(1), Gen5Operand.Vector(0)), EndProgram(4)), Lowering.EmulatedWave64);
+        Assert.DoesNotContain("wave64ExchangeParity", module.Names);
+    }
+
+    [Theory]
+    [InlineData(Lowering.Wave32, 64u)]
+    [InlineData(Lowering.NativeWave64, 64u)]
+    [InlineData(Lowering.NativeWave64, 256u)]
+    public void NativeLoweringsAddNoBarriersAroundSharedMemory(Lowering lowering, uint threads)
+    {
+        var module = Compile(DataShareProgram(false), lowering, threads);
+        Assert.Equal(0, module.Count(SpirvOp.ControlBarrier));
+    }
+
+    [Theory]
+    [InlineData(false, 2)]
+    [InlineData(true, 2)]
+    public void EmulatedWave64_OrdersSharedMemoryPhasesWithinOneWave(bool explicitBarrier, int expectedBarriers)
+    {
+        var module = Compile(DataShareProgram(explicitBarrier), Lowering.EmulatedWave64, threads: 64);
+        Assert.Equal(expectedBarriers, module.Count(SpirvOp.ControlBarrier));
+    }
+
+    [Fact]
+    public void EmulatedWave64_MultiWaveGroupsUseNoWorkgroupBarrierForSharedMemoryPhases()
+    {
+        var module = Compile(DataShareProgram(false), Lowering.EmulatedWave64, threads: 128);
+        Assert.Equal(0, module.Count(SpirvOp.ControlBarrier));
+    }
+
+    [Fact]
+    public void EmulatedWave64_MultiWaveExchangeMeetsAtAnArrivalCounter()
+    {
+        var single = Compile(ReadFirstLaneProgram(), Lowering.EmulatedWave64, threads: 64);
+        var multi = Compile(ReadFirstLaneProgram(), Lowering.EmulatedWave64, threads: 256);
+        Assert.DoesNotContain("wave64RendezvousEpoch", single.Names);
+        Assert.Contains("wave64RendezvousEpoch", multi.Names);
+        Assert.True(multi.Count(SpirvOp.AtomicIAdd) > single.Count(SpirvOp.AtomicIAdd));
+    }
+
+    [Theory]
+    [InlineData(Lowering.Wave32)]
+    [InlineData(Lowering.NativeWave64)]
+    public void NativeLoweringsNeverDeclareARendezvous(Lowering lowering)
+    {
+        var module = Compile(ReadFirstLaneProgram(), lowering, threads: 256);
+        Assert.DoesNotContain("wave64RendezvousEpoch", module.Names);
+        Assert.Equal(0, module.Count(SpirvOp.AtomicIAdd));
+    }
+
+    [Fact]
+    public void HostWave64SupportDoesNotChangeAWave32Shader()
+    {
+        var program = ReadFirstLaneProgram();
+        var (plan, resources, layout) = Prepare(program);
+        byte[] Compile32(bool hostWave64)
+        {
+            var configured = new ShaderCompileRequest(plan, resources, layout)
+            {
+                WaveSize = 32,
+                HostWave64Supported = hostWave64,
+                LocalSizeX = 64,
+            };
+            Assert.True(Gen5SpirvTranslator.TryCompileProgram(configured, out var shader, out var error), error);
+            return shader.Spirv.ToArray();
+        }
+
+        Assert.Equal(Compile32(false), Compile32(true));
+    }
+
+    [Theory]
+    [InlineData(Lowering.Wave32)]
+    [InlineData(Lowering.NativeWave64)]
+    [InlineData(Lowering.EmulatedWave64)]
+    public void ReadLaneBroadcastsFromTheRequestedLane(Lowering lowering)
+    {
+        var module = Compile(ReadLaneProgram(33), lowering);
+        Assert.True(
+            module.Count(SpirvOp.GroupNonUniformBroadcast) > 0 || module.Count(SpirvOp.GroupNonUniformShuffle) > 0 ||
+            module.Count(SpirvOp.Store) > 0);
+    }
+
+    [Theory]
+    [InlineData(Lowering.Wave32)]
+    [InlineData(Lowering.NativeWave64)]
+    [InlineData(Lowering.EmulatedWave64)]
+    public void LoweringsAreDeterministic(Lowering lowering)
+    {
+        var first = Compile(CompareProgram(), lowering);
+        var second = Compile(CompareProgram(), lowering);
+        Assert.Equal(first.Instructions.Count, second.Instructions.Count);
+    }
+}

tests/SharpEmu.ShaderCompiler.Tests/Resources/ResourceMaterializationCacheTests.cs

@@ -153,6 +153,26 @@ public sealed class ResourceMaterializationCacheTests
         Assert.Equal(0, cache.Hits);
     }
 
+    private static ShaderResourcePlan PlanWithSpareUserData() =>
+        ShaderResourcePlan.Extract(DirectImageTableTests.CreateWaveIndexedDescriptorProgram(), ShaderStage.Compute, Hash, 0, 4);
+
+    [Fact]
+    public void AUserDataWordTheProgramNeverReadsDoesNotSplitTheEntry()
+    {
+        var plan = PlanWithSpareUserData();
+        var heap = new Heap();
+        var cache = new ResourceMaterializationCache();
+        Assert.True(Run(cache, plan, heap, [0x1000, 0, 5, 6], out var first, out _));
+        Assert.True(Run(cache, plan, heap, [0x1000, 0, 7, 8], out var second, out _));
+        Assert.Equal((1, 1), (cache.Hits, cache.Misses));
+        // The hit carries the second draw's user data and leaves the first draw's snapshot alone.
+        Assert.Equal(new uint[] { 0x1000, 0, 7, 8 }, second.UserData);
+        Assert.Equal(new uint[] { 0x1000, 0, 5, 6 }, first.UserData);
+        Assert.Same(first.Images, second.Images);
+        Assert.True(Run(cache, plan, heap, [0x1000, 0, 5, 6], out var third, out _));
+        Assert.Same(first, third);
+    }
+
     [Fact]
     public void ADifferentDrawKeyIsADifferentEntry()
     {
@@ -264,6 +284,276 @@ public sealed class ResourceMaterializationCacheTests
         Assert.Equal(0, cache.TableRefreshes);
     }
 
+    // A V# taken verbatim from s[4:7], the way titles pass per-draw constants.
+    private static ShaderResourcePlan DirectBufferPlan() =>
+        Extract(Program(BufferLoad(0, 4), EndProgram(8)), userDataCount: 8);
+
+    // The same V#, whose base pair s[4:5] is also the address of a scalar load.
+    private static ShaderResourcePlan SteeringBufferPlan() =>
+        Extract(Program(ScalarLoad(0, 4, 16), Vop1(8, "VMovB32", 0, Gen5Operand.Scalar(16)), BufferStore(12, 4), EndProgram(20)),
+            userDataCount: 8);
+
+    private static bool RunWords(ResourceMaterializationCache cache, ShaderResourcePlan plan, TestWordMemory memory, uint[] userData,
+        out ResourceSnapshot snapshot)
+    {
+        snapshot = new ResourceSnapshot();
+        var specialization = new ResourceSpecialization();
+        return cache.Materialize(plan, Inputs(userData, memory.Read, memory.Read), (address, destination, _) =>
+        {
+            for (var offset = 0; offset < destination.Length; offset += 4)
+            {
+                if (!memory.Read(address + (ulong)offset, out var word))
+                    return false;
+                BitConverter.TryWriteBytes(destination[offset..], word);
+            }
+
+            return true;
+        }, ref snapshot, ref specialization, out _);
+    }
+
+    private static uint[] Reference(ShaderResourcePlan plan, TestWordMemory memory, uint[] userData)
+    {
+        var snapshot = new ResourceSnapshot();
+        var specialization = new ResourceSpecialization();
+        Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref snapshot, ref specialization));
+        return snapshot.Buffers[0];
+    }
+
+    private static uint[] LinearConstants(int draw) => [0, 0, 0, 0, 0x10000u + (uint)draw * 0x100u, 0, 256, 0x2_0000];
+
+    [Fact]
+    public void ALinearlyAllocatedBufferBaseStopsMissingOnceConfirmed()
+    {
+        var plan = DirectBufferPlan();
+        var memory = new TestWordMemory();
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 16; draw++)
+        {
+            var userData = LinearConstants(draw);
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            Assert.Equal(Reference(plan, memory, userData), snapshot.Buffers[0]);
+            Assert.Equal(userData, snapshot.UserData);
+        }
+
+        // One miss to build, three to confirm the word; every later draw is a hit.
+        Assert.Equal((12, 4), (cache.Hits, cache.Misses));
+    }
+
+    [Fact]
+    public void InterleavedObjectsConfirmTheirOwnBases()
+    {
+        // Constants V# in s[4:7], a second buffer in s[8:11] whose size tells the two objects apart.
+        var plan = Extract(Program(BufferLoad(0, 4), BufferLoad(8, 8), EndProgram(16)), userDataCount: 12);
+        var memory = new TestWordMemory();
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 32; draw++)
+        {
+            uint[] userData = [0, 0, 0, 0, 0x10000u + (uint)draw * 0x100u, 0, 256, 0x2_0000, 0x40000, 0, draw % 2 == 0 ? 256u : 512u, 0x2_0000];
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            var expected = new ResourceSnapshot();
+            var expectedSpecialization = new ResourceSpecialization();
+            Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+            Assert.Equal(expected.Buffers.Select(buffer => buffer.ToArray()), snapshot.Buffers.Select(buffer => buffer.ToArray()));
+        }
+
+        Assert.True(cache.Hits >= 20, $"hits={cache.Hits} misses={cache.Misses}");
+    }
+
+    // Pointer s[0:1] moves on every draw; the program loads four constants through it.
+    private static uint[] PointerConstants(int draw) => [0x10000u + (uint)draw * 0x100u, 0, 0, 0, 0x30000, 0, 256, 0x2_0000, 0];
+
+    [Fact]
+    public void AMovingScalarLoadPointerStopsMissingOnceConfirmed()
+    {
+        var (plan, _, _) = Prepare(FlattenedReadReuseTests.RepeatedReadProgram(4), userDataCount: 9);
+        // The same four words repeat every 0x100 bytes, so every pointer finds the same constants.
+        var memory = new TestWordMemory { Base = 0x10000, Words = Enumerable.Range(0, 0x1000).Select(index => (uint)(index % 64) * 3 + 1).ToArray() };
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 16; draw++)
+        {
+            var userData = PointerConstants(draw);
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            var expected = new ResourceSnapshot();
+            var expectedSpecialization = new ResourceSpecialization();
+            Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+            Assert.Equal(expected.FlattenedResourceTable, snapshot.FlattenedResourceTable);
+            Assert.Equal(expected.Buffers.Select(buffer => buffer.ToArray()), snapshot.Buffers.Select(buffer => buffer.ToArray()));
+        }
+
+        Assert.Equal((12, 4), (cache.Hits, cache.Misses));
+
+        // Other constants behind the next pointer: the draw reads them instead of reusing the table.
+        var moved = PointerConstants(30);
+        memory.At(moved[0] + 8) = 0xBEEF;
+        Assert.True(RunWords(cache, plan, memory, moved, out var reread));
+        Assert.Equal(0xBEEFu, reread.FlattenedResourceTable[2]);
+        Assert.Equal(12, cache.Hits);
+    }
+
+    [Fact]
+    public void ARebasedHitLeavesTheCachedSnapshotAlone()
+    {
+        var plan = DirectBufferPlan();
+        var memory = new TestWordMemory();
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 4; draw++)
+            Assert.True(RunWords(cache, plan, memory, LinearConstants(draw), out _));
+
+        Assert.True(RunWords(cache, plan, memory, LinearConstants(3), out var built));
+        Assert.True(RunWords(cache, plan, memory, LinearConstants(9), out var patched));
+        Assert.Equal(0x10900u, patched.Buffers[0][0]);
+        Assert.Equal(0x10300u, built.Buffers[0][0]);
+        Assert.NotSame(built.Buffers, patched.Buffers);
+        Assert.True(RunWords(cache, plan, memory, LinearConstants(3), out var again));
+        Assert.Equal(0x10300u, again.Buffers[0][0]);
+    }
+
+    [Fact]
+    public void AReadThatMovesWithTheBaseAndKeepsItsBytesIsReused()
+    {
+        var plan = SteeringBufferPlan();
+        var memory = new TestWordMemory { Base = 0x10000, Words = Enumerable.Repeat(0x1234u, 0x1000).ToArray() };
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 16; draw++)
+        {
+            var userData = LinearConstants(draw);
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            var expected = new ResourceSnapshot();
+            var expectedSpecialization = new ResourceSpecialization();
+            Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+            Assert.Equal(expected.Buffers[0], snapshot.Buffers[0]);
+            Assert.Equal(expected.FlattenedResourceTable, snapshot.FlattenedResourceTable);
+        }
+
+        Assert.Equal((12, 4), (cache.Hits, cache.Misses));
+
+        // The moved read now finds other bytes at the new base: the draw must not reuse the table.
+        var changed = LinearConstants(20);
+        memory.At(changed[4]) = 0x5678;
+        Assert.True(RunWords(cache, plan, memory, changed, out var refreshed));
+        Assert.Equal(0x5678u, refreshed.FlattenedResourceTable[0]);
+        Assert.Equal(12, cache.Hits);
+    }
+
+    [Fact]
+    public void MovedConstantsThatChangeEveryDrawAreReadAgain()
+    {
+        var plan = SteeringBufferPlan();
+        var memory = new TestWordMemory { Base = 0x10000, Words = Enumerable.Range(0, 0x1000).Select(index => (uint)index * 7).ToArray() };
+        var cache = new ResourceMaterializationCache();
+        for (var draw = 0; draw < 16; draw++)
+        {
+            var userData = LinearConstants(draw);
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            Assert.Equal(Reference(plan, memory, userData), snapshot.Buffers[0]);
+            var expected = new ResourceSnapshot();
+            var expectedSpecialization = new ResourceSpecialization();
+            Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+            Assert.Equal(expected.FlattenedResourceTable, snapshot.FlattenedResourceTable);
+        }
+
+        // The constants differ at every base, so nothing is reused as is; once the base is
+        // confirmed, the moved table is evaluated again instead of the whole walk.
+        Assert.Equal(0, cache.Hits);
+        Assert.True(cache.TableRefreshes >= 10, $"refreshes={cache.TableRefreshes} misses={cache.Misses}");
+    }
+
+    // The V# comes from a descriptor table at s[0:1]; its constants are read through it.
+    private static ShaderResourcePlan MemoryDescriptorPlan() =>
+        Extract(Program(ScalarLoad(0, 0, 8, 4), ScalarBufferLoad(4, 8, 12, 4), Vop1(8, "VMovB32", 0, Gen5Operand.Scalar(12)),
+            BufferLoad(12, 8), EndProgram(16)), userDataCount: 2);
+
+    // The title rewrites the table every draw: a freshly allocated base, and new constants there.
+    private static void WriteDraw(TestWordMemory memory, int draw)
+    {
+        var bufferBase = 0x10000u + (uint)draw * 0x100u;
+        memory.At(0x1000) = bufferBase;
+        memory.At(0x1004) = 0;
+        memory.At(0x1008) = 256;
+        memory.At(0x100C) = 0x2_0000;
+        for (var word = 0u; word < 4; word++)
+            memory.At(bufferBase + word * 4) = (uint)draw * 10 + word;
+    }
+
+    [Fact]
+    public void ABaseRewrittenInGuestMemoryStopsMissingOnceConfirmed()
+    {
+        var plan = MemoryDescriptorPlan();
+        var memory = new TestWordMemory { Base = 0x1000, Words = new uint[0x8000] };
+        var cache = new ResourceMaterializationCache();
+        uint[] userData = [0x1000, 0];
+        for (var draw = 0; draw < 16; draw++)
+        {
+            WriteDraw(memory, draw);
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            var expected = new ResourceSnapshot();
+            var expectedSpecialization = new ResourceSpecialization();
+            Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+            Assert.Equal(expected.Buffers.Select(buffer => buffer.ToArray()), snapshot.Buffers.Select(buffer => buffer.ToArray()));
+            Assert.Equal(expected.FlattenedResourceTable, snapshot.FlattenedResourceTable);
+        }
+
+        // One miss to build, three to confirm the base; the later draws patch it.
+        Assert.True(cache.Hits >= 10, $"hits={cache.Hits} misses={cache.Misses}");
+
+        // Another descriptor word changed with the base: the draw is materialized again.
+        WriteDraw(memory, 20);
+        memory.At(0x1008) = 512;
+        var hits = cache.Hits;
+        Assert.True(RunWords(cache, plan, memory, userData, out var resized));
+        Assert.Equal(hits, cache.Hits);
+        Assert.Equal(512u, resized.Buffers[0][2]);
+    }
+
+    private static void AssertMatchesFullWalk(ShaderResourcePlan plan, TestWordMemory memory, uint[] userData, ResourceSnapshot snapshot)
+    {
+        var expected = new ResourceSnapshot();
+        var expectedSpecialization = new ResourceSpecialization();
+        Assert.True(ResourceMaterializer.Materialize(plan, Inputs(userData, memory.Read, memory.Read), ref expected, ref expectedSpecialization));
+        Assert.Equal(expected.Buffers.Select(buffer => buffer.ToArray()), snapshot.Buffers.Select(buffer => buffer.ToArray()));
+        Assert.Equal(expected.FlattenedResourceTable, snapshot.FlattenedResourceTable);
+    }
+
+    [Fact]
+    public void ATableConstantBesideARewrittenBaseIsCopiedForTheDraw()
+    {
+        // The V# and a constant both come from the table at s[0:1]; the constant changes every draw.
+        var plan = Extract(Program(ScalarLoad(0, 0, 8, 4), ScalarLoad(4, 0, 16, 1, immediateOffset: 16),
+            ScalarBufferLoad(8, 8, 12, 4), Vop1(12, "VMovB32", 0, Gen5Operand.Scalar(12)),
+            Vop1(16, "VMovB32", 1, Gen5Operand.Scalar(16)), BufferLoad(20, 8), EndProgram(24)), userDataCount: 2);
+        var memory = new TestWordMemory { Base = 0x1000, Words = new uint[0x8000] };
+        var cache = new ResourceMaterializationCache();
+        uint[] userData = [0x1000, 0];
+        for (var draw = 0; draw < 16; draw++)
+        {
+            WriteDraw(memory, draw);
+            memory.At(0x1010) = (uint)draw * 3 + 1;
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            AssertMatchesFullWalk(plan, memory, userData, snapshot);
+        }
+
+        Assert.True(cache.Hits >= 10, $"hits={cache.Hits} misses={cache.Misses}");
+    }
+
+    [Fact]
+    public void AlternatingObjectsWithRewrittenBasesRebaseTheirOwnVariant()
+    {
+        // Two objects share the draw key and differ in their buffer size; both move their base.
+        var plan = MemoryDescriptorPlan();
+        var memory = new TestWordMemory { Base = 0x1000, Words = new uint[0x8000] };
+        var cache = new ResourceMaterializationCache();
+        uint[] userData = [0x1000, 0];
+        for (var draw = 0; draw < 32; draw++)
+        {
+            WriteDraw(memory, draw);
+            memory.At(0x1008) = draw % 2 == 0 ? 256u : 512u;
+            Assert.True(RunWords(cache, plan, memory, userData, out var snapshot));
+            AssertMatchesFullWalk(plan, memory, userData, snapshot);
+        }
+
+        Assert.True(cache.Hits >= 20, $"hits={cache.Hits} misses={cache.Misses}");
+    }
+
     [Fact]
     public void AFullGenerationKeepsRecentEntries()
     {