[CPU] Fix Sema ORBIS_GEN2_ERROR_BUSY loop (#621)

* fix(memory): reserve only large regions

* Potential fix for pull request finding

README.md

@@ -25,6 +25,14 @@ SPDX-License-Identifier: GPL-2.0-or-later
 
 ---
 
+<p align="center">
+  <a href="#support">
+    <img src="https://img.shields.io/badge/Support-GitHub%20Sponsors%20%26%20Crypto-EA4AAA?style=for-the-badge&logo=githubsponsors&logoColor=white" alt="Support SharpEmu">
+  </a>
+</p>
+
+---
+
 > [!NOTE]  
 > SharpEmu supports Windows x64, Linux x64, and macOS x64. Apple Silicon Macs
 > can run the macOS x64 build through Rosetta 2, and Windows on ARM devices
@@ -136,6 +144,18 @@ Provided valuable references for filesystem handling and low-level C# implementa
 
 - [**GPL-2.0 license**](https://github.com/sharpemu/sharpemu/blob/main/LICENSE)
 
+## Support
+
+Support SharpEmu via GitHub Sponsors or cryptocurrency. Every contribution helps fund ongoing development and long-term maintenance. GitHub Sponsors is the preferred way to support the project, but cryptocurrency donations are also appreciated.
+
+### ETH/USDT
+
+`0xF315F5d986c790bB3A58DbE60F1B2760997dEd82`
+
+### BTC
+
+`bc1qmr9k8899njys5ny63xsues4jgmkk96erslrkmv`
+
 ## Contributing
 
 Before opening an issue or pull request, please read our contribution guidelines:

SharpEmu.slnx

@@ -21,6 +21,7 @@ SPDX-License-Identifier: GPL-2.0-or-later
   <Folder Name="/tests/">
     <Project Path="tests/SharpEmu.Libs.Tests/SharpEmu.Libs.Tests.csproj" />
     <Project Path="tests/SharpEmu.ShaderCompiler.Metal.Tests/SharpEmu.ShaderCompiler.Metal.Tests.csproj" />
+    <Project Path="tests/SharpEmu.ShaderCompiler.Tests/SharpEmu.ShaderCompiler.Tests.csproj" />
     <Project Path="tests/SharpEmu.SourceGenerators.Tests/SharpEmu.SourceGenerators.Tests.csproj" />
   </Folder>
 </Solution>

scripts/ps5_names.txt

@@ -153133,6 +153133,7 @@ scePsmlMfsrGetContextBufferRequirement800M3_2
 scePsmlMfsrGetDispatchMfsrPacket1000
 scePsmlMfsrGetDispatchMfsrPacket1100
 scePsmlMfsrGetDispatchMfsrPacketSizeInDwords
+scePsmlMfsrGetDispatchMfsrPacket900
 scePsmlMfsrGetMipmapBias
 scePsmlMfsrGetSharedResourcesInitRequirement
 scePsmlMfsrInit

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

@@ -19,6 +19,9 @@ public sealed partial class DirectExecutionBackend
 	private static int _lazyCommitTraceCount;
 	private static int _guestAllocatorHoleRecoveries;
 	private static int _auxiliaryThreadExecuteFaultRecoveries;
+	private static int _auxiliaryThreadExecuteFaultSkips;
+	private nint _workerAbortStack;
+	private const uint WorkerAbortStackSize = 0x10000u;
 
 	private unsafe void SetupExceptionHandler()
 	{
@@ -435,18 +438,91 @@ public sealed partial class DirectExecutionBackend
 		void* contextRecord,
 		ulong rip)
 	{
-		if (exceptionRecord->ExceptionCode != 3221225477u ||
-			rip >= 0x0000000800000000UL ||
-			_activeGuestThreadState is not { Name: "tbb_thead" } activeThread)
+		if (exceptionRecord->ExceptionCode != 3221225477u)
 		{
 			return false;
 		}
 
+		// Prefer ThreadStatic active state; fall back to host-thread name when
+		// concurrent TBB AVs race logging (tLT61: recover skipped, then Fatal).
+		GuestThreadState? activeThread = _activeGuestThreadState;
+		if (activeThread is null || activeThread.Name != "tbb_thead")
+		{
+			var hostName = Thread.CurrentThread.Name;
+			if (hostName is null ||
+				!hostName.StartsWith("SharpEmu-tbb_thead", StringComparison.Ordinal))
+			{
+				return false;
+			}
+
+			activeThread = FindGuestThreadStateByHostThreadId(unchecked((int)GetCurrentThreadId()));
+			if (activeThread is null || activeThread.Name != "tbb_thead")
+			{
+				var skip = Interlocked.Increment(ref _auxiliaryThreadExecuteFaultSkips);
+				if (skip <= 8 || skip % 64 == 0)
+				{
+					Console.Error.WriteLine(
+						$"[LOADER][WARN] tbb_recover skip #{skip}: rip=0x{rip:X16} " +
+						$"host='{hostName}' active={(activeThread?.Name ?? "null")}");
+					Console.Error.Flush();
+				}
+				return false;
+			}
+		}
+
 		var hostExit = ActiveEntryReturnSentinelRip;
 		if (hostExit < 0x10000)
 		{
 			hostExit = unchecked((ulong)_guestReturnStub);
 		}
+
+			// Prefer worker-abort (SetEvent + ExitThread) over host_exit→RunEpilogue:
+		// the latter FailFasts the process after TBB recover (tLT28/30 silent die).
+		// Do NOT abandon mutexes here — managed HLE from inside VEH can re-enter
+		// and Fatal (tLT73). NativeGuestExecutor.Run abandons after detecting abort.
+		var abortRip = unchecked((ulong)_workerAbortStub);
+		if (abortRip >= 0x10000)
+		{
+			// Do NOT SetEvent from managed VEH: that wakes the renter which may
+			// TerminateThread while this thread is still inside VEH return
+			// (tLTA2: recover logged, no respawning, process die). Abort stub
+			// SetEvent's only after CONTINUE_EXECUTION resumes at park.
+
+			// Prefer the entry-stub-saved host RSP (real CreateThread stack).
+			// Do not treat mid-range host stacks as guest — Astro worker stacks
+			// often sit in 0x02xxxxxx_xxxx and were wrongly replaced with a
+			// shared VirtualAlloc abort stack (concurrent TBB AV → die).
+			var hostRspSlot = TlsGetValue(_hostRspSlotTlsIndex);
+			ulong hostRsp = 0;
+			if (hostRspSlot != 0)
+			{
+				hostRsp = *(ulong*)hostRspSlot;
+			}
+
+			if (hostRsp < 0x10000)
+			{
+				hostRsp = EnsureWorkerAbortStackRsp();
+			}
+
+			if (hostRsp >= 0x10000)
+			{
+				WriteCtxU64(contextRecord, 152, hostRsp & ~0xFUL);
+			}
+
+			WriteCtxU64(contextRecord, 120, 0);
+			WriteCtxU64(contextRecord, 248, abortRip);
+			var recovery = Interlocked.Increment(ref _auxiliaryThreadExecuteFaultRecoveries);
+			Console.Error.WriteLine(
+				$"[LOADER][WARN] Recovered auxiliary TBB execute fault #{recovery}: " +
+				$"thread=0x{activeThread.ThreadHandle:X16} target=0x{rip:X16} " +
+				$"host_rsp=0x{hostRsp:X16} -> worker_abort=0x{abortRip:X16}");
+			Console.Error.WriteLine(
+				"[LOADER][INFO] tbb_recover: parking native worker (SetEvent+park); " +
+				"renter will TerminateThread+respawn — avoids ExitThread after VEH");
+			Console.Error.Flush();
+			return true;
+		}
+
 		if (hostExit < 0x10000)
 		{
 			Console.Error.WriteLine(
@@ -458,13 +534,57 @@ public sealed partial class DirectExecutionBackend
 		_ = TryPatchActiveGuestReturnSlot(hostExit);
 		WriteCtxU64(contextRecord, 120, 0);
 		WriteCtxU64(contextRecord, 248, hostExit);
-		var recovery = Interlocked.Increment(ref _auxiliaryThreadExecuteFaultRecoveries);
+		var recoveryFallback = Interlocked.Increment(ref _auxiliaryThreadExecuteFaultRecoveries);
 		Console.Error.WriteLine(
-			$"[LOADER][WARN] Recovered auxiliary TBB execute fault #{recovery}: " +
+			$"[LOADER][WARN] Recovered auxiliary TBB execute fault #{recoveryFallback}: " +
 			$"thread=0x{activeThread.ThreadHandle:X16} target=0x{rip:X16} -> host_exit=0x{hostExit:X16}");
+		Console.Error.WriteLine(
+			"[LOADER][INFO] tbb_recover: resumed at host_exit (abort stub unavailable); " +
+			"subsequent FastFail/CLR must not re-enter managed VEH " +
+			"(live trampoline pre-filters 0xC0000409 / 0xE0434352)");
+		Console.Error.Flush();
 		return true;
 	}
 
+	private GuestThreadState? FindGuestThreadStateByHostThreadId(int hostThreadId)
+	{
+		if (hostThreadId == 0)
+		{
+			return null;
+		}
+
+		try
+		{
+			foreach (var thread in SnapshotGuestThreads())
+			{
+				if (Volatile.Read(ref thread.HostThreadId) == hostThreadId)
+				{
+					return thread;
+				}
+			}
+		}
+		catch
+		{
+		}
+
+		return null;
+	}
+
+	private unsafe ulong EnsureWorkerAbortStackRsp()
+	{
+		if (_workerAbortStack == 0)
+		{
+			_workerAbortStack = (nint)VirtualAlloc(null, WorkerAbortStackSize, 12288u, 4u);
+			if (_workerAbortStack == 0)
+			{
+				return 0;
+			}
+		}
+
+		// Grow-down stack: hand out near the top with alignment headroom.
+		return (ulong)(_workerAbortStack + (nint)WorkerAbortStackSize - 0x100) & ~0xFUL;
+	}
+
 	private unsafe bool TryRecoverGuestInt41(uint exceptionCode, void* contextRecord, ulong rip)
 	{
 		if (!_ignoreGuestInt41 || exceptionCode != 3221225477u || rip < 0x10000)

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

@@ -1449,6 +1449,9 @@ public sealed partial class DirectExecutionBackend
 		var expectedSemaphoreTrywaitAgain =
 			string.Equals(nid, "H2a+IN9TP0E", StringComparison.Ordinal) &&
 			result == OrbisGen2Result.ORBIS_GEN2_ERROR_TRY_AGAIN;
+		var expectedPollSemaBusy =
+			string.Equals(nid, "12wOHk8ywb0", StringComparison.Ordinal) &&
+			result == OrbisGen2Result.ORBIS_GEN2_ERROR_BUSY;
 		var expectedNetAcceptWouldBlock =
 			string.Equals(nid, "PIWqhn9oSxc", StringComparison.Ordinal) &&
 			resultValue == unchecked((int)0x80410123);
@@ -1463,6 +1466,7 @@ public sealed partial class DirectExecutionBackend
 			!expectedEqueueTimeout &&
 			!expectedMutexTrylockBusy &&
 			!expectedSemaphoreTrywaitAgain &&
+			!expectedPollSemaBusy &&
 			!expectedNetAcceptWouldBlock &&
 			!expectedUserServiceNoEvent &&
 			!expectedPrivacyInvalidParameter)

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

@@ -29,9 +29,29 @@ public sealed partial class DirectExecutionBackend
 	private static readonly bool NativeGuestWorkersDisabled =
 		string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_DISABLE_NATIVE_GUEST_WORKERS"), "1", StringComparison.Ordinal);
 
+	// Cap concurrent native-worker Runs. Astro's tbb_thead burst overlaps many
+	// UnmanagedCallersOnly prologues; a large prewarm + unbounded concurrency
+	// FailFasts (0xC0000409) mid-storm with no VEH breadcrumb. Pool size and
+	// in-flight Runs are separate knobs.
+	private static readonly int NativeWorkerMaxConcurrent = ReadNativeWorkerMaxConcurrent();
+
+	private static int ReadNativeWorkerMaxConcurrent()
+	{
+		if (int.TryParse(
+			    Environment.GetEnvironmentVariable("SHARPEMU_NATIVE_WORKER_MAX_CONCURRENT"),
+			    out var parsed) &&
+		    parsed > 0)
+		{
+			return Math.Clamp(parsed, 1, 64);
+		}
+
+		return 2;
+	}
+
 	private readonly object _nativeWorkerGate = new();
 	private readonly List<NativeGuestExecutor> _allNativeWorkers = new();
 	private readonly Stack<NativeGuestExecutor> _idleNativeWorkers = new();
+	private readonly SemaphoreSlim _nativeWorkerRunLimiter = new(NativeWorkerMaxConcurrent);
 	private bool _nativeWorkersDisposed;
 	private int _nativeWorkerCreationFailedLogged;
 
@@ -49,6 +69,9 @@ public sealed partial class DirectExecutionBackend
 	[DllImport("kernel32.dll", SetLastError = true)]
 	private static extern uint WaitForSingleObject(nint hHandle, uint dwMilliseconds);
 
+	[DllImport("kernel32.dll", SetLastError = true)]
+	private static extern bool TerminateThread(nint hThread, uint dwExitCode);
+
 	// Runs an emitted guest entry stub. Preferred path is a pooled native worker
 	// thread; falls back to the historical inline calli (guest frames above this
 	// thread's managed frames) when workers are disabled or unavailable.
@@ -56,40 +79,148 @@ public sealed partial class DirectExecutionBackend
 	// Callers set the Active* thread-statics before emitting the stub and read the
 	// yield/forced-exit flags right after this returns, so the worker outcome is
 	// copied back into this thread's statics before returning.
-	private unsafe int RunGuestEntryStub(void* entryStub, ulong hostRspSlot)
+	private unsafe int RunGuestEntryStub(void* entryStub, ulong hostRspSlot, bool requireNativeWorker = false)
 	{
-		var worker = RentNativeGuestExecutor();
-		if (worker is null)
-		{
-			TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot);
-			return CallNativeEntry(entryStub);
-		}
+		// Limit in-flight native Runs before renting so the idle pool is not
+		// drained by threads blocked on the concurrency gate.
+		_nativeWorkerRunLimiter.Wait();
+		NativeGuestExecutor? worker = null;
 		try
 		{
-			var state = _activeGuestThreadState;
-			var nativeReturn = worker.Run(
-				_activeCpuContext!,
-				state,
-				GuestThreadExecution.CurrentGuestThreadHandle,
-				_activeEntryReturnSentinelRip,
-				_activeGuestReturnSlotAddress,
-				(nint)hostRspSlot,
-				(nint)entryStub,
-				state?.AffinityMask ?? 0,
-				out var yieldRequested,
-				out var yieldReason,
-				out var forcedExit);
-			_activeGuestThreadYieldRequested = yieldRequested;
-			_activeGuestThreadYieldReason = yieldReason;
-			_activeForcedGuestExit = forcedExit;
-			return nativeReturn;
+			// Astro can spawn a burst of tbb_thead while workers are still in
+			// TerminateThread+respawn. Wait for a native worker — never fall back
+			// to managed inline (FailFast) and never throw (uncaught throw mid-
+			// storm was a silent process die).
+			var maxAttempts = requireNativeWorker ? 500 : 48;
+			for (var attempt = 0; attempt < maxAttempts; attempt++)
+			{
+				worker = RentNativeGuestExecutor();
+				if (worker is not null)
+				{
+					break;
+				}
+
+				if (!requireNativeWorker)
+				{
+					break;
+				}
+
+				Thread.Sleep(attempt < 32 ? 1 : 4);
+			}
+
+			if (worker is null)
+			{
+				if (requireNativeWorker)
+				{
+					var n = Interlocked.Increment(ref _tbbNativeWorkerRefuseCount);
+					if (n <= 8 || n % 32 == 0)
+					{
+						Console.Error.WriteLine(
+							$"[LOADER][ERROR] tbb_native_worker unavailable #{n} after {maxAttempts} attempts; " +
+							"skipping run (no managed inline, no throw)");
+						Console.Error.Flush();
+					}
+
+					_activeGuestThreadYieldRequested = true;
+					_activeGuestThreadYieldReason = "tbb_native_worker_unavailable";
+					_activeForcedGuestExit = true;
+					return unchecked((int)0x80020012);
+				}
+
+				TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot);
+				return CallNativeEntry(entryStub);
+			}
+
+			try
+			{
+				var state = _activeGuestThreadState;
+				if (state is { Name: "tbb_thead" })
+				{
+					var n = Interlocked.Increment(ref _tbbNativeRunEnterCount);
+					if (n <= 12 || n % 64 == 0)
+					{
+						Console.Error.WriteLine(
+							$"[LOADER][INFO] tbb_run_enter #{n} native_tid_pending handle=0x{state.ThreadHandle:X16} " +
+							$"max_concurrent={NativeWorkerMaxConcurrent}");
+						Console.Error.Flush();
+					}
+				}
+
+				var nativeReturn = worker.Run(
+					_activeCpuContext!,
+					state,
+					GuestThreadExecution.CurrentGuestThreadHandle,
+					_activeEntryReturnSentinelRip,
+					_activeGuestReturnSlotAddress,
+					(nint)hostRspSlot,
+					(nint)entryStub,
+					state?.AffinityMask ?? 0,
+					out var yieldRequested,
+					out var yieldReason,
+					out var forcedExit);
+				_activeGuestThreadYieldRequested = yieldRequested;
+				_activeGuestThreadYieldReason = yieldReason;
+				_activeForcedGuestExit = forcedExit;
+				return nativeReturn;
+			}
+			finally
+			{
+				ReturnNativeGuestExecutor(worker);
+			}
 		}
 		finally
 		{
-			ReturnNativeGuestExecutor(worker);
+			_nativeWorkerRunLimiter.Release();
 		}
 	}
 
+	private static int _tbbNativeRunEnterCount;
+	private static int _tbbNativeWorkerRefuseCount;
+	internal static int _tbbWorkerPrologueFaultCount;
+
+	private void PrewarmNativeGuestWorkers(int count)
+	{
+		if (!OperatingSystem.IsWindows() || NativeGuestWorkersDisabled || count <= 0)
+		{
+			return;
+		}
+
+		var warmed = new List<NativeGuestExecutor>(count);
+		for (var i = 0; i < count; i++)
+		{
+			var worker = NativeGuestExecutor.TryCreate(this);
+			if (worker is null)
+			{
+				break;
+			}
+
+			warmed.Add(worker);
+		}
+
+		lock (_nativeWorkerGate)
+		{
+			if (_nativeWorkersDisposed)
+			{
+				foreach (var worker in warmed)
+				{
+					worker.Dispose();
+				}
+				return;
+			}
+
+			foreach (var worker in warmed)
+			{
+				_allNativeWorkers.Add(worker);
+				_idleNativeWorkers.Push(worker);
+			}
+		}
+
+		Console.Error.WriteLine(
+			$"[LOADER][INFO] Native guest workers prewarmed: {warmed.Count}/{count} " +
+			$"max_concurrent={NativeWorkerMaxConcurrent}");
+		Console.Error.Flush();
+	}
+
 	private NativeGuestExecutor? RentNativeGuestExecutor()
 	{
 		// NativeGuestExecutor emits a Win32 wait loop and creates it with
@@ -400,6 +531,22 @@ public sealed partial class DirectExecutionBackend
 				return false;
 			}
 			FlushInstructionCache(GetCurrentProcess(), _loopStub, LoopStubSize);
+			return StartWorkerThread();
+		}
+
+		private bool RestartWorkerThread()
+		{
+			if (_loopStub == null || _controlBlock == null)
+			{
+				return false;
+			}
+
+			*(int*)_controlBlock = 0;
+			return StartWorkerThread();
+		}
+
+		private bool StartWorkerThread()
+		{
 			_threadHandle = CreateThread(
 				0,
 				WorkerStackReservation,
@@ -445,6 +592,49 @@ public sealed partial class DirectExecutionBackend
 			_runForcedExit = false;
 			SignalWorkAvailable();
 			WaitWorkCompleted();
+
+			// Normal path: RunEpilogue/ExitRun clears _entered before SetEvent(done).
+			// TBB abort stub SetEvent's without ExitRun — _entered stays true.
+			if (_entered)
+			{
+				var waitRc = WaitForSingleObject(_threadHandle, 500u);
+				Console.Error.WriteLine(
+					$"[LOADER][WARN] Native guest worker tid={_nativeThreadId} aborted during run; " +
+					$"wait_rc=0x{waitRc:X8} respawning");
+				Console.Error.Flush();
+				if (_runState is { } abortedState)
+				{
+					_ = GuestThreadExecution.NotifyGuestThreadAbandoned(
+						abortedState.ThreadHandle,
+						"tbb_worker_abort");
+					Volatile.Write(ref abortedState.HostThreadId, _prevHostThreadId);
+				}
+				_entered = false;
+				if (_threadHandle != 0)
+				{
+					// Abort stub parks (no ExitThread). Force-kill the parked OS
+					// thread so we can recreate the loop without process teardown.
+					if (waitRc != 0u)
+					{
+						_ = TerminateThread(_threadHandle, unchecked((uint)(-1)));
+						_ = WaitForSingleObject(_threadHandle, 1000u);
+					}
+					CloseHandle(_threadHandle);
+					_threadHandle = 0;
+					_nativeThreadId = 0;
+				}
+				if (!RestartWorkerThread())
+				{
+					_runPrologueFailed = true;
+				}
+				else
+				{
+					_runPrologueFailed = false;
+					_runForcedExit = true;
+					_runNativeResult = 0;
+				}
+			}
+
 			_runContext = null;
 			_runState = null;
 			yieldRequested = _runYieldRequested;
@@ -452,7 +642,22 @@ public sealed partial class DirectExecutionBackend
 			forcedExit = _runForcedExit;
 			if (_runPrologueFailed)
 			{
-				throw new InvalidOperationException("Native guest worker failed to bind the run ambient (prologue fault)");
+				// Never throw out of the native-worker rent path: an uncaught
+				// exception mid-TBB storm kills the process with no FailFast
+				// breadcrumb.
+				var n = Interlocked.Increment(ref _tbbWorkerPrologueFaultCount);
+				if (n <= 8 || n % 32 == 0)
+				{
+					Console.Error.WriteLine(
+						$"[LOADER][ERROR] tbb_worker prologue fault #{n}; soft-fail run " +
+						$"(tid={_nativeThreadId})");
+					Console.Error.Flush();
+				}
+
+				yieldRequested = true;
+				yieldReason = "tbb_worker_prologue_fault";
+				forcedExit = true;
+				return unchecked((int)0x80020012);
 			}
 			return _runNativeResult;
 		}
@@ -546,6 +751,18 @@ public sealed partial class DirectExecutionBackend
 			_activeGuestThreadState = _runState;
 			backend.BindTlsBase(_runContext!);
 			TlsSetValue(backend._hostRspSlotTlsIndex, _runHostRspSlot);
+			if (backend._workerDoneEventTlsIndex != uint.MaxValue)
+			{
+				nint doneHandle = OperatingSystem.IsWindows()
+					? _workCompleted!.SafeWaitHandle.DangerousGetHandle()
+					: _doneSemaphore;
+				TlsSetValue(backend._workerDoneEventTlsIndex, doneHandle);
+			}
+			if (backend._tbbAbortEligibleTlsIndex != uint.MaxValue)
+			{
+				nint eligible = _runState is { Name: "tbb_thead" } ? 1 : 0;
+				TlsSetValue(backend._tbbAbortEligibleTlsIndex, eligible);
+			}
 			if (_runState is { } state)
 			{
 				_prevHostThreadId = Volatile.Read(ref state.HostThreadId);
@@ -580,6 +797,14 @@ public sealed partial class DirectExecutionBackend
 				Volatile.Write(ref state.HostThreadId, _prevHostThreadId);
 			}
 			TlsSetValue(_backend._hostRspSlotTlsIndex, _prevHostRspSlot);
+			if (_backend._workerDoneEventTlsIndex != uint.MaxValue)
+			{
+				TlsSetValue(_backend._workerDoneEventTlsIndex, 0);
+			}
+			if (_backend._tbbAbortEligibleTlsIndex != uint.MaxValue)
+			{
+				TlsSetValue(_backend._tbbAbortEligibleTlsIndex, 0);
+			}
 			GuestThreadExecution.RestoreGuestThread(_prevGuestThreadHandle);
 			_activeExecutionBackend = _prevBackend;
 			_activeCpuContext = _prevContext;

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

@@ -214,6 +214,15 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 
 	private nint _guestReturnStub;
 
+	private nint _workerAbortStub;
+	private nint _vehManagedEntryLock;
+
+	private uint _workerDoneEventTlsIndex = uint.MaxValue;
+
+	private uint _tbbAbortEligibleTlsIndex = uint.MaxValue;
+
+	private nint _setEventAddress;
+
 	private nint _rawExceptionHandler;
 
 	private nint _rawExceptionHandlerStub;
@@ -1041,6 +1050,8 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		_selfHandlePtr = GCHandle.ToIntPtr(_selfHandle);
 		_guestTlsBaseTlsIndex = TlsAlloc();
 		_hostRspSlotTlsIndex = TlsAlloc();
+		_workerDoneEventTlsIndex = OperatingSystem.IsWindows() ? TlsAlloc() : uint.MaxValue;
+		_tbbAbortEligibleTlsIndex = OperatingSystem.IsWindows() ? TlsAlloc() : uint.MaxValue;
 		if (_guestTlsBaseTlsIndex == uint.MaxValue || _hostRspSlotTlsIndex == uint.MaxValue)
 		{
 			throw new OutOfMemoryException("Failed to allocate native TLS slots");
@@ -1064,6 +1075,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			{
 				throw new InvalidOperationException("Failed to resolve kernel32 thread timing functions");
 			}
+			_setEventAddress = kernel32 != 0 ? GetProcAddress(kernel32, "SetEvent") : 0;
 		}
 		else
 		{
@@ -1087,13 +1099,29 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		{
 			throw new OutOfMemoryException("Failed to allocate host stack slot storage");
 		}
+		_vehManagedEntryLock = (nint)VirtualAlloc(null, 64u, 12288u, 4u);
+		if (_vehManagedEntryLock == 0)
+		{
+			throw new OutOfMemoryException("Failed to allocate VEH managed-entry lock");
+		}
+		// owner (nint) + depth (int); recursive — nested VEH on same thread must reenter.
+		*(nint*)_vehManagedEntryLock = 0;
+		*(int*)(_vehManagedEntryLock + sizeof(nint)) = 0;
 		_unresolvedReturnStub = CreateUnresolvedReturnStub();
 		_guestReturnStub = CreateGuestReturnStub();
 		if (_guestReturnStub == 0)
 		{
 			throw new OutOfMemoryException("Failed to allocate guest return stub");
 		}
+		_workerAbortStub = CreateWorkerAbortStub();
+		if (_workerAbortStub == 0 && OperatingSystem.IsWindows())
+		{
+			Console.Error.WriteLine(
+				"[LOADER][WARN] Worker abort stub unavailable; TBB execute-fault recover will use host_exit");
+		}
 		SetupExceptionHandler();
+		// Cover the Astro TBB spawn storm (often 8–12 concurrent tbb_thead).
+		PrewarmNativeGuestWorkers(Math.Max(NativeWorkerMaxConcurrent, 4));
 	}
 
 	public bool TryExecute(CpuContext context, ulong entryPoint, Generation generation, IReadOnlyDictionary<ulong, string> importStubs, IReadOnlyDictionary<string, ulong> runtimeSymbols, CpuExecutionOptions executionOptions, out OrbisGen2Result result)
@@ -2409,9 +2437,147 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		return (nint)ptr;
 	}
 
+	/// <summary>
+	/// After a TBB execute-fault, VEH redirects here on a host stack.
+	/// SetEvent(done) then park forever — ExitThread from VEH CONTINUE_EXECUTION
+	/// was taking down the whole process (recover logged, no respawning). The
+	/// renter TerminateThread's the parked worker and respawns a clean loop.
+	/// </summary>
+	private unsafe nint CreateWorkerAbortStub()
+	{
+		if (!OperatingSystem.IsWindows() ||
+			_workerDoneEventTlsIndex == uint.MaxValue ||
+			_tlsGetValueAddress == 0 ||
+			_setEventAddress == 0)
+		{
+			return 0;
+		}
+
+		nint kernel32 = GetModuleHandle("kernel32.dll");
+		nint getStdHandle = kernel32 != 0 ? GetProcAddress(kernel32, "GetStdHandle") : 0;
+		nint writeFile = kernel32 != 0 ? GetProcAddress(kernel32, "WriteFile") : 0;
+		nint flushFileBuffers = kernel32 != 0 ? GetProcAddress(kernel32, "FlushFileBuffers") : 0;
+
+		const uint stubSize = 256u;
+		void* ptr = VirtualAlloc(null, stubSize, 12288u, 4u);
+		if (ptr == null)
+		{
+			return 0;
+		}
+
+		byte* code = (byte*)ptr;
+		int offset = 0;
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+		EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28); // sub rsp, 0x28
+		EmitByte(code, ref offset, 0xB9);
+		EmitUInt32(code, ref offset, _workerDoneEventTlsIndex); // mov ecx, tls
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+		*(nint*)(code + offset) = _tlsGetValueAddress;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0); // call TlsGetValue
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85);
+		EmitByte(code, ref offset, 0xC0); // test rax, rax
+		EmitByte(code, ref offset, 0x74); EmitByte(code, ref offset, 0x0F); // jz skip SetEvent (15 bytes)
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+		EmitByte(code, ref offset, 0xC1); // mov rcx, rax
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+		*(nint*)(code + offset) = _setEventAddress;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0); // call SetEvent
+
+		// Breadcrumb on host stack (survives silent teardown better than managed log).
+		int msgAbsSlot = -1;
+		if (getStdHandle != 0 && writeFile != 0)
+		{
+			ReadOnlySpan<byte> msg = "[LOADER][WARN] tbb_abort_stub SetEvent+park\n"u8;
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x20); // extra shadow for WriteFile args
+			EmitByte(code, ref offset, 0xB9); EmitUInt32(code, ref offset, unchecked((uint)-12));
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = getStdHandle;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0xC1); // mov rcx, handle
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0xC3); // mov rbx, handle (nonvolatile for flush)
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			msgAbsSlot = offset;
+			*(nint*)(code + offset) = 0;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0xC2); // mov rdx, msg
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xB8);
+			EmitUInt32(code, ref offset, (uint)msg.Length);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20); // lea r9, [rsp+0x20]
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20); EmitUInt32(code, ref offset, 0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x28); EmitUInt32(code, ref offset, 0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = writeFile;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			if (flushFileBuffers != 0)
+			{
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+				EmitByte(code, ref offset, 0xD9); // mov rcx, rbx
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+				*(nint*)(code + offset) = flushFileBuffers;
+				offset += sizeof(nint);
+				EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			}
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x20);
+		}
+
+		// Park: do not ExitThread (process-wide silent die after VEH redirect).
+		int parkOffset = offset;
+		EmitByte(code, ref offset, 0xF3); EmitByte(code, ref offset, 0x90); // pause
+		EmitByte(code, ref offset, 0xEB);
+		EmitByte(code, ref offset, unchecked((byte)(parkOffset - (offset + 1)))); // jmp park
+
+		if (msgAbsSlot >= 0)
+		{
+			ReadOnlySpan<byte> msgEmbed = "[LOADER][WARN] tbb_abort_stub SetEvent+park\n"u8;
+			*(nint*)(code + msgAbsSlot) = (nint)ptr + offset;
+			for (int i = 0; i < msgEmbed.Length; i++)
+			{
+				EmitByte(code, ref offset, msgEmbed[i]);
+			}
+		}
+
+		if (offset > (int)stubSize)
+		{
+			Console.Error.WriteLine(
+				$"[LOADER][ERROR] Worker abort stub overflow: used={offset} cap={stubSize}");
+			VirtualFree(ptr, 0u, 32768u);
+			return 0;
+		}
+
+		uint oldProtect = default;
+		if (!VirtualProtect(ptr, stubSize, 32u, &oldProtect))
+		{
+			VirtualFree(ptr, 0u, 32768u);
+			return 0;
+		}
+		FlushInstructionCache(GetCurrentProcess(), ptr, (nuint)offset);
+		return (nint)ptr;
+	}
+
 	private unsafe nint CreateExceptionHandlerTrampoline(nint managedHandler)
 	{
-		const uint stubSize = 256u;
+		// Live VEH trampoline used by SetupExceptionHandler. Must pre-filter
+		// FastFail / CLR / MSVC C++ / stack-overflow the same way as
+		// WindowsFaultHandling.CreateHandlerThunk: entering managed VEH while
+		// the thread is in cooperative GC mode fail-fasts with
+		// "UnmanagedCallersOnly method from managed code" (tLT18–22).
+		// Extra headroom for native tbb abort + recursive managed-entry spinlock.
+		const uint stubSize = 2048u;
 		void* ptr = VirtualAlloc(null, stubSize, 12288u, 64u);
 		if (ptr == null)
 		{
@@ -2420,10 +2586,305 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 
 		byte* code = (byte*)ptr;
 		int offset = 0;
+
+		ReadOnlySpan<uint> nonManagedExceptionCodes =
+		[
+			0xE0434352u, // CLR managed exception
+			0xE06D7363u, // MSVC C++ exception
+			0xC0000409u, // STATUS_STACK_BUFFER_OVERRUN / FailFast
+			0xC00000FDu, // STATUS_STACK_OVERFLOW
+		];
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x01); // mov rax, [rcx]
+		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x00); // mov eax, [rax] ExceptionCode
+		var passJumpOffsets = stackalloc int[nonManagedExceptionCodes.Length];
+		int fastFailJumpSlot = -1;
+		for (int i = 0; i < nonManagedExceptionCodes.Length; i++)
+		{
+			EmitByte(code, ref offset, 0x3D);
+			EmitUInt32(code, ref offset, nonManagedExceptionCodes[i]);
+			EmitByte(code, ref offset, 0x74);
+			passJumpOffsets[i] = offset;
+			EmitByte(code, ref offset, 0x00);
+			if (nonManagedExceptionCodes[i] == 0xC0000409u)
+			{
+				fastFailJumpSlot = i;
+			}
+		}
+
+		EmitByte(code, ref offset, 0xE9); // jmp mainBody (rel32; FastFail breadcrumb sits between)
+		var mainBodyJumpSlot = offset;
+		EmitUInt32(code, ref offset, 0u);
+
+		int passOffset = offset;
+		EmitByte(code, ref offset, 0x31); EmitByte(code, ref offset, 0xC0); // xor eax, eax
+		EmitByte(code, ref offset, 0xC3);
+
+		int fastFailPassOffset = offset;
+		var fastFailLogInstalled = false;
+		nint kernel32 = GetModuleHandle("kernel32.dll");
+		nint getStdHandle = kernel32 != 0 ? GetProcAddress(kernel32, "GetStdHandle") : 0;
+		nint writeFile = kernel32 != 0 ? GetProcAddress(kernel32, "WriteFile") : 0;
+		if (fastFailJumpSlot >= 0 && getStdHandle != 0 && writeFile != 0)
+		{
+			// Prefix + Context.Rip hex (AMD64 CONTEXT.Rip @ 0xF8) + newline.
+			// Keep in sync with WindowsFaultHandling.CreateHandlerThunk.
+			ReadOnlySpan<byte> msg =
+				"[LOADER][FATAL] VEH_PASS FastFail 0xC0000409 (live trampoline; skip managed VEH) rip=0x"u8;
+			ReadOnlySpan<byte> hexDigits = "0123456789ABCDEF"u8;
+			// rcx=EXCEPTION_POINTERS*: capture Rip into r10 before clobbering.
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x41);
+			EmitByte(code, ref offset, 0x08); // mov rax, [rcx+8] ContextRecord*
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x90);
+			EmitUInt32(code, ref offset, 0xF8u); // mov r10, [rax+0xF8] Rip
+			EmitByte(code, ref offset, 0x50);
+			EmitByte(code, ref offset, 0x51);
+			EmitByte(code, ref offset, 0x52);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x50);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x51);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x52); // push r10 (Rip)
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x40); // sub rsp, 0x40 (hex buf @ +0x30)
+			EmitByte(code, ref offset, 0xB9); EmitUInt32(code, ref offset, unchecked((uint)-12));
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = getStdHandle;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x28); // mov [rsp+0x28], rax  stderr handle
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC1);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			var msgAbsSlot = offset;
+			*(nint*)(code + offset) = 0;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC2);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xB8);
+			EmitUInt32(code, ref offset, (uint)msg.Length);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20); EmitUInt32(code, ref offset, 0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x38); EmitUInt32(code, ref offset, 0); // lpOverlapped slot
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = writeFile;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+
+			// Hex-encode Rip. Stack after sub 0x40: [rsp+0x40]=saved Rip (push r10).
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B);
+			EmitByte(code, ref offset, 0x54); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x40); // mov r10, [rsp+0x40]
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB8);
+			var hexDigitsAbsSlot = offset;
+			*(nint*)(code + offset) = 0;
+			offset += sizeof(nint); // mov r8, hexDigits
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+			EmitByte(code, ref offset, 0x5C); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x30); // lea r11, [rsp+0x30] hex out
+			EmitByte(code, ref offset, 0xB9); EmitUInt32(code, ref offset, 16u); // ecx = 16 nibbles
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xD0); // mov rax, r10
+			int hexLoopOffset = offset;
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC1); EmitByte(code, ref offset, 0xC0);
+			EmitByte(code, ref offset, 0x04); // rol rax, 4
+			EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC2); // mov edx, eax
+			EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xE2); EmitByte(code, ref offset, 0x0F); // and edx, 0xF
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0xB6);
+			EmitByte(code, ref offset, 0x14); EmitByte(code, ref offset, 0x10); // movzx edx, byte [r8+rdx]
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x88); EmitByte(code, ref offset, 0x13); // mov [r11], dl
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xC3); // inc r11
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xC9); // dec ecx
+			EmitByte(code, ref offset, 0x75);
+			EmitByte(code, ref offset, unchecked((byte)(hexLoopOffset - (offset + 1)))); // jnz hexLoop (rel8)
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xC6); EmitByte(code, ref offset, 0x03);
+			EmitByte(code, ref offset, 0x0A); // mov byte [r11], '\n'
+
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x28); // mov rcx, [rsp+0x28] stderr
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8D);
+			EmitByte(code, ref offset, 0x54); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x30); // lea rdx, [rsp+0x30]
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xB8);
+			EmitUInt32(code, ref offset, 17u); // 16 hex + newline
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x20); EmitUInt32(code, ref offset, 0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, 0x38); EmitUInt32(code, ref offset, 0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = writeFile;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+
+			// Flush redirected stderr so FastFail rip survives process teardown.
+			nint flushFileBuffers = kernel32 != 0 ? GetProcAddress(kernel32, "FlushFileBuffers") : 0;
+			if (flushFileBuffers != 0)
+			{
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B);
+				EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+				EmitByte(code, ref offset, 0x28); // mov rcx, stderr
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+				*(nint*)(code + offset) = flushFileBuffers;
+				offset += sizeof(nint);
+				EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			}
+
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x40);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5A); // pop r10
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x59);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x58);
+			EmitByte(code, ref offset, 0x5A);
+			EmitByte(code, ref offset, 0x59);
+			EmitByte(code, ref offset, 0x58);
+			EmitByte(code, ref offset, 0x31); EmitByte(code, ref offset, 0xC0);
+			EmitByte(code, ref offset, 0xC3);
+
+			var msgOffset = offset;
+			for (int i = 0; i < msg.Length; i++)
+			{
+				EmitByte(code, ref offset, msg[i]);
+			}
+
+			var hexDigitsOffset = offset;
+			for (int i = 0; i < hexDigits.Length; i++)
+			{
+				EmitByte(code, ref offset, hexDigits[i]);
+			}
+
+			*(nint*)(code + msgAbsSlot) = (nint)ptr + msgOffset;
+			*(nint*)(code + hexDigitsAbsSlot) = (nint)ptr + hexDigitsOffset;
+			code[passJumpOffsets[fastFailJumpSlot]] =
+				checked((byte)(fastFailPassOffset - (passJumpOffsets[fastFailJumpSlot] + 1)));
+			fastFailLogInstalled = true;
+		}
+
+		int mainBodyOffset = offset;
+		*(int*)(code + mainBodyJumpSlot) = mainBodyOffset - (mainBodyJumpSlot + sizeof(int));
+		for (int i = 0; i < nonManagedExceptionCodes.Length; i++)
+		{
+			if (i == fastFailJumpSlot && fastFailLogInstalled)
+			{
+				continue;
+			}
+
+			code[passJumpOffsets[i]] = checked((byte)(passOffset - (passJumpOffsets[i] + 1)));
+		}
+
 		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x54); // push r12
 		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x55); // push r13
 		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE4); // mov r12, rsp
 		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xCD); // mov r13, rcx
+
+		// Native worker EXECUTE-AV abort without managed VEH.
+		// Do NOT catch read/write AVs — workers need managed lazy-commit (tLTJ
+		// silent-die when every worker AV was aborted). Execute faults on
+		// tbb_thead are the concurrent-managed FailFast case (tLTC).
+		int tbbFallthroughJump = -1;
+		if (_workerAbortStub != 0 &&
+			_tlsGetValueAddress != 0 &&
+			_hostRspSlotTlsIndex != uint.MaxValue)
+		{
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x8B);
+			EmitByte(code, ref offset, 0x45); EmitByte(code, ref offset, 0x00); // mov rax, [r13]
+			EmitByte(code, ref offset, 0x81); EmitByte(code, ref offset, 0x38);
+			EmitUInt32(code, ref offset, 0xC0000005u); // cmp dword [rax], AV
+			EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+			tbbFallthroughJump = offset;
+			EmitUInt32(code, ref offset, 0u); // jne fallthrough
+
+			// ExceptionInformation[0] == 8 → EXECUTE (DEP). Offset 32 on x64 RECORD.
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xB8); EmitUInt32(code, ref offset, 32u);
+			EmitByte(code, ref offset, 0x08); // cmp qword [rax+32], 8
+			EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+			var tbbNotExecuteJump = offset;
+			EmitUInt32(code, ref offset, 0u);
+
+			int tbbNotEligibleJump = -1;
+			if (_tbbAbortEligibleTlsIndex != uint.MaxValue)
+			{
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+				EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+				EmitByte(code, ref offset, 0xB9);
+				EmitUInt32(code, ref offset, _tbbAbortEligibleTlsIndex);
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+				*(nint*)(code + offset) = _tlsGetValueAddress;
+				offset += sizeof(nint);
+				EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+				EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
+				EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85);
+				EmitByte(code, ref offset, 0xC0);
+				EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+				tbbNotEligibleJump = offset;
+				EmitUInt32(code, ref offset, 0u);
+			}
+
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+			EmitByte(code, ref offset, 0xB9);
+			EmitUInt32(code, ref offset, _hostRspSlotTlsIndex);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = _tlsGetValueAddress;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85);
+			EmitByte(code, ref offset, 0xC0);
+			EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+			var tbbNoHostRspJump = offset;
+			EmitUInt32(code, ref offset, 0u);
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B);
+			EmitByte(code, ref offset, 0x00); // mov r8, [rax] hostRsp
+			EmitByte(code, ref offset, 0x4D); EmitByte(code, ref offset, 0x85);
+			EmitByte(code, ref offset, 0xC0);
+			EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+			var tbbZeroRspJump = offset;
+			EmitUInt32(code, ref offset, 0u);
+
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x83);
+			EmitByte(code, ref offset, 0xE0); EmitByte(code, ref offset, 0xF0); // and r8, ~0xF
+			EmitByte(code, ref offset, 0x4D); EmitByte(code, ref offset, 0x8B);
+			EmitByte(code, ref offset, 0x4D); EmitByte(code, ref offset, 0x08); // mov r9, [r13+8]
+			EmitByte(code, ref offset, 0x4D); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0x81); EmitUInt32(code, ref offset, 0x98u); // Context.Rsp
+			EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+			*(nint*)(code + offset) = _workerAbortStub;
+			offset += sizeof(nint);
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0x81); EmitUInt32(code, ref offset, 0xF8u); // Context.Rip
+			EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xC7);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x78);
+			EmitUInt32(code, ref offset, 0u); // Context.Rax = 0
+
+			EmitByte(code, ref offset, 0xB8); EmitUInt32(code, ref offset, unchecked((uint)-1));
+			EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89);
+			EmitByte(code, ref offset, 0xE4); // mov rsp, r12
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5D);
+			EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5C);
+			EmitByte(code, ref offset, 0xC3);
+
+			int tbbFallthroughOffset = offset;
+			*(int*)(code + tbbFallthroughJump) = tbbFallthroughOffset - (tbbFallthroughJump + sizeof(int));
+			*(int*)(code + tbbNotExecuteJump) = tbbFallthroughOffset - (tbbNotExecuteJump + sizeof(int));
+			if (tbbNotEligibleJump >= 0)
+			{
+				*(int*)(code + tbbNotEligibleJump) = tbbFallthroughOffset - (tbbNotEligibleJump + sizeof(int));
+			}
+			*(int*)(code + tbbNoHostRspJump) = tbbFallthroughOffset - (tbbNoHostRspJump + sizeof(int));
+			*(int*)(code + tbbZeroRspJump) = tbbFallthroughOffset - (tbbZeroRspJump + sizeof(int));
+		}
+
 		EmitByte(code, ref offset, 0x65); EmitByte(code, ref offset, 0x48); // mov rax, gs:[8]
 		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x04); EmitByte(code, ref offset, 0x25);
 		EmitUInt32(code, ref offset, 8u);
@@ -2440,11 +2901,67 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		EmitUInt32(code, ref offset, 0u);
 
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+		// Serialize managed VEH entry (recursive spinlock). Concurrent UnmanagedCallersOnly
+		// FailFast was the tLTQ silent mid-TBB pattern (enter without abort breadcrumb).
+		// Lock layout: [0]=owner UniqueThread (nint), [8]=depth (int).
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB9);
+		*(nint*)(code + offset) = _vehManagedEntryLock;
+		offset += sizeof(nint); // mov r9, lock*
+		EmitByte(code, ref offset, 0x65); EmitByte(code, ref offset, 0x4C);
+		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x14);
+		EmitByte(code, ref offset, 0x25); EmitUInt32(code, ref offset, 0x48u); // mov r10, gs:[0x48]
+		int hostAcquireSpin = offset;
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x01); // mov rax, [r9]
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x39); EmitByte(code, ref offset, 0xD0); // cmp rax, r10
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+		int hostMineJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85); EmitByte(code, ref offset, 0xC0); // test rax, rax
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int hostPauseJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0xF0); EmitByte(code, ref offset, 0x4C);
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0xB1); EmitByte(code, ref offset, 0x11); // lock cmpxchg [r9], r10
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int hostRetryJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xC7);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x08);
+		EmitUInt32(code, ref offset, 1u); // mov dword [r9+8], 1
+		EmitByte(code, ref offset, 0xE9);
+		int hostGotJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		int hostPauseOffset = offset;
+		EmitByte(code, ref offset, 0xF3); EmitByte(code, ref offset, 0x90); // pause
+		EmitByte(code, ref offset, 0xE9);
+		int hostPauseBackJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		int hostMineOffset = offset;
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xFF);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x08); // inc dword [r9+8]
+		int hostGotOffset = offset;
+		*(int*)(code + hostMineJump) = hostMineOffset - (hostMineJump + sizeof(int));
+		*(int*)(code + hostPauseJump) = hostPauseOffset - (hostPauseJump + sizeof(int));
+		*(int*)(code + hostRetryJump) = hostAcquireSpin - (hostRetryJump + sizeof(int));
+		*(int*)(code + hostGotJump) = hostGotOffset - (hostGotJump + sizeof(int));
+		*(int*)(code + hostPauseBackJump) = hostAcquireSpin - (hostPauseBackJump + sizeof(int));
 		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE9); // mov rcx, r13
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
 		*(nint*)(code + offset) = managedHandler;
 		offset += sizeof(nint);
 		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB9);
+		*(nint*)(code + offset) = _vehManagedEntryLock;
+		offset += sizeof(nint); // mov r9, lock*
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xFF);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x08); // dec dword [r9+8]
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int hostStillJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xC7);
+		EmitByte(code, ref offset, 0x01); EmitUInt32(code, ref offset, 0u); // mov qword [r9], 0
+		int hostStillOffset = offset;
+		*(int*)(code + hostStillJump) = hostStillOffset - (hostStillJump + sizeof(int));
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
 		EmitByte(code, ref offset, 0xE9);
 		int hostRestoreJump = offset;
@@ -2470,11 +2987,64 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		EmitUInt32(code, ref offset, 0u);
 		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xDC); // mov rsp, r11
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB9);
+		*(nint*)(code + offset) = _vehManagedEntryLock;
+		offset += sizeof(nint); // mov r9, lock*
+		EmitByte(code, ref offset, 0x65); EmitByte(code, ref offset, 0x4C);
+		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x14);
+		EmitByte(code, ref offset, 0x25); EmitUInt32(code, ref offset, 0x48u); // mov r10, gs:[0x48]
+		int guestAcquireSpin = offset;
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x01); // mov rax, [r9]
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x39); EmitByte(code, ref offset, 0xD0); // cmp rax, r10
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+		int guestMineJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85); EmitByte(code, ref offset, 0xC0);
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int guestPauseJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0xF0); EmitByte(code, ref offset, 0x4C);
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0xB1); EmitByte(code, ref offset, 0x11);
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int guestRetryJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xC7);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x08);
+		EmitUInt32(code, ref offset, 1u);
+		EmitByte(code, ref offset, 0xE9);
+		int guestGotJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		int guestPauseOffset = offset;
+		EmitByte(code, ref offset, 0xF3); EmitByte(code, ref offset, 0x90);
+		EmitByte(code, ref offset, 0xE9);
+		int guestPauseBackJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		int guestMineOffset = offset;
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xFF);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x08);
+		int guestGotOffset = offset;
+		*(int*)(code + guestMineJump) = guestMineOffset - (guestMineJump + sizeof(int));
+		*(int*)(code + guestPauseJump) = guestPauseOffset - (guestPauseJump + sizeof(int));
+		*(int*)(code + guestRetryJump) = guestAcquireSpin - (guestRetryJump + sizeof(int));
+		*(int*)(code + guestGotJump) = guestGotOffset - (guestGotJump + sizeof(int));
+		*(int*)(code + guestPauseBackJump) = guestAcquireSpin - (guestPauseBackJump + sizeof(int));
 		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE9); // mov rcx, r13
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
 		*(nint*)(code + offset) = managedHandler;
 		offset += sizeof(nint);
 		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB9);
+		*(nint*)(code + offset) = _vehManagedEntryLock;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xFF);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x08); // dec dword [r9+8]
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x85);
+		int guestStillJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xC7);
+		EmitByte(code, ref offset, 0x01); EmitUInt32(code, ref offset, 0u);
+		int guestStillOffset = offset;
+		*(int*)(code + guestStillJump) = guestStillOffset - (guestStillJump + sizeof(int));
 		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
 		EmitByte(code, ref offset, 0xE9);
 		int guestRestoreJump = offset;
@@ -2495,6 +3065,18 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		*(int*)(code + missingHostStackJump) = passThroughOffset - (missingHostStackJump + sizeof(int));
 		*(int*)(code + guestRestoreJump) = restoreOffset - (guestRestoreJump + sizeof(int));
 
+		if (offset > (int)stubSize)
+		{
+			Console.Error.WriteLine(
+				$"[LOADER][ERROR] Exception handler trampoline overflow: used={offset} cap={stubSize}");
+			VirtualFree(ptr, 0, 0x8000u);
+			return 0;
+		}
+
+		Console.Error.WriteLine(
+			$"[LOADER][INFO] VEH trampoline built: bytes={offset} native_worker_abort=" +
+			$"{(_workerAbortStub != 0 && _hostRspSlotTlsIndex != uint.MaxValue)}");
+
 		uint oldProtect = default;
 		VirtualProtect(ptr, stubSize, 32u, &oldProtect);
 		FlushInstructionCache(GetCurrentProcess(), ptr, (nuint)offset);
@@ -5149,16 +5731,27 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				return GuestNativeCallExitReason.Exception;
 			}
 			FlushInstructionCache(GetCurrentProcess(), ptr, stubSize);
-			if (!TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot))
-			{
-				reason = "failed to bind host-RSP storage for guest thread stub";
-				return GuestNativeCallExitReason.Exception;
-			}
 			ActiveGuestThreadYieldRequested = false;
 			ActiveGuestThreadYieldReason = null;
 			try
 			{
-				var nativeReturn = CallNativeEntry(ptr);
+				// TBB execute-AV recover needs native-worker TLS (eligible/done).
+				// Other guests stay on CallNativeEntry — full native-worker migration
+				// increased splash hangs / UnmanagedCallersOnly (tLTN/tLTO).
+				int nativeReturn;
+				if (name == "tbb_thead")
+				{
+					nativeReturn = RunGuestEntryStub(ptr, hostRspSlot, requireNativeWorker: true);
+				}
+				else
+				{
+					if (!TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot))
+					{
+						reason = "failed to bind host-RSP storage for guest thread stub";
+						return GuestNativeCallExitReason.Exception;
+					}
+					nativeReturn = CallNativeEntry(ptr);
+				}
 				if (ActiveGuestThreadYieldRequested)
 				{
 					reason = ActiveGuestThreadYieldReason ?? "guest thread blocked";
@@ -5304,16 +5897,24 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				return GuestNativeCallExitReason.Exception;
 			}
 			FlushInstructionCache(GetCurrentProcess(), ptr, stubSize);
-			if (!TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot))
-			{
-				reason = "failed to bind host-RSP storage for guest continuation stub";
-				return GuestNativeCallExitReason.Exception;
-			}
 			ActiveGuestThreadYieldRequested = false;
 			ActiveGuestThreadYieldReason = null;
 			try
 			{
-				var nativeReturn = CallNativeEntry(ptr);
+				int nativeReturn;
+				if (name == "tbb_thead")
+				{
+					nativeReturn = RunGuestEntryStub(ptr, hostRspSlot, requireNativeWorker: true);
+				}
+				else
+				{
+					if (!TlsSetValue(_hostRspSlotTlsIndex, (nint)hostRspSlot))
+					{
+						reason = "failed to bind host-RSP storage for guest continuation stub";
+						return GuestNativeCallExitReason.Exception;
+					}
+					nativeReturn = CallNativeEntry(ptr);
+				}
 				if (ActiveGuestThreadYieldRequested)
 				{
 					reason = ActiveGuestThreadYieldReason ?? "guest thread blocked";
@@ -6446,6 +7047,16 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			VirtualFree((void*)_hostRspSlotStorage, 0u, 32768u);
 			_hostRspSlotStorage = 0;
 		}
+		if (_vehManagedEntryLock != 0)
+		{
+			VirtualFree((void*)_vehManagedEntryLock, 0u, 32768u);
+			_vehManagedEntryLock = 0;
+		}
+		if (_workerAbortStack != 0)
+		{
+			VirtualFree((void*)_workerAbortStack, 0u, 32768u);
+			_workerAbortStack = 0;
+		}
 		if (_guestTlsBaseTlsIndex != uint.MaxValue)
 		{
 			TlsFree(_guestTlsBaseTlsIndex);
@@ -6456,6 +7067,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			TlsFree(_hostRspSlotTlsIndex);
 			_hostRspSlotTlsIndex = uint.MaxValue;
 		}
+		if (_workerDoneEventTlsIndex != uint.MaxValue)
+		{
+			TlsFree(_workerDoneEventTlsIndex);
+			_workerDoneEventTlsIndex = uint.MaxValue;
+		}
 		if (_unresolvedReturnStub != 0)
 		{
 			VirtualFree((void*)_unresolvedReturnStub, 0u, 32768u);
@@ -6466,6 +7082,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			VirtualFree((void*)_guestReturnStub, 0u, 32768u);
 			_guestReturnStub = 0;
 		}
+		if (_workerAbortStub != 0)
+		{
+			VirtualFree((void*)_workerAbortStub, 0u, 32768u);
+			_workerAbortStub = 0;
+		}
 		if (_guestContextTransferStub != 0)
 		{
 			VirtualFree((void*)_guestContextTransferStub, 0u, 32768u);

src/SharpEmu.Core/Cpu/Native/Windows/WindowsFaultHandling.cs

@@ -24,7 +24,7 @@ internal sealed unsafe partial class WindowsFaultHandling : IHostFaultHandling
 
     public nint CreateHandlerThunk(nint managedCallback, uint hostRspSwitchTlsSlot, nint tlsGetValueAddress)
     {
-        const uint stubSize = 256u;
+        const uint stubSize = 1024u;
         void* ptr = (void*)_memory.Allocate(0, stubSize, HostPageProtection.ReadWriteExecute);
         if (ptr == null)
         {
@@ -43,11 +43,15 @@ internal sealed unsafe partial class WindowsFaultHandling : IHostFaultHandling
         // managed code; the CLR's own VEH handles its exceptions. MSVC C++ exceptions
         // (Vulkan drivers, host CRT) are excluded too: the managed handler only ever
         // returned CONTINUE_SEARCH for them.
+        //
+        // FastFail (0xC0000409) is logged from this native path only: managed VEH never
+        // sees it (tLT18–21 silent exits after TBB AV recovery).
         ReadOnlySpan<uint> nonManagedExceptionCodes =
             [WindowsFaultCodes.ClrManagedException, 0xE06D7363u, WindowsFaultCodes.FastFail, WindowsFaultCodes.StackOverflow];
         EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x01); // mov rax, [rcx] (ExceptionRecord*)
         EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x00);                                   // mov eax, [rax] (ExceptionCode)
         var passJumpOffsets = stackalloc int[nonManagedExceptionCodes.Length];
+        int fastFailJumpSlot = -1;
         for (int i = 0; i < nonManagedExceptionCodes.Length; i++)
         {
             EmitByte(code, ref offset, 0x3D);                                                                 // cmp eax, imm32
@@ -55,13 +59,162 @@ internal sealed unsafe partial class WindowsFaultHandling : IHostFaultHandling
             EmitByte(code, ref offset, 0x74);                                                                 // je pass
             passJumpOffsets[i] = offset;
             EmitByte(code, ref offset, 0x00);
+            if (nonManagedExceptionCodes[i] == WindowsFaultCodes.FastFail)
+            {
+                fastFailJumpSlot = i;
+            }
         }
-        EmitByte(code, ref offset, 0xEB); EmitByte(code, ref offset, 0x03);                                   // jmp over pass block
+        EmitByte(code, ref offset, 0xE9); // jmp mainBody rel32 (FastFail breadcrumb sits between)
+        var mainBodyJumpSlot = offset;
+        EmitUInt32(code, ref offset, 0u);
+
         int passOffset = offset;
         EmitByte(code, ref offset, 0x31); EmitByte(code, ref offset, 0xC0);                                   // pass: xor eax, eax (EXCEPTION_CONTINUE_SEARCH)
         EmitByte(code, ref offset, 0xC3);                                                                     // ret
+
+        // FastFail: native stderr breadcrumb with Context.Rip (no managed entry), then CONTINUE_SEARCH.
+        // Keep in sync with DirectExecutionBackend.CreateExceptionHandlerTrampoline.
+        int fastFailPassOffset = offset;
+        var fastFailLogInstalled = false;
+        if (fastFailJumpSlot >= 0 &&
+            NativeLibrary.TryLoad("kernel32.dll", out var kernel32) &&
+            NativeLibrary.TryGetExport(kernel32, "GetStdHandle", out var getStdHandle) &&
+            NativeLibrary.TryGetExport(kernel32, "WriteFile", out var writeFile))
+        {
+            ReadOnlySpan<byte> msg =
+                "[LOADER][FATAL] VEH_PASS FastFail 0xC0000409 (native; no managed VEH) rip=0x"u8;
+            ReadOnlySpan<byte> hexDigits = "0123456789ABCDEF"u8;
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x41);
+            EmitByte(code, ref offset, 0x08); // mov rax, [rcx+8]
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x90);
+            EmitUInt32(code, ref offset, 0xF8u); // mov r10, [rax+0xF8]
+            EmitByte(code, ref offset, 0x50); // push rax
+            EmitByte(code, ref offset, 0x51); // push rcx
+            EmitByte(code, ref offset, 0x52); // push rdx
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x50); // push r8
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x51); // push r9
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x52); // push r10
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+            EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x40); // sub rsp, 0x40
+            EmitByte(code, ref offset, 0xB9); EmitUInt32(code, ref offset, unchecked((uint)-12));
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+            *(nint*)(code + offset) = getStdHandle;
+            offset += sizeof(nint);
+            EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89);
+            EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x28); // mov [rsp+0x28], rax
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC1);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+            var msgAbsSlot = offset;
+            *(nint*)(code + offset) = 0;
+            offset += sizeof(nint);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC2);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xB8);
+            EmitUInt32(code, ref offset, (uint)msg.Length);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x20);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+            EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x20); EmitUInt32(code, ref offset, 0);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+            EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x38); EmitUInt32(code, ref offset, 0);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+            *(nint*)(code + offset) = writeFile;
+            offset += sizeof(nint);
+            EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B);
+            EmitByte(code, ref offset, 0x54); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x40); // mov r10, [rsp+0x40]
+            EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xB8);
+            var hexDigitsAbsSlot = offset;
+            *(nint*)(code + offset) = 0;
+            offset += sizeof(nint);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+            EmitByte(code, ref offset, 0x5C); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x30); // lea r11, [rsp+0x30]
+            EmitByte(code, ref offset, 0xB9); EmitUInt32(code, ref offset, 16u);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xD0);
+            int hexLoopOffset = offset;
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC1); EmitByte(code, ref offset, 0xC0);
+            EmitByte(code, ref offset, 0x04);
+            EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xC2);
+            EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xE2); EmitByte(code, ref offset, 0x0F);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0xB6);
+            EmitByte(code, ref offset, 0x14); EmitByte(code, ref offset, 0x10);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x88); EmitByte(code, ref offset, 0x13);
+            EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xC3);
+            EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xC9);
+            EmitByte(code, ref offset, 0x75);
+            EmitByte(code, ref offset, unchecked((byte)(hexLoopOffset - (offset + 1)))); // jnz rel8
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xC6); EmitByte(code, ref offset, 0x03);
+            EmitByte(code, ref offset, 0x0A);
+
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8B);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x28);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x8D);
+            EmitByte(code, ref offset, 0x54); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x30);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0xB8);
+            EmitUInt32(code, ref offset, 17u);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8D);
+            EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x20);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+            EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x20); EmitUInt32(code, ref offset, 0);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xC7);
+            EmitByte(code, ref offset, 0x44); EmitByte(code, ref offset, 0x24);
+            EmitByte(code, ref offset, 0x38); EmitUInt32(code, ref offset, 0);
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+            *(nint*)(code + offset) = writeFile;
+            offset += sizeof(nint);
+            EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+
+            EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83);
+            EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x40);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5A);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x59);
+            EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x58);
+            EmitByte(code, ref offset, 0x5A);
+            EmitByte(code, ref offset, 0x59);
+            EmitByte(code, ref offset, 0x58);
+            EmitByte(code, ref offset, 0x31); EmitByte(code, ref offset, 0xC0);
+            EmitByte(code, ref offset, 0xC3);
+
+            var msgOffset = offset;
+            for (int i = 0; i < msg.Length; i++)
+            {
+                EmitByte(code, ref offset, msg[i]);
+            }
+
+            var hexDigitsOffset = offset;
+            for (int i = 0; i < hexDigits.Length; i++)
+            {
+                EmitByte(code, ref offset, hexDigits[i]);
+            }
+
+            *(nint*)(code + msgAbsSlot) = (nint)ptr + msgOffset;
+            *(nint*)(code + hexDigitsAbsSlot) = (nint)ptr + hexDigitsOffset;
+            code[passJumpOffsets[fastFailJumpSlot]] =
+                checked((byte)(fastFailPassOffset - (passJumpOffsets[fastFailJumpSlot] + 1)));
+            fastFailLogInstalled = true;
+        }
+
+        int mainBodyOffset = offset;
+        *(int*)(code + mainBodyJumpSlot) = mainBodyOffset - (mainBodyJumpSlot + sizeof(int));
+
         for (int i = 0; i < nonManagedExceptionCodes.Length; i++)
         {
+            if (i == fastFailJumpSlot && fastFailLogInstalled)
+            {
+                continue;
+            }
+
             code[passJumpOffsets[i]] = checked((byte)(passOffset - (passJumpOffsets[i] + 1)));
         }
         EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x54); // push r12

src/SharpEmu.Core/Memory/PhysicalVirtualMemory.cs

@@ -238,7 +238,15 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         var alignedSize = (size + 0xFFF) & ~0xFFFUL;
         var protection = executable ? PAGE_EXECUTE_READWRITE : PAGE_READWRITE;
         var hostProtection = executable ? HostPageProtection.ReadWriteExecute : HostPageProtection.ReadWrite;
-        var result = _hostMemory.Allocate(desiredAddress, alignedSize, hostProtection);
+
+        // Reserve address space only for very large non-executable regions; commit is done lazily later.
+        var reservedOnly = !executable &&
+            alignedSize >= LargeDataReserveThreshold &&
+            alignedSize > FullCommitRegionLimit;
+
+        var result = reservedOnly
+            ? _hostMemory.Reserve(desiredAddress, alignedSize, HostPageProtection.ReadWrite)
+            : _hostMemory.Allocate(desiredAddress, alignedSize, hostProtection);
         if (result == 0)
         {
             return false;
@@ -252,6 +260,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             return false;
         }
 
+        var state = reservedOnly ? ReserveRegion(actualAddress, alignedSize) : "n/a";
+
         _gate.EnterWriteLock();
         try
         {
@@ -260,7 +270,7 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 VirtualAddress = actualAddress,
                 Size = alignedSize,
                 IsExecutable = executable,
-                IsReservedOnly = false,
+                IsReservedOnly = reservedOnly,
                 Protection = protection
             });
         }
@@ -269,6 +279,7 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             _gate.ExitWriteLock();
         }
 
+
         var allocationKind = executable ? "executable memory" : "data memory";
         TraceVmem($"Allocated exact {allocationKind}: 0x{actualAddress:X16} - 0x{actualAddress + alignedSize:X16} ({alignedSize} bytes)");
         return true;
@@ -361,44 +372,7 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
 
         var actualAddress = result;
 
-        var lazyPrimeState = "n/a";
-        if (reservedOnly)
-        {
-            var primeBytes = Math.Min(alignedSize, LazyReservePrimeBytes);
-            if (primeBytes != 0)
-            {
-                ulong committedBytes = 0;
-                while (committedBytes < primeBytes)
-                {
-                    var remaining = primeBytes - committedBytes;
-                    var chunkBytes = Math.Min(remaining, LazyReservePrimeChunkBytes);
-                    var commitAddress = actualAddress + committedBytes;
-                    if (!_hostMemory.Commit(commitAddress, chunkBytes, HostPageProtection.ReadWrite))
-                    {
-                        break;
-                    }
-
-                    committedBytes += chunkBytes;
-                }
-
-                if (committedBytes != 0)
-                {
-                    lazyPrimeState = committedBytes == primeBytes
-                        ? $"ok:{committedBytes:X}"
-                        : $"partial:{committedBytes:X}/{primeBytes:X}";
-                    TraceVmem($"Primed lazy region: 0x{actualAddress:X16} - 0x{actualAddress + committedBytes:X16} ({committedBytes} bytes)");
-                }
-                else
-                {
-                    lazyPrimeState = $"fail:{primeBytes:X}";
-                    TraceVmem($"Failed to prime lazy region at 0x{actualAddress:X16} ({primeBytes} bytes), continuing with on-demand commit");
-                }
-            }
-            else
-            {
-                lazyPrimeState = "skip:0";
-            }
-        }
+        var lazyPrimeState = reservedOnly ? ReserveRegion(actualAddress, alignedSize) : "n/a";
 
         _gate.EnterWriteLock();
         try
@@ -425,6 +399,41 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         return actualAddress;
     }
 
+    private string ReserveRegion(ulong actualAddress, ulong alignedSize)
+    {
+        var primeBytes = Math.Min(alignedSize, LazyReservePrimeBytes);
+        if (primeBytes == 0)
+        {
+            return "skip:0";
+        }
+
+        ulong committedBytes = 0;
+        while (committedBytes < primeBytes)
+        {
+            var remaining = primeBytes - committedBytes;
+            var chunkBytes = Math.Min(remaining, LazyReservePrimeChunkBytes);
+            var commitAddress = actualAddress + committedBytes;
+            if (!_hostMemory.Commit(commitAddress, chunkBytes, HostPageProtection.ReadWrite))
+            {
+                break;
+            }
+
+            committedBytes += chunkBytes;
+        }
+
+        if (committedBytes != 0)
+        {
+            var state = committedBytes == primeBytes
+                ? $"ok:{committedBytes:X}"
+                : $"partial:{committedBytes:X}/{primeBytes:X}";
+            TraceVmem($"region: 0x{actualAddress:X16} - 0x{actualAddress + committedBytes:X16} ({committedBytes} bytes)");
+            return state;
+        }
+
+        TraceVmem($"Failed to reserve region at 0x{actualAddress:X16} ({primeBytes} bytes)!");
+        return $"fail:{primeBytes:X}";
+    }
+
     public bool TryBackFixedRange(ulong address, ulong size, bool executable)
     {
         if (size == 0)

src/SharpEmu.HLE/GuestThreadExecution.cs

@@ -221,6 +221,29 @@ public static class GuestThreadExecution
 
     public static IGuestThreadScheduler? Scheduler { get; set; }
 
+    /// <summary>
+    /// Fired when a guest thread is torn down without a clean pthread_exit
+    /// (e.g. TBB execute-AV → worker_abort). Libs use this to abandon mutexes.
+    /// </summary>
+    public static event Func<ulong, string, int>? GuestThreadAbandoned;
+
+    public static int NotifyGuestThreadAbandoned(ulong threadHandle, string reason)
+    {
+        if (threadHandle == 0 || GuestThreadAbandoned is null)
+        {
+            return 0;
+        }
+
+        try
+        {
+            return GuestThreadAbandoned.Invoke(threadHandle, reason);
+        }
+        catch
+        {
+            return 0;
+        }
+    }
+
     public static bool IsGuestThread => _currentGuestThreadHandle != 0;
 
     public static ulong CurrentGuestThreadHandle => _currentGuestThreadHandle;

src/SharpEmu.Libs/Agc/GnmTiling.cs

@@ -6,6 +6,55 @@ using System.Runtime.CompilerServices;
 
 namespace SharpEmu.Libs.Agc;
 
+/// <summary>Which in-block address equation a <see cref="DetileParams"/> carries.</summary>
+internal enum DetileEquation
+{
+    /// <summary>Unsupported mode/format; caller must use the CPU path or raw upload.</summary>
+    None,
+
+    /// <summary>Exact AddrLib XOR equation (RDNA2 modes 5/9/24/27): factored X/Y terms.</summary>
+    ExactXor,
+
+    /// <summary>Other modes: a precomputed in-block Morton/standard element-offset table.</summary>
+    BlockTable,
+}
+
+/// <summary>
+/// Backend-agnostic description of how to deswizzle one surface, produced by
+/// <see cref="GnmTiling.GetDetileParams"/>. Holds only plain integers and small
+/// int[] tables — no host graphics-API types — so it can cross the guest-GPU
+/// backend seam and drive a Vulkan (SPIR-V) or Metal (MSL) detile compute kernel
+/// identically to the CPU <see cref="GnmTiling.TryDetile"/> fallback. The single
+/// shared addressing formula both consume is:
+/// <code>
+///   inBlockByte = Equation == ExactXor
+///       ? XByteTerm[x &amp; XMask] ^ YByteTerm[y &amp; YMask]
+///       : BlockTable[(y % BlockHeight) * BlockWidth + (x % BlockWidth)] * BytesPerElement;
+///   srcByte = ((y / BlockHeight) * BlocksPerRow + (x / BlockWidth)) * BlockBytes + inBlockByte;
+/// </code>
+/// </summary>
+internal readonly record struct DetileParams(
+    DetileEquation Equation,
+    int ElementsWide,
+    int ElementsHigh,
+    int BytesPerElement,
+    int BlockWidth,
+    int BlockHeight,
+    int BlockElements,
+    int BlockBytes,
+    int BlocksPerRow,
+    // ExactXor: within-block BYTE offset = XByteTerm[x & XMask] ^ YByteTerm[y & YMask].
+    int[] XByteTerm,
+    int XMask,
+    int[] YByteTerm,
+    int YMask,
+    // BlockTable: within-block ELEMENT offset = BlockTable[inBlockY * BlockWidth + inBlockX].
+    int[] BlockTable)
+{
+    /// <summary>False when the mode/format is not GPU-portable (Equation == None).</summary>
+    public bool IsSupported => Equation != DetileEquation.None;
+}
+
 /// <summary>
 /// Deswizzles RDNA2 (GFX10) tiled texture surfaces into linear layout so they
 /// can be uploaded to Vulkan. PS5 stores most textures in a swizzled layout
@@ -501,6 +550,141 @@ internal static unsafe class GnmTiling
         return true;
     }
 
+    /// <summary>
+    /// Computes the detile parameters for a surface without performing the copy,
+    /// so a GPU compute kernel can run the deswizzle instead of the CPU. Returns
+    /// <see cref="DetileParams.IsSupported"/> == false (Equation == None) when the
+    /// mode/format is not GPU-portable, so the caller keeps the CPU
+    /// <see cref="TryDetile"/> path or a raw upload. Reuses the same helpers and
+    /// caches as <see cref="TryDetile"/>, so the two never disagree on addressing.
+    /// </summary>
+    public static DetileParams GetDetileParams(
+        uint swizzleMode,
+        int bytesPerElement,
+        int elementsWide,
+        int elementsHigh)
+    {
+        if (!ShouldDetile(swizzleMode) ||
+            bytesPerElement <= 0 ||
+            elementsWide <= 0 ||
+            elementsHigh <= 0 ||
+            !TryGetSwizzleKind(swizzleMode, out var kind, out var blockBytes))
+        {
+            return default;
+        }
+
+        var bppLog2 = BitLog2((uint)bytesPerElement);
+        if (bppLog2 < 0)
+        {
+            return default;
+        }
+
+        var blockElements = blockBytes >> bppLog2;
+        var (blockWidth, blockHeight) = SquareBlockDimensions(blockElements);
+        if (blockWidth == 0 || blockHeight == 0)
+        {
+            return default;
+        }
+
+        var blocksPerRow = (elementsWide + blockWidth - 1) / blockWidth;
+
+        if (TryGetExactXorPattern(swizzleMode, bppLog2, out var pattern))
+        {
+            var terms = _patternTermCache.GetOrAdd(
+                (swizzleMode, bppLog2),
+                _ => CreatePatternTerms(pattern));
+            return new DetileParams(
+                DetileEquation.ExactXor,
+                elementsWide,
+                elementsHigh,
+                bytesPerElement,
+                blockWidth,
+                blockHeight,
+                blockElements,
+                blockBytes,
+                blocksPerRow,
+                terms.X,
+                terms.XMask,
+                terms.Y,
+                terms.YMask,
+                []);
+        }
+
+        var blockTable = _blockTableCache.GetOrAdd(
+            (kind, blockWidth, blockHeight),
+            static key => CreateBlockTable(key.Kind, key.Width, key.Height));
+        return new DetileParams(
+            DetileEquation.BlockTable,
+            elementsWide,
+            elementsHigh,
+            bytesPerElement,
+            blockWidth,
+            blockHeight,
+            blockElements,
+            blockBytes,
+            blocksPerRow,
+            [],
+            0,
+            [],
+            0,
+            blockTable);
+    }
+
+    /// <summary>
+    /// CPU deswizzle driven entirely by a resolved <see cref="DetileParams"/> — the
+    /// exact addressing the Vulkan/Metal compute kernel runs per texel, so a
+    /// backend that packaged <paramref name="parameters"/> for the GPU path can
+    /// fall back to this without re-deriving the swizzle. Copies
+    /// <c>ElementsWide * ElementsHigh</c> elements from <paramref name="tiled"/>
+    /// into <paramref name="linear"/>; returns false when unsupported or the output
+    /// span is too small. Out-of-range source elements are left zero (matching the
+    /// reference), so a truncated <paramref name="tiled"/> degrades gracefully.
+    /// </summary>
+    public static bool DetileWithParams(
+        in DetileParams parameters,
+        ReadOnlySpan<byte> tiled,
+        Span<byte> linear)
+    {
+        if (!parameters.IsSupported)
+        {
+            return false;
+        }
+
+        var width = parameters.ElementsWide;
+        var height = parameters.ElementsHigh;
+        var bpp = parameters.BytesPerElement;
+        var requiredLinear = (long)width * height * bpp;
+        if (width <= 0 || height <= 0 || bpp <= 0 || linear.Length < requiredLinear)
+        {
+            return false;
+        }
+
+        var isExactXor = parameters.Equation == DetileEquation.ExactXor;
+        for (var y = 0; y < height; y++)
+        {
+            var blockY = y / parameters.BlockHeight;
+            var inY = y % parameters.BlockHeight;
+            var yTerm = isExactXor ? parameters.YByteTerm[y & parameters.YMask] : 0;
+            for (var x = 0; x < width; x++)
+            {
+                var blockX = x / parameters.BlockWidth;
+                var inBlockByte = isExactXor
+                    ? parameters.XByteTerm[x & parameters.XMask] ^ yTerm
+                    : parameters.BlockTable[inY * parameters.BlockWidth + (x % parameters.BlockWidth)] * bpp;
+                var srcByte = ((long)blockY * parameters.BlocksPerRow + blockX) * parameters.BlockBytes + inBlockByte;
+                var dstByte = ((long)y * width + x) * bpp;
+                if (srcByte < 0 || srcByte + bpp > tiled.Length)
+                {
+                    continue;
+                }
+
+                tiled.Slice((int)srcByte, bpp).CopyTo(linear.Slice((int)dstByte, bpp));
+            }
+        }
+
+        return true;
+    }
+
     private enum SwizzleKind
     {
         Standard,

src/SharpEmu.Libs/Audio/AudioOutExports.cs

@@ -4,6 +4,7 @@
 using SharpEmu.HLE;
 using SharpEmu.HLE.Host;
 using System.Buffers;
+using System.Buffers.Binary;
 using System.Collections.Concurrent;
 using System.Diagnostics;
 
@@ -11,8 +12,17 @@ namespace SharpEmu.Libs.Audio;
 
 public static class AudioOutExports
 {
+    private const int AudioOutOutputParamSize = 16;
+    private const int AudioOutMaximumOutputCount = 25;
+
+    internal const int AudioOutErrorInvalidPort = unchecked((int)0x80260003);
+    internal const int AudioOutErrorInvalidPointer = unchecked((int)0x80260004);
+    internal const int AudioOutErrorPortFull = unchecked((int)0x80260005);
+    internal const int AudioOutErrorInvalidSize = unchecked((int)0x80260006);
+
     private static readonly ConcurrentDictionary<int, PortState> Ports = new();
     private static int _nextPortHandle;
+    private static Func<uint, IHostAudioStream?>? _streamFactoryForTests;
 
     // Diagnostic: confirm sceAudioOutOutput is actually called and whether the
     // guest submits real samples or silence. Gated so it costs nothing when off.
@@ -56,6 +66,7 @@ public static class AudioOutExports
         public int BytesPerSample { get; }
         public bool IsFloat { get; }
         public IHostAudioStream? Backend { get; }
+        public object SubmissionGate { get; } = new();
         public volatile float Volume = 1.0f;
         public int BufferByteLength =>
             checked((int)BufferLength * Channels * BytesPerSample);
@@ -83,7 +94,24 @@ public static class AudioOutExports
             }
         }
 
-        public void Dispose() => Backend?.Dispose();
+        public void Dispose()
+        {
+            lock (SubmissionGate)
+            {
+                Backend?.Dispose();
+            }
+        }
+    }
+
+    private readonly record struct OutputDescriptor(int Handle, ulong SourceAddress);
+
+    private struct ResolvedOutput
+    {
+        public int Handle;
+        public ulong SourceAddress;
+        public PortState Port;
+        public byte[]? HostBuffer;
+        public int HostBufferLength;
     }
 
     [SysAbiExport(
@@ -115,9 +143,18 @@ public static class AudioOutExports
         string backendName;
         try
         {
-            var audio = HostPlatform.Current.Audio;
-            backend = audio.OpenStereoPcm16Stream(frequency);
-            backendName = audio.BackendName;
+            var streamFactory = Volatile.Read(ref _streamFactoryForTests);
+            if (streamFactory is not null)
+            {
+                backend = streamFactory(frequency);
+                backendName = "test";
+            }
+            else
+            {
+                var audio = HostPlatform.Current.Audio;
+                backend = audio.OpenStereoPcm16Stream(frequency);
+                backendName = audio.BackendName;
+            }
         }
         catch (Exception exception)
         {
@@ -192,6 +229,46 @@ public static class AudioOutExports
         return ctx.SetReturn(0);
     }
 
+    [SysAbiExport(
+        Nid = "w3PdaSTSwGE",
+        ExportName = "sceAudioOutOutputs",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceAudioOut")]
+    public static int AudioOutOutputs(CpuContext ctx)
+    {
+        var parameterAddress = ctx[CpuRegister.Rdi];
+        var outputCount = unchecked((uint)ctx[CpuRegister.Rsi]);
+        if (outputCount == 0 || outputCount > AudioOutMaximumOutputCount)
+        {
+            return ctx.SetReturn(AudioOutErrorPortFull);
+        }
+
+        if (parameterAddress == 0)
+        {
+            return ctx.SetReturn(AudioOutErrorInvalidPointer);
+        }
+
+        var count = checked((int)outputCount);
+        Span<byte> parameterBytes =
+            stackalloc byte[AudioOutMaximumOutputCount * AudioOutOutputParamSize];
+        parameterBytes = parameterBytes[..checked(count * AudioOutOutputParamSize)];
+        if (!ctx.Memory.TryRead(parameterAddress, parameterBytes))
+        {
+            return ctx.SetReturn(AudioOutErrorInvalidPointer);
+        }
+
+        Span<OutputDescriptor> descriptors = stackalloc OutputDescriptor[count];
+        for (var i = 0; i < count; i++)
+        {
+            var entry = parameterBytes.Slice(i * AudioOutOutputParamSize, AudioOutOutputParamSize);
+            descriptors[i] = new OutputDescriptor(
+                BinaryPrimitives.ReadInt32LittleEndian(entry),
+                BinaryPrimitives.ReadUInt64LittleEndian(entry[8..]));
+        }
+
+        return ctx.SetReturn(SubmitOutputs(ctx, descriptors));
+    }
+
     [SysAbiExport(
         Nid = "QOQtbeDqsT4",
         ExportName = "sceAudioOutOutput",
@@ -225,16 +302,7 @@ public static class AudioOutExports
                 return ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
             }
 
-            if (_traceOutput)
-            {
-                var n = Interlocked.Increment(ref _outputCount);
-                if (n <= 8 || n % 200 == 0)
-                {
-                    var peak = PeakAmplitude(source, port.IsFloat, port.BytesPerSample);
-                    Console.Error.WriteLine(
-                        $"[LOADER][TRACE] audioout.output#{n} handle={handle} bytes={source.Length} ch={port.Channels} float={port.IsFloat} vol={port.Volume:F2} peak={peak:F4} backend={(port.Backend is null ? "none" : "coreaudio")}");
-                }
-            }
+            TraceOutput(handle, port, source);
 
             if (port.Backend is null)
             {
@@ -272,6 +340,184 @@ public static class AudioOutExports
         }
     }
 
+    private static int SubmitOutputs(CpuContext ctx, ReadOnlySpan<OutputDescriptor> descriptors)
+    {
+        var resolvedArray = ArrayPool<ResolvedOutput>.Shared.Rent(descriptors.Length);
+        var resolved = resolvedArray.AsSpan(0, descriptors.Length);
+        resolved.Clear();
+
+        Span<int> lockOrder = stackalloc int[descriptors.Length];
+        var acquiredLocks = 0;
+        try
+        {
+            uint bufferLength = 0;
+            for (var i = 0; i < descriptors.Length; i++)
+            {
+                var descriptor = descriptors[i];
+                for (var previous = 0; previous < i; previous++)
+                {
+                    if (resolved[previous].Handle == descriptor.Handle)
+                    {
+                        return AudioOutErrorInvalidPort;
+                    }
+                }
+
+                if (!Ports.TryGetValue(descriptor.Handle, out var port))
+                {
+                    return _shutdown ? 0 : AudioOutErrorInvalidPort;
+                }
+
+                if (i == 0)
+                {
+                    bufferLength = port.BufferLength;
+                }
+                else if (port.BufferLength != bufferLength)
+                {
+                    return AudioOutErrorInvalidSize;
+                }
+
+                resolved[i].Handle = descriptor.Handle;
+                resolved[i].SourceAddress = descriptor.SourceAddress;
+                resolved[i].Port = port;
+                lockOrder[i] = i;
+            }
+
+            // Every batch takes port locks in handle order. Two guest threads can
+            // submit overlapping batches in a different descriptor order without
+            // deadlocking each other.
+            for (var i = 1; i < lockOrder.Length; i++)
+            {
+                var index = lockOrder[i];
+                var position = i;
+                while (position > 0 &&
+                       resolved[lockOrder[position - 1]].Handle > resolved[index].Handle)
+                {
+                    lockOrder[position] = lockOrder[position - 1];
+                    position--;
+                }
+
+                lockOrder[position] = index;
+            }
+
+            for (; acquiredLocks < lockOrder.Length; acquiredLocks++)
+            {
+                Monitor.Enter(resolved[lockOrder[acquiredLocks]].Port.SubmissionGate);
+            }
+
+            // AudioOutClose removes the handle before waiting for SubmissionGate.
+            // Recheck after acquiring all gates so a close racing this batch cannot
+            // turn a validated submission into a write to a disposed backend.
+            for (var i = 0; i < resolved.Length; i++)
+            {
+                if (!Ports.TryGetValue(resolved[i].Handle, out var current) ||
+                    !ReferenceEquals(current, resolved[i].Port))
+                {
+                    return _shutdown ? 0 : AudioOutErrorInvalidPort;
+                }
+            }
+
+            // Stage every guest buffer before the first host submission. A bad
+            // pointer in a later descriptor therefore cannot partially enqueue the
+            // earlier ports.
+            for (var i = 0; i < resolved.Length; i++)
+            {
+                ref var output = ref resolved[i];
+                if (output.SourceAddress == 0)
+                {
+                    continue;
+                }
+
+                var sourceBuffer = ArrayPool<byte>.Shared.Rent(output.Port.BufferByteLength);
+                try
+                {
+                    var source = sourceBuffer.AsSpan(0, output.Port.BufferByteLength);
+                    if (!ctx.Memory.TryRead(output.SourceAddress, source))
+                    {
+                        return AudioOutErrorInvalidPointer;
+                    }
+
+                    TraceOutput(output.Handle, output.Port, source);
+
+                    output.HostBufferLength = checked(
+                        (int)output.Port.BufferLength * AudioPcmConversion.OutputFrameSize);
+                    output.HostBuffer = ArrayPool<byte>.Shared.Rent(output.HostBufferLength);
+                    AudioPcmConversion.ConvertToStereoPcm16(
+                        source,
+                        output.HostBuffer.AsSpan(0, output.HostBufferLength),
+                        checked((int)output.Port.BufferLength),
+                        output.Port.Channels,
+                        output.Port.BytesPerSample,
+                        output.Port.IsFloat,
+                        output.Port.Volume);
+                }
+                finally
+                {
+                    ArrayPool<byte>.Shared.Return(sourceBuffer);
+                }
+            }
+
+            PortState? pacingPort = null;
+            for (var i = 0; i < resolved.Length; i++)
+            {
+                ref var output = ref resolved[i];
+                if (output.HostBuffer is null ||
+                    output.Port.Backend is null ||
+                    !output.Port.Backend.Submit(
+                        output.HostBuffer.AsSpan(0, output.HostBufferLength)))
+                {
+                    if (pacingPort is null ||
+                        HasLongerBufferDuration(output.Port, pacingPort))
+                    {
+                        pacingPort = output.Port;
+                    }
+                }
+            }
+
+            // A batch is one guest scheduling point. When one or more ports have
+            // no usable backend, pace once using the longest affected buffer rather
+            // than sleeping once per port.
+            pacingPort?.PaceSilence();
+            return checked((int)resolved[0].Port.BufferLength);
+        }
+        finally
+        {
+            for (var i = acquiredLocks - 1; i >= 0; i--)
+            {
+                Monitor.Exit(resolved[lockOrder[i]].Port.SubmissionGate);
+            }
+
+            for (var i = 0; i < resolved.Length; i++)
+            {
+                if (resolved[i].HostBuffer is { } hostBuffer)
+                {
+                    ArrayPool<byte>.Shared.Return(hostBuffer);
+                }
+            }
+
+            ArrayPool<ResolvedOutput>.Shared.Return(resolvedArray, clearArray: true);
+        }
+    }
+
+    private static bool HasLongerBufferDuration(PortState candidate, PortState current) =>
+        (ulong)candidate.BufferLength * current.Frequency >
+        (ulong)current.BufferLength * candidate.Frequency;
+
+    private static void TraceOutput(int handle, PortState port, ReadOnlySpan<byte> source)
+    {
+        if (!_traceOutput)
+        {
+            return;
+        }
+
+        var n = Interlocked.Increment(ref _outputCount);
+        if (n <= 8 || n % 200 == 0)
+        {
+            var peak = PeakAmplitude(source, port.IsFloat, port.BytesPerSample);
+            Console.Error.WriteLine(
+                $"[LOADER][TRACE] audioout.output#{n} handle={handle} bytes={source.Length} ch={port.Channels} float={port.IsFloat} vol={port.Volume:F2} peak={peak:F4} backend={(port.Backend is null ? "none" : "coreaudio")}");
+        }
+    }
+
     [SysAbiExport(
         Nid = "b+uAV89IlxE",
         ExportName = "sceAudioOutSetVolume",
@@ -362,6 +608,25 @@ public static class AudioOutExports
         }
     }
 
+    internal static void SetStreamFactoryForTests(Func<uint, IHostAudioStream?>? streamFactory) =>
+        Volatile.Write(ref _streamFactoryForTests, streamFactory);
+
+    internal static void ResetForTests()
+    {
+        foreach (var handle in Ports.Keys)
+        {
+            if (Ports.TryRemove(handle, out var port))
+            {
+                port.Dispose();
+            }
+        }
+
+        _nextPortHandle = 0;
+        _outputCount = 0;
+        Volatile.Write(ref _shutdown, false);
+        Volatile.Write(ref _streamFactoryForTests, null);
+    }
+
     private static bool _shutdown;
 
     private static bool TryGetFormat(

src/SharpEmu.Libs/Gpu/GuestGpuTypes.cs

@@ -1,6 +1,8 @@
 // Copyright (C) 2026 SharpEmu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+using SharpEmu.Libs.Agc;
+
 namespace SharpEmu.Libs.Gpu;
 
 // The types that cross the guest-GPU backend seam. Every field is either a neutral
@@ -10,7 +12,8 @@ namespace SharpEmu.Libs.Gpu;
 // translation for its API.
 
 /// <summary>A guest texture referenced by a draw or dispatch. Format/NumberType/
-/// TileMode/DstSelect are raw guest descriptor codes.</summary>
+/// TileMode/DstSelect/Type are raw guest descriptor codes. Depth is the
+/// normalized volume depth (one for non-3D resources).</summary>
 internal sealed record GuestDrawTexture(
     ulong Address,
     uint Width,
@@ -32,7 +35,15 @@ internal sealed record GuestDrawTexture(
     // from; -1 when the range is untracked or the pixels were not read here.
     long WriteGeneration = -1,
     bool ArrayedView = false,
-    uint ArrayLayers = 1);
+    uint ArrayLayers = 1,
+    uint Type = 9,
+    uint Depth = 1,
+    // GPU-detile opt-in (SHARPEMU_GPU_DETILE): when Detile is non-null the AGC
+    // layer skipped the CPU deswizzle and shipped the raw TILED bytes here in
+    // TiledSource; the Vulkan backend detiles them on the GPU. RgbaPixels is
+    // empty in that case. Both are neutral (no host graphics-API values).
+    byte[]? TiledSource = null,
+    DetileParams? Detile = null);
 
 /// <summary>Raw guest sampler descriptor dwords, copied verbatim from guest memory.</summary>
 internal readonly record struct GuestSampler(
@@ -55,7 +66,9 @@ internal readonly record struct TextureContentIdentity(
     uint Pitch,
     GuestSampler Sampler,
     bool Arrayed = false,
-    uint ArrayLayers = 1);
+    uint ArrayLayers = 1,
+    uint Type = 9,
+    uint Depth = 1);
 
 internal sealed record GuestMemoryBuffer(
     ulong BaseAddress,

src/SharpEmu.Libs/Gpu/Metal/MetalDetilePass.cs

@@ -0,0 +1,313 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Numerics;
+using SharpEmu.Libs.Agc;
+using SharpEmu.ShaderCompiler.Metal;
+
+namespace SharpEmu.Libs.Gpu.Metal;
+
+/// <summary>
+/// Metal twin of <c>VulkanDetilePass</c>: runs the ExactXor detile equation from
+/// <see cref="GnmTiling.GetDetileParams"/> as a Metal compute kernel
+/// (<see cref="MslFixedShaders.CreateDetileCompute"/>), writing a linear buffer
+/// and blitting it into the sampled texture.
+///
+/// <see cref="RecordDetile"/> records the compute dispatch + blit onto a caller's
+/// command buffer and returns its transient buffers for the caller to release
+/// once that command buffer completes — the async, non-blocking shape (Metal
+/// hazard-tracks the compute-write → blit-read → sample dependency automatically,
+/// so no manual barriers are needed).
+///
+/// Only ExactXor 4-bytes/element surfaces are handled. NOTE: authored on Windows;
+/// the MSL and every Metal call here are <b>Mac-untested</b> — mirrors the
+/// verified Vulkan logic and the existing Metal message-send conventions, but
+/// must be validated on a real Metal device.
+/// </summary>
+internal sealed unsafe class MetalDetilePass : IDisposable
+{
+    private const uint LocalSize = 8;
+    private const int PushConstantUints = 11;
+
+    private readonly nint _device;
+    private nint _pipelineState;
+    private bool _initialized;
+    private bool _disposed;
+
+    public MetalDetilePass(nint device)
+    {
+        _device = device;
+    }
+
+    public static bool Supports(in DetileParams parameters) =>
+        (parameters.Equation == DetileEquation.ExactXor ||
+         parameters.Equation == DetileEquation.BlockTable) &&
+        parameters.BytesPerElement is 4 or 8 or 16;
+
+    /// <summary>
+    /// Records the deswizzle of <paramref name="tiled"/> into
+    /// <paramref name="texture"/> (<paramref name="texelWidth"/> x
+    /// <paramref name="texelHeight"/> texels x <paramref name="layers"/> slices)
+    /// onto <paramref name="commandBuffer"/>. The kernel iterates the element grid
+    /// from <paramref name="parameters"/> (for block-compressed formats a 4x4 block
+    /// is one element). Does not commit; the caller releases
+    /// <paramref name="transientBuffers"/> when the command buffer completes.
+    /// Returns false (empty transients) when unsupported or the pipeline could not
+    /// be built.
+    /// </summary>
+    public bool RecordDetile(
+        nint commandBuffer,
+        nint texture,
+        uint texelWidth,
+        uint texelHeight,
+        uint layers,
+        ReadOnlySpan<byte> tiled,
+        in DetileParams parameters,
+        out nint[] transientBuffers)
+    {
+        transientBuffers = [];
+        var bytesPerElement = (uint)parameters.BytesPerElement;
+        if (_disposed || commandBuffer == 0 || texture == 0 ||
+            !Supports(parameters) || texelWidth == 0 || texelHeight == 0 || layers == 0 || tiled.IsEmpty ||
+            tiled.Length % (int)(layers * bytesPerElement) != 0 ||
+            !EnsurePipeline())
+        {
+            return false;
+        }
+
+        var elementsWide = (uint)parameters.ElementsWide;
+        var elementsHigh = (uint)parameters.ElementsHigh;
+        var uintsPerElement = bytesPerElement / sizeof(uint);
+
+        // Array slices are packed contiguously in the tiled buffer; each slice's
+        // element stride is the whole buffer split evenly by layer.
+        var srcSliceElements = (uint)((ulong)tiled.Length / bytesPerElement / layers);
+
+        // Binding 1 carries the within-block offset table. ExactXor: element-shifted
+        // X/Y byte terms. BlockTable: GetDetileParams' block table (already element
+        // offsets) in binding 1, a placeholder in binding 2. The two equations index
+        // different-sized buffers, so the kernel branches and reads only one.
+        uint[] xTerm;
+        uint[] yTerm;
+        uint equationValue;
+        if (parameters.Equation == DetileEquation.BlockTable)
+        {
+            xTerm = new uint[parameters.BlockTable.Length];
+            for (var index = 0; index < xTerm.Length; index++)
+            {
+                xTerm[index] = (uint)parameters.BlockTable[index];
+            }
+
+            yTerm = [0];
+            equationValue = 1;
+        }
+        else
+        {
+            var shift = BitOperations.TrailingZeroCount((uint)parameters.BytesPerElement);
+            xTerm = ToElementTerms(parameters.XByteTerm, shift);
+            yTerm = ToElementTerms(parameters.YByteTerm, shift);
+            equationValue = 0;
+        }
+
+        var newBufferWithBytes = MetalNative.Selector("newBufferWithBytes:length:options:");
+        var newBufferWithLength = MetalNative.Selector("newBufferWithLength:options:");
+
+        nint tiledBuffer;
+        nint xBuffer;
+        nint yBuffer;
+        fixed (byte* tiledPointer = tiled)
+        {
+            tiledBuffer = MetalNative.SendBuffer(
+                _device, newBufferWithBytes, (nint)tiledPointer, (nuint)tiled.Length, 0);
+        }
+
+        fixed (uint* xPointer = xTerm)
+        {
+            xBuffer = MetalNative.SendBuffer(
+                _device, newBufferWithBytes, (nint)xPointer, (nuint)xTerm.Length * sizeof(uint), 0);
+        }
+
+        fixed (uint* yPointer = yTerm)
+        {
+            yBuffer = MetalNative.SendBuffer(
+                _device, newBufferWithBytes, (nint)yPointer, (nuint)yTerm.Length * sizeof(uint), 0);
+        }
+
+        var outputBytes = (nuint)elementsWide * elementsHigh * bytesPerElement * layers;
+        var outputBuffer = MetalNative.SendNewBuffer(_device, newBufferWithLength, outputBytes, 0);
+
+        Span<uint> push =
+        [
+            elementsWide,
+            elementsHigh,
+            (uint)parameters.BlockWidth,
+            (uint)parameters.BlockHeight,
+            (uint)parameters.BlockElements,
+            (uint)parameters.BlocksPerRow,
+            (uint)parameters.XMask,
+            (uint)parameters.YMask,
+            srcSliceElements,
+            equationValue,
+            uintsPerElement,
+        ];
+        nint paramsBuffer;
+        fixed (uint* pushPointer = push)
+        {
+            paramsBuffer = MetalNative.SendBuffer(
+                _device, newBufferWithBytes, (nint)pushPointer, (nuint)PushConstantUints * sizeof(uint), 0);
+        }
+
+        if (tiledBuffer == 0 || xBuffer == 0 || yBuffer == 0 || outputBuffer == 0 || paramsBuffer == 0)
+        {
+            ReleaseAll(tiledBuffer, xBuffer, yBuffer, outputBuffer, paramsBuffer);
+            return false;
+        }
+
+        // Compute encoder: one thread per texel.
+        var setBuffer = MetalNative.Selector("setBuffer:offset:atIndex:");
+        var encoder = MetalNative.Send(commandBuffer, MetalNative.Selector("computeCommandEncoder"));
+        MetalNative.Send(encoder, MetalNative.Selector("setComputePipelineState:"), _pipelineState);
+        MetalNative.SendSetBuffer(encoder, setBuffer, tiledBuffer, 0, 0);
+        MetalNative.SendSetBuffer(encoder, setBuffer, xBuffer, 0, 1);
+        MetalNative.SendSetBuffer(encoder, setBuffer, yBuffer, 0, 2);
+        MetalNative.SendSetBuffer(encoder, setBuffer, outputBuffer, 0, 3);
+        MetalNative.SendSetBuffer(encoder, setBuffer, paramsBuffer, 0, 4);
+
+        // X is widened by uintsPerElement (each thread copies one word); one
+        // grid-Z layer per array slice.
+        var threadgroups = new MtlSize
+        {
+            Width = (nuint)((elementsWide * uintsPerElement + LocalSize - 1) / LocalSize),
+            Height = (nuint)((elementsHigh + LocalSize - 1) / LocalSize),
+            Depth = layers,
+        };
+        var threadsPerThreadgroup = new MtlSize { Width = LocalSize, Height = LocalSize, Depth = 1 };
+        MetalNative.SendDispatch(
+            encoder,
+            MetalNative.Selector("dispatchThreadgroups:threadsPerThreadgroup:"),
+            threadgroups,
+            threadsPerThreadgroup);
+        MetalNative.SendVoid(encoder, MetalNative.Selector("endEncoding"));
+
+        // Blit the layer-major linear output buffer into the sampled texture, one
+        // slice per array layer (Metal copyFromBuffer targets a single slice). The
+        // buffer is element/block-packed (row stride = elementsWide*bpp); the copy
+        // region is in texels. Metal tracks the compute-write -> blit-read hazard.
+        var blit = MetalNative.Send(commandBuffer, MetalNative.Selector("blitCommandEncoder"));
+        var copySelector = MetalNative.Selector(
+            "copyFromBuffer:sourceOffset:sourceBytesPerRow:sourceBytesPerImage:sourceSize:" +
+            "toTexture:destinationSlice:destinationLevel:destinationOrigin:");
+        var sliceBytes = (nuint)elementsWide * elementsHigh * bytesPerElement;
+        var rowBytes = (nuint)elementsWide * bytesPerElement;
+        for (uint layer = 0; layer < layers; layer++)
+        {
+            MetalNative.SendCopyBufferToTexture(
+                blit,
+                copySelector,
+                outputBuffer,
+                (nuint)layer * sliceBytes,
+                rowBytes,
+                sliceBytes,
+                new MtlSize { Width = texelWidth, Height = texelHeight, Depth = 1 },
+                texture,
+                layer,
+                0,
+                new MtlOrigin { X = 0, Y = 0, Z = 0 });
+        }
+
+        MetalNative.SendVoid(blit, MetalNative.Selector("endEncoding"));
+
+        transientBuffers = [tiledBuffer, xBuffer, yBuffer, outputBuffer, paramsBuffer];
+        return true;
+    }
+
+    private bool EnsurePipeline()
+    {
+        if (_initialized)
+        {
+            return _pipelineState != 0;
+        }
+
+        _initialized = true;
+
+        var options = MetalNative.Send(
+            MetalNative.Send(MetalNative.Class("MTLCompileOptions"), MetalNative.Selector("alloc")),
+            MetalNative.Selector("init"));
+        MetalNative.SendVoidBool(options, MetalNative.Selector("setFastMathEnabled:"), false);
+
+        nint libraryError = 0;
+        var library = MetalNative.Send(
+            _device,
+            MetalNative.Selector("newLibraryWithSource:options:error:"),
+            MetalNative.NsString(MslFixedShaders.CreateDetileCompute()),
+            options,
+            ref libraryError);
+        if (library == 0)
+        {
+            Console.Error.WriteLine(
+                $"[GPU-DETILE] Metal detile library compile failed: {MetalNative.DescribeError(libraryError)}");
+            return false;
+        }
+
+        var function = MetalNative.Send(
+            library, MetalNative.Selector("newFunctionWithName:"), MetalNative.NsString("detile_cs"));
+        if (function == 0)
+        {
+            Console.Error.WriteLine("[GPU-DETILE] Metal detile function 'detile_cs' not found.");
+            return false;
+        }
+
+        nint pipelineError = 0;
+        _pipelineState = MetalNative.Send(
+            _device,
+            MetalNative.Selector("newComputePipelineStateWithFunction:error:"),
+            function,
+            ref pipelineError);
+        if (_pipelineState == 0)
+        {
+            Console.Error.WriteLine(
+                $"[GPU-DETILE] Metal detile pipeline failed: {MetalNative.DescribeError(pipelineError)}");
+            return false;
+        }
+
+        return true;
+    }
+
+    private static uint[] ToElementTerms(int[] byteTerms, int shift)
+    {
+        var terms = new uint[byteTerms.Length];
+        for (var index = 0; index < byteTerms.Length; index++)
+        {
+            terms[index] = (uint)byteTerms[index] >> shift;
+        }
+
+        return terms;
+    }
+
+    private static void ReleaseAll(params nint[] objects)
+    {
+        var release = MetalNative.Selector("release");
+        foreach (var handle in objects)
+        {
+            if (handle != 0)
+            {
+                MetalNative.SendVoid(handle, release);
+            }
+        }
+    }
+
+    public void Dispose()
+    {
+        if (_disposed)
+        {
+            return;
+        }
+
+        _disposed = true;
+        if (_pipelineState != 0)
+        {
+            MetalNative.SendVoid(_pipelineState, MetalNative.Selector("release"));
+            _pipelineState = 0;
+        }
+    }
+}

src/SharpEmu.Libs/Gpu/Metal/MetalVideoPresenter.Draws.cs

@@ -1,6 +1,7 @@
 // Copyright (C) 2026 SharpEmu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+using SharpEmu.Libs.Agc;
 using SharpEmu.ShaderCompiler;
 using SharpEmu.ShaderCompiler.Metal;
 
@@ -1583,6 +1584,45 @@ internal static partial class MetalVideoPresenter
             return cached;
         }
 
+        // Default-on GPU detile packages the tiled source + resolved DetileParams
+        // with empty RgbaPixels so a backend can deswizzle on the GPU. The Metal
+        // GPU compute pass (MetalDetilePass / detile_compute.msl) is the intended
+        // equivalent of VulkanDetilePass, but it is Mac-untested, so the active
+        // Metal path CPU-detiles here via GnmTiling.DetileWithParams — the exact
+        // same DetileParams addressing the kernel runs — restoring the linear-upload
+        // behavior Metal had before GPU detile existed, with no regression.
+        if (texture.RgbaPixels.Length == 0 &&
+            texture.TiledSource is { } tiledSource &&
+            texture.Detile is { } detileParameters)
+        {
+            // The tiled source packs the array slices contiguously (one per layer);
+            // detile each into its layer-major linear region so the reconstructed
+            // pixels match what the CPU array-upload path produced pre-GPU-detile.
+            // A plain 2D texture is just one layer.
+            var layers = Math.Max((int)texture.ArrayLayers, 1);
+            var sliceLinearBytes =
+                detileParameters.ElementsWide * detileParameters.ElementsHigh * detileParameters.BytesPerElement;
+            var sliceTiledBytes = tiledSource.Length / layers;
+            var linear = new byte[sliceLinearBytes * layers];
+            var detiledAll = true;
+            for (var layer = 0; layer < layers; layer++)
+            {
+                if (!GnmTiling.DetileWithParams(
+                        detileParameters,
+                        tiledSource.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
+                        linear.AsSpan(layer * sliceLinearBytes, sliceLinearBytes)))
+                {
+                    detiledAll = false;
+                    break;
+                }
+            }
+
+            if (detiledAll)
+            {
+                texture = texture with { RgbaPixels = linear };
+            }
+        }
+
         // AGC ships the raw (detiled) source texels; create the texture in the
         // guest's native format — Mac-family GPUs sample BC blocks directly —
         // and size expectations with the same block-aware math AGC used.

src/SharpEmu.Libs/Json/JsonExports.cs

@@ -80,6 +80,18 @@ public static class JsonExports
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
+    [SysAbiExport(
+        Nid = "PR5k1penBLM",
+        ExportName = "_ZN3sce4Json11Initializer9terminateEv",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceJson")]
+    public static int InitializerTerminate(CpuContext ctx)
+    {
+        TraceJson("Initializer.terminate", ctx[CpuRegister.Rdi], 0);
+        ctx[CpuRegister.Rax] = 0;
+        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+    }
+
     [SysAbiExport(
         Nid = "Cxwy7wHq4J0",
         ExportName = "_ZN3sce4Json11Initializer10initializeEPKNS0_13InitParameterE",

src/SharpEmu.Libs/Kernel/KernelPthreadCompatExports.cs

@@ -153,6 +153,64 @@ public static class KernelPthreadCompatExports
     static KernelPthreadCompatExports()
     {
         RunSynchronizationSelfChecks();
+        GuestThreadExecution.GuestThreadAbandoned += AbandonMutexesOwnedByThread;
+    }
+
+    /// <summary>
+    /// Force-release mutexes still owned by a guest thread that is being torn
+    /// down without a clean unlock (TBB worker_abort, abrupt exit). Otherwise
+    /// waiters can spin forever and block splash→first GPU submit.
+    /// </summary>
+    public static int AbandonMutexesOwnedByThread(ulong threadId, string reason)
+    {
+        if (threadId == 0)
+        {
+            return 0;
+        }
+
+        var released = 0;
+        var wakeKeys = new List<string>();
+        foreach (var pair in _mutexStates)
+        {
+            var state = pair.Value;
+            string? wakeKey = null;
+            lock (state)
+            {
+                if (state.OwnerThreadId != threadId || state.RecursionCount <= 0)
+                {
+                    continue;
+                }
+
+                state.OwnerThreadId = 0;
+                state.RecursionCount = 0;
+                wakeKey = state.Waiters.First?.Value.Cooperative == true
+                    ? state.Waiters.First.Value.WakeKey
+                    : null;
+                Monitor.PulseAll(state);
+                released++;
+                Console.Error.WriteLine(
+                    $"[LOADER][WARN] pthread_mutex_abandon mutex=0x{pair.Key:X16} " +
+                    $"owner={KernelPthreadState.DescribeThreadHandle(threadId)} " +
+                    $"reason={reason} waiters={state.Waiters.Count}");
+            }
+
+            if (wakeKey is not null)
+            {
+                wakeKeys.Add(wakeKey);
+            }
+        }
+
+        foreach (var wakeKey in wakeKeys)
+        {
+            _ = GuestThreadExecution.Scheduler?.WakeBlockedThreads(wakeKey, 1);
+        }
+
+        if (released > 0)
+        {
+            Console.Error.Flush();
+        }
+
+        return released;
     }
 
     [SysAbiExport(

src/SharpEmu.Libs/Kernel/KernelPthreadState.cs

@@ -27,8 +27,12 @@ internal static class KernelPthreadState
     internal static ulong GetCurrentThreadHandle()
     {
         var guestThreadHandle = GuestThreadExecution.CurrentGuestThreadHandle;
-        if (guestThreadHandle != 0 && TryGetThreadIdentity(guestThreadHandle, out _))
+        // Prefer the bound guest handle even when it is not yet in Threads.
+        // Falling through to a synthetic ThreadStatic handle while a guest
+        // thread is bound causes mutex owner mismatches (unlock PERM → hang).
+        if (guestThreadHandle != 0)
         {
+            EnsureGuestThreadIdentity(guestThreadHandle);
             return guestThreadHandle;
         }
 
@@ -39,15 +43,27 @@ internal static class KernelPthreadState
     internal static ulong GetCurrentThreadUniqueId()
     {
         var guestThreadHandle = GuestThreadExecution.CurrentGuestThreadHandle;
-        if (guestThreadHandle != 0 && TryGetThreadIdentity(guestThreadHandle, out var identity))
+        if (guestThreadHandle != 0)
         {
-            return identity.UniqueId;
+            return EnsureGuestThreadIdentity(guestThreadHandle).UniqueId;
         }
 
         EnsureCurrentThreadRegistered();
         return _currentThreadUniqueId;
     }
 
+    internal static string DescribeThreadHandle(ulong threadHandle)
+    {
+        if (threadHandle == 0)
+        {
+            return "none";
+        }
+
+        return TryGetThreadIdentity(threadHandle, out var identity)
+            ? $"0x{threadHandle:X16}('{identity.Name}')"
+            : $"0x{threadHandle:X16}";
+    }
+
     internal static ulong CreateThreadHandle(string name)
     {
         var uniqueId = unchecked((ulong)Interlocked.Increment(ref _nextUniqueThreadId));
@@ -59,6 +75,18 @@ internal static class KernelPthreadState
         return Threads.TryGetValue(threadHandle, out identity);
     }
 
+    private static ThreadIdentity EnsureGuestThreadIdentity(ulong guestThreadHandle)
+    {
+        if (Threads.TryGetValue(guestThreadHandle, out var existing))
+        {
+            return existing;
+        }
+
+        var uniqueId = unchecked((ulong)Interlocked.Increment(ref _nextUniqueThreadId));
+        var identity = new ThreadIdentity(uniqueId, $"Guest-0x{guestThreadHandle:X}");
+        return Threads.GetOrAdd(guestThreadHandle, identity);
+    }
+
     private static void EnsureCurrentThreadRegistered()
     {
         if (_currentThreadHandle != 0)

src/SharpEmu.Libs/Network/NetExports.cs

@@ -755,6 +755,52 @@ public static class NetExports
         return true;
     }
 
+    [SysAbiExport(
+        Nid = "8Kcp5d-q1Uo",
+        ExportName = "sceNetInetPton",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceNet")]
+    public static int NetInetPton(CpuContext ctx)
+    {
+        var addressFamily = unchecked((int)ctx[CpuRegister.Rdi]);
+        var sourceAddress = ctx[CpuRegister.Rsi];
+        var destinationAddress = ctx[CpuRegister.Rdx];
+        if (sourceAddress == 0 || destinationAddress == 0)
+        {
+            return SetNetError(ctx, NetErrorInvalidArgument, NetErrnoInvalidArgument);
+        }
+
+        if (!TryReadUtf8Z(ctx, sourceAddress, MaxNameLength, out var source))
+        {
+            return SetNetError(ctx, NetErrorInvalidArgument, NetErrnoInvalidArgument);
+        }
+
+        var family = addressFamily switch
+        {
+            2 => AddressFamily.InterNetwork,      // AF_INET
+            28 => AddressFamily.InterNetworkV6,   // AF_INET6
+            _ => AddressFamily.Unknown,
+        };
+        if (family == AddressFamily.Unknown ||
+            !IPAddress.TryParse(source, out var parsed) ||
+            parsed.AddressFamily != family)
+        {
+            // Match BSD inet_pton: return 0 for a parseable-family miss.
+            ctx[CpuRegister.Rax] = 0;
+            return 0;
+        }
+
+        var bytes = parsed.GetAddressBytes();
+        if (!ctx.Memory.TryWrite(destinationAddress, bytes))
+        {
+            return SetNetError(ctx, NetErrorInvalidArgument, NetErrnoInvalidArgument);
+        }
+
+        TraceNet("inet_pton", addressFamily, sourceAddress, destinationAddress, (ulong)bytes.Length);
+        ctx[CpuRegister.Rax] = 1;
+        return 1;
+    }
+
     private static void TraceNet(string operation, int id, ulong arg0, ulong arg1, ulong arg2)
     {
         if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_NET"), "1", StringComparison.Ordinal))

src/SharpEmu.Libs/Psml/PsmlExports.cs

@@ -0,0 +1,365 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.HLE;
+
+namespace SharpEmu.Libs.Psml;
+
+public static class PsmlExports
+{
+    // Empirically for Astro Bot (PPSA21567):
+    //   [0] must be 0x80 (sizeThis / r9); any other value в†’ Allocate length=0
+    //   AllocateMainDirectMemory(length=[0]*[16], alignment=[8], type=0xC)
+    // So put desired byte size in [8] (becomes alignment) and page size in [16].
+    private const ulong RequirementStructSize = 0x80;
+    private const ulong SharedResourcesPageSize = 0x10000;
+    private const ulong SharedResourcesBufferSizeBytes = 0x2000000;
+    private const ulong SharedResourcesContextSizeBytes = 0x100000;
+    private const ulong ContextBufferSizeBytes = 0x800000;
+    private const ulong ContextBufferAuxSizeBytes = 0x100000;
+
+    private static int _mfsrInitialized;
+    private static readonly Lock SharedResourcesGate = new();
+    private static readonly Dictionary<ulong, SharedResourcesState> SharedResourcesByDescriptor = new();
+    private static readonly Lock ContextGate = new();
+    private static readonly Dictionary<ulong, ContextState> ContextsByAddress = new();
+
+    [SysAbiExport(
+        Nid = "3WVD91e12ZQ",
+        ExportName = "scePsmlMfsrInit",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrInit(CpuContext ctx)
+    {
+        var arg0 = ctx[CpuRegister.Rdi];
+        var arg1 = ctx[CpuRegister.Rsi];
+        var arg2 = ctx[CpuRegister.Rdx];
+        Interlocked.Exchange(ref _mfsrInitialized, 1);
+        TracePsml($"mfsr_init arg0=0x{arg0:X} arg1=0x{arg1:X} arg2=0x{arg2:X}");
+        return ctx.SetReturn(0);
+    }
+
+    [SysAbiExport(
+        Nid = "+2KpvixvL6E",
+        ExportName = "scePsmlMfsrGetSharedResourcesInitRequirement",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetSharedResourcesInitRequirement(CpuContext ctx)
+    {
+        var bufferRequirementAddress = ctx[CpuRegister.Rdi];
+        var contextRequirementAddress = ctx[CpuRegister.Rsi];
+        var flags = ctx[CpuRegister.Rdx];
+        var configAddress = ctx[CpuRegister.Rcx];
+        if (bufferRequirementAddress == 0 || contextRequirementAddress == 0)
+        {
+            return ctx.SetReturn(OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT);
+        }
+
+        if (!WriteMemoryRequirement(ctx, bufferRequirementAddress, SharedResourcesBufferSizeBytes) ||
+            !WriteMemoryRequirement(ctx, contextRequirementAddress, SharedResourcesContextSizeBytes))
+        {
+            return ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        TracePsml(
+            $"mfsr_get_shared_resources_init_requirement buf=0x{bufferRequirementAddress:X16} " +
+            $"ctx=0x{contextRequirementAddress:X16} flags=0x{flags:X} config=0x{configAddress:X16} " +
+            $"buf_size=0x{SharedResourcesBufferSizeBytes:X} ctx_size=0x{SharedResourcesContextSizeBytes:X}");
+        return ctx.SetReturn(0);
+    }
+
+    [SysAbiExport(
+        Nid = "eWoKNeB6V-k",
+        ExportName = "scePsmlMfsrCreateSharedResources",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrCreateSharedResources(CpuContext ctx)
+    {
+        var descriptorAddress = ctx[CpuRegister.Rdi];
+        var contextRequirementAddress = ctx[CpuRegister.Rsi];
+        var directMemoryAddress = ctx[CpuRegister.Rdx];
+        if (descriptorAddress == 0 || contextRequirementAddress == 0 || directMemoryAddress == 0)
+        {
+            return ctx.SetReturn(OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT);
+        }
+
+        var contextSizeBytes = ReadRequirementSize(ctx, contextRequirementAddress, SharedResourcesContextSizeBytes);
+        var state = new SharedResourcesState(
+            DescriptorAddress: descriptorAddress,
+            DirectMemoryAddress: directMemoryAddress,
+            BufferSizeBytes: SharedResourcesBufferSizeBytes,
+            ContextSizeBytes: contextSizeBytes,
+            PageSizeBytes: SharedResourcesPageSize);
+
+        if (!WriteSharedResourcesDescriptor(ctx, state))
+        {
+            return ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        lock (SharedResourcesGate)
+        {
+            SharedResourcesByDescriptor[descriptorAddress] = state;
+        }
+
+        TracePsml(
+            $"mfsr_create_shared_resources desc=0x{descriptorAddress:X16} req=0x{contextRequirementAddress:X16} " +
+            $"direct=0x{directMemoryAddress:X16} buf_size=0x{state.BufferSizeBytes:X} " +
+            $"ctx_size=0x{state.ContextSizeBytes:X} page=0x{state.PageSizeBytes:X}");
+        return ctx.SetReturn(0);
+    }
+
+    [SysAbiExport(
+        Nid = "ArakEpzsZo0",
+        ExportName = "scePsmlMfsrGetContextBufferRequirement800M3_2",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetContextBufferRequirement800M3_2(CpuContext ctx)
+    {
+        var bufferRequirementAddress = ctx[CpuRegister.Rdi];
+        var contextRequirementAddress = ctx[CpuRegister.Rsi];
+        var directMemoryAddress = ctx[CpuRegister.Rdx];
+        if (bufferRequirementAddress == 0 || contextRequirementAddress == 0 || directMemoryAddress == 0)
+        {
+            return ctx.SetReturn(OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT);
+        }
+
+        if (!WriteMemoryRequirement(ctx, bufferRequirementAddress, ContextBufferSizeBytes) ||
+            !WriteMemoryRequirement(ctx, contextRequirementAddress, ContextBufferAuxSizeBytes))
+        {
+            return ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        TracePsml(
+            $"mfsr_get_context_buffer_requirement_800m3_2 buf=0x{bufferRequirementAddress:X16} " +
+            $"ctx=0x{contextRequirementAddress:X16} direct=0x{directMemoryAddress:X16} " +
+            $"buf_size=0x{ContextBufferSizeBytes:X} aux_size=0x{ContextBufferAuxSizeBytes:X}");
+        return ctx.SetReturn(0);
+    }
+
+    [SysAbiExport(
+        Nid = "gxv3i+MTEzU",
+        ExportName = "scePsmlMfsrCreateContext800M3_2",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrCreateContext800M3_2(CpuContext ctx)
+    {
+        var contextAddress = ctx[CpuRegister.Rdi];
+        var requirementAddress = ctx[CpuRegister.Rsi];
+        var structSize = ctx[CpuRegister.Rdx];
+        var sharedDirectMemory = ctx[CpuRegister.Rcx];
+        var pageSize = ctx[CpuRegister.R8];
+        if (contextAddress == 0 || sharedDirectMemory == 0)
+        {
+            return ctx.SetReturn(OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT);
+        }
+
+        var sharedState = TryFindSharedResources(sharedDirectMemory, contextAddress);
+        var effectiveStructSize = structSize != 0 ? structSize : RequirementStructSize;
+        var effectivePageSize = pageSize != 0 ? pageSize : SharedResourcesPageSize;
+        var sharedDescriptor = sharedState?.DescriptorAddress ?? contextAddress - 0x30;
+        var bufferSize = sharedState?.BufferSizeBytes ?? ContextBufferSizeBytes;
+
+        var state = new ContextState(
+            ContextAddress: contextAddress,
+            SharedResourcesDescriptor: sharedDescriptor,
+            DirectMemoryAddress: sharedDirectMemory,
+            BufferSizeBytes: bufferSize,
+            PageSizeBytes: effectivePageSize,
+            StructSizeBytes: effectiveStructSize);
+
+        if (!WriteContextObject(ctx, state))
+        {
+            return ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        lock (ContextGate)
+        {
+            ContextsByAddress[contextAddress] = state;
+        }
+
+        TracePsml(
+            $"mfsr_create_context_800m3_2 ctx=0x{contextAddress:X16} req=0x{requirementAddress:X16} " +
+            $"struct=0x{effectiveStructSize:X} direct=0x{sharedDirectMemory:X16} " +
+            $"shared_desc=0x{sharedDescriptor:X16} buf_size=0x{bufferSize:X} page=0x{effectivePageSize:X}");
+        return ctx.SetReturn(0);
+    }
+
+    private static bool WriteMemoryRequirement(CpuContext ctx, ulong address, ulong sizeBytes)
+    {
+        return ctx.TryWriteUInt64(address, RequirementStructSize) &&
+               ctx.TryWriteUInt64(address + 0x08, sizeBytes) &&
+               ctx.TryWriteUInt64(address + 0x10, SharedResourcesPageSize);
+    }
+
+    private static ulong ReadRequirementSize(CpuContext ctx, ulong address, ulong fallback)
+    {
+        if (!ctx.TryReadUInt64(address + 0x08, out var sizeBytes) || sizeBytes == 0)
+        {
+            return fallback;
+        }
+
+        // Guest requirement structs use size @8; reject pointer-like garbage.
+        return sizeBytes > 0x1000_0000UL ? fallback : sizeBytes;
+    }
+
+    private static SharedResourcesState? TryFindSharedResources(ulong directMemoryAddress, ulong contextAddress)
+    {
+        lock (SharedResourcesGate)
+        {
+            foreach (var state in SharedResourcesByDescriptor.Values)
+            {
+                if (state.DirectMemoryAddress == directMemoryAddress)
+                {
+                    return state;
+                }
+            }
+
+            var inferredDescriptor = contextAddress >= 0x30 ? contextAddress - 0x30 : 0;
+            if (inferredDescriptor != 0 &&
+                SharedResourcesByDescriptor.TryGetValue(inferredDescriptor, out var byDescriptor))
+            {
+                return byDescriptor;
+            }
+        }
+
+        return null;
+    }
+
+    private static bool WriteContextObject(CpuContext ctx, ContextState state)
+    {
+        return ctx.TryWriteUInt64(state.ContextAddress + 0x00, state.StructSizeBytes) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x08, state.DirectMemoryAddress) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x10, state.SharedResourcesDescriptor) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x18, state.BufferSizeBytes) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x20, state.PageSizeBytes) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x28, state.ContextAddress) &&
+               ctx.TryWriteUInt64(state.ContextAddress + 0x30, state.DirectMemoryAddress);
+    }
+
+    private static bool WriteSharedResourcesDescriptor(CpuContext ctx, SharedResourcesState state)
+    {
+        // Stamp a compact self-describing blob so follow-up PSML calls can treat the
+        // descriptor as initialized guest memory instead of an all-zero placeholder.
+        return ctx.TryWriteUInt64(state.DescriptorAddress + 0x00, RequirementStructSize) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x08, state.DirectMemoryAddress) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x10, state.DirectMemoryAddress) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x18, state.BufferSizeBytes) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x20, state.ContextSizeBytes) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x28, state.PageSizeBytes) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x30, state.DirectMemoryAddress) &&
+               ctx.TryWriteUInt64(state.DescriptorAddress + 0x38, state.BufferSizeBytes + state.ContextSizeBytes);
+    }
+
+
+    // Astro logo path: SizeInDwords then GetDispatchMfsrPacket900 (RUNLFro+qok).
+    // Unresolved 900 returns non-zero and trips GfxRenderStagePSSR.cpp:266.
+    private const int SoftPacketSizeInDwords = 0x80;
+    private const ulong SoftPacketSizeBytes = SoftPacketSizeInDwords * 4UL;
+
+    [SysAbiExport(
+        Nid = "AHalTX9wFZY",
+        ExportName = "scePsmlMfsrGetDispatchMfsrPacketSizeInDwords",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetDispatchMfsrPacketSizeInDwords(CpuContext ctx)
+    {
+        // Logo/PSSR path: guest asserts ret == 0, then calls GetDispatchMfsrPacket900
+        // with rdi=SoftPacketSizeInDwords (observed 0x80). Returning the size in rax
+        // tripped :266 and skipped the 900 call (tDISP-s6). SCE_OK keeps the chain.
+        var arg0 = ctx[CpuRegister.Rdi];
+        TracePsml(
+            $"mfsr_get_dispatch_packet_size_dwords arg0=0x{arg0:X} " +
+            $"size=0x{SoftPacketSizeInDwords:X} ret=0");
+        return ctx.SetReturn(0);
+    }
+
+    [SysAbiExport(
+        Nid = "RUNLFro+qok",
+        ExportName = "scePsmlMfsrGetDispatchMfsrPacket900",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetDispatchMfsrPacket900(CpuContext ctx) =>
+        SoftGetDispatchMfsrPacket(ctx, "900");
+
+    [SysAbiExport(
+        Nid = "s2psNHUIdjk",
+        ExportName = "scePsmlMfsrGetDispatchMfsrPacket1000",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetDispatchMfsrPacket1000(CpuContext ctx) =>
+        SoftGetDispatchMfsrPacket(ctx, "1000");
+
+    [SysAbiExport(
+        Nid = "94iBp3KvIuI",
+        ExportName = "scePsmlMfsrGetDispatchMfsrPacket1100",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libScePsml")]
+    public static int PsmlMfsrGetDispatchMfsrPacket1100(CpuContext ctx) =>
+        SoftGetDispatchMfsrPacket(ctx, "1100");
+
+    private static int SoftGetDispatchMfsrPacket(CpuContext ctx, string variant)
+    {
+        var arg0 = ctx[CpuRegister.Rdi];
+        var arg1 = ctx[CpuRegister.Rsi];
+        var arg2 = ctx[CpuRegister.Rdx];
+        var arg3 = ctx[CpuRegister.Rcx];
+        // Logo call shape (tDISP-s3): rdi=size_dwords (0x80), rsi/rdx = guest
+        // packet/param buffers. Clear the first mapped buffer arg.
+        var packetAddress = 0UL;
+        foreach (var candidate in new[] { arg1, arg2, arg3 })
+        {
+            if (candidate >= 0x10000)
+            {
+                packetAddress = candidate;
+                break;
+            }
+        }
+
+        var cleared = packetAddress != 0 && TryClearGuestBuffer(ctx, packetAddress, SoftPacketSizeBytes);
+        TracePsml(
+            $"mfsr_get_dispatch_packet_{variant} a0=0x{arg0:X16} a1=0x{arg1:X16} " +
+            $"a2=0x{arg2:X16} a3=0x{arg3:X16} packet=0x{packetAddress:X16} cleared={cleared}");
+        return ctx.SetReturn(0);
+    }
+
+    private static bool TryClearGuestBuffer(CpuContext ctx, ulong address, ulong length)
+    {
+        Span<byte> zeroes = stackalloc byte[4096];
+        zeroes.Clear();
+        for (ulong offset = 0; offset < length;)
+        {
+            var chunkSize = (int)Math.Min((ulong)zeroes.Length, length - offset);
+            if (!ctx.Memory.TryWrite(address + offset, zeroes[..chunkSize]))
+            {
+                return false;
+            }
+
+            offset += unchecked((uint)chunkSize);
+        }
+
+        return true;
+    }
+
+    private static void TracePsml(string message)
+    {
+        if (string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_PSML"), "1", StringComparison.Ordinal))
+        {
+            Console.Error.WriteLine($"[LOADER][TRACE] psml.{message}");
+        }
+    }
+
+    private readonly record struct SharedResourcesState(
+        ulong DescriptorAddress,
+        ulong DirectMemoryAddress,
+        ulong BufferSizeBytes,
+        ulong ContextSizeBytes,
+        ulong PageSizeBytes);
+
+    private readonly record struct ContextState(
+        ulong ContextAddress,
+        ulong SharedResourcesDescriptor,
+        ulong DirectMemoryAddress,
+        ulong BufferSizeBytes,
+        ulong PageSizeBytes,
+        ulong StructSizeBytes);
+}

src/SharpEmu.Libs/Stubs/GameServiceStubs.cs

@@ -59,4 +59,65 @@ public static class GameServiceStubs
     public static int NpUniversalDataSystemCreateEvent(CpuContext ctx) => OkWithHandle(ctx, CpuRegister.Rdi);
     public static int NpUniversalDataSystemPostEvent(CpuContext ctx) => Ok(ctx);
     public static int NpUniversalDataSystemDestroyEvent(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "47UAEuQl+iI", ExportName = "sceNpUniversalDataSystemTerminate",
+        Target = Generation.Gen5, LibraryName = "libSceNpUniversalDataSystem")]
+    public static int NpUniversalDataSystemTerminate(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "0HBYxYAjmf0", ExportName = "sceNpGameIntentTerminate",
+        Target = Generation.Gen5, LibraryName = "libSceNpGameIntent")]
+    public static int NpGameIntentTerminate(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "jqb7HntFQFc", ExportName = "sceWebBrowserDialogInitialize",
+        Target = Generation.Gen5, LibraryName = "libSceWebBrowserDialog")]
+    public static int WebBrowserDialogInitialize(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "ocHtyBwHfys", ExportName = "sceWebBrowserDialogTerminate",
+        Target = Generation.Gen5, LibraryName = "libSceWebBrowserDialog")]
+    public static int WebBrowserDialogTerminate(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "kvYEw2lBndk", ExportName = "sceGameLiveStreamingInitialize",
+        Target = Generation.Gen5, LibraryName = "libSceGameLiveStreaming")]
+    public static int GameLiveStreamingInitialize(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "isruqthpYcw", ExportName = "sceSharePlayInitialize",
+        Target = Generation.Gen5, LibraryName = "libSceSharePlay")]
+    public static int SharePlayInitialize(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "0IL1keINExQ", ExportName = "sceShareTerminate",
+        Target = Generation.Gen5, LibraryName = "libSceShareUtility")]
+    public static int ShareTerminate(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "YBiIdcDPrxs", ExportName = "sceShareFeaturePermit",
+        Target = Generation.Gen5, LibraryName = "libSceShareUtility")]
+    public static int ShareFeaturePermit(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "9TrhuGzberQ", ExportName = "sceVoiceInit",
+        Target = Generation.Gen5, LibraryName = "libSceVoice")]
+    public static int VoiceInit(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "clyKUyi3RYU", ExportName = "sceVoiceSetThreadsParams",
+        Target = Generation.Gen5, LibraryName = "libSceVoice")]
+    public static int VoiceSetThreadsParams(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "dPj4ZtRcIWk", ExportName = "sceContentSearchInit",
+        Target = Generation.Gen5, LibraryName = "libSceContentSearch")]
+    public static int ContentSearchInit(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "zoxb0wEChEM", ExportName = "sceContentDeleteInitialize",
+        Target = Generation.Gen5, LibraryName = "libSceContentDelete")]
+    public static int ContentDeleteInitialize(CpuContext ctx) => Ok(ctx);
+
+    [SysAbiExport(Nid = "Fc8qxlKINYQ", ExportName = "sceVideoRecordingSetInfo",
+        Target = Generation.Gen5, LibraryName = "libSceVideoRecording")]
+    public static int VideoRecordingSetInfo(CpuContext ctx) => Ok(ctx);
+
+    // Captured from GTA V Enhanced (PPSA04264); not in the public NID catalog.
+    // Side-effect-free success — same as unresolved stub behavior that kept boot
+    // moving; reverse the ABI before writing guest memory.
+    #pragma warning disable SHEM006
+    [SysAbiExport(Nid = "Ikfdt-rIqCE", ExportName = "sceUnknownIkfdt",
+        Target = Generation.Gen5, LibraryName = "libKernel")]
+    public static int UnknownIkfdt(CpuContext ctx) => Ok(ctx);
+    #pragma warning restore SHEM006
 }

src/SharpEmu.Libs/SystemService/SystemServiceExports.cs

@@ -18,6 +18,7 @@ public static class SystemServiceExports
     private const int TitleIdFieldSize = 0x10;
 
     private static string? _mainAppTitleId;
+    private static int _noticeScreenSkipFlag;
 
     public static void ConfigureApplicationInfo(string? titleId)
     {
@@ -37,9 +38,11 @@ public static class SystemServiceExports
             return ctx.SetReturn(OrbisSystemServiceErrorParameter);
         }
 
-        // No system notice screen to skip in the emulator; report "do not skip".
+        // Keep the flag state even though the emulator does not display the
+        // system notice screen. Titles use this service as a normal preference
+        // store and expect a later get to observe the value they set.
         Span<byte> flagBytes = stackalloc byte[1];
-        flagBytes[0] = 0;
+        flagBytes[0] = unchecked((byte)Volatile.Read(ref _noticeScreenSkipFlag));
         return ctx.Memory.TryWrite(flagAddress, flagBytes)
             ? ctx.SetReturn(0)
             : ctx.SetReturn((int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
@@ -61,8 +64,13 @@ public static class SystemServiceExports
         ExportName = "sceSystemServiceSetNoticeScreenSkipFlag",
         Target = Generation.Gen5,
         LibraryName = "libSceSystemService")]
-    public static int SystemServiceSetNoticeScreenSkipFlag(CpuContext ctx) =>
-        ctx.SetReturn(0);
+    public static int SystemServiceSetNoticeScreenSkipFlag(CpuContext ctx)
+    {
+        // The native API takes the flag value in the first argument. Treat any
+        // non-zero value as true, matching the bool-like PS5 ABI.
+        Volatile.Write(ref _noticeScreenSkipFlag, ctx[CpuRegister.Rdi] != 0 ? 1 : 0);
+        return ctx.SetReturn(0);
+    }
 
     [SysAbiExport(
         Nid = "4veE0XiIugA",
@@ -231,4 +239,7 @@ public static class SystemServiceExports
         Target = Generation.Gen4 | Generation.Gen5,
         LibraryName = "libSceSystemService")]
     public static int SystemServiceReportAbnormalTermination(CpuContext ctx) => ctx.SetReturn(0);
+
+    internal static void ResetForTests() =>
+        Volatile.Write(ref _noticeScreenSkipFlag, 0);
 }

src/SharpEmu.Libs/VideoOut/VulkanDetilePass.cs

@@ -0,0 +1,894 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using SharpEmu.Libs.Agc;
+using SharpEmu.ShaderCompiler.Vulkan;
+using Silk.NET.Vulkan;
+using VkBuffer = Silk.NET.Vulkan.Buffer;
+
+namespace SharpEmu.Libs.VideoOut;
+
+/// <summary>
+/// Self-contained GPU deswizzle pass: runs the ExactXor detile equation from
+/// <see cref="GnmTiling.GetDetileParams"/> as a Vulkan compute shader
+/// (<see cref="SpirvFixedShaders.CreateDetileCompute"/>), writing a linear buffer
+/// and copying it into a sampled image — the GPU equivalent of the CPU
+/// <c>GnmTiling.TryDetile</c> + staging upload.
+///
+/// Two entry points share the same (verified) recording:
+/// <see cref="DetileIntoImage"/> is a self-contained one-shot (submit + wait) used
+/// by the isolation self-test; <see cref="RecordDetile"/> records into a caller's
+/// command buffer and hands back its transient buffers + descriptor pool for the
+/// caller to retire with that command buffer's fence — the render-path variant,
+/// which must never block the render thread.
+///
+/// Only ExactXor 4-bytes/element surfaces are handled; <see cref="Supports"/> lets
+/// the caller fall back to the CPU path for everything else.
+/// </summary>
+internal sealed unsafe class VulkanDetilePass : IDisposable
+{
+    private const uint LocalSize = 8;
+    private const uint PushConstantBytes = 11 * sizeof(uint);
+
+    private readonly Vk _vk;
+    private readonly Device _device;
+    private readonly Queue _queue;
+    private readonly PhysicalDevice _physicalDevice;
+    private readonly uint _queueFamilyIndex;
+
+    private ShaderModule _shaderModule;
+    private DescriptorSetLayout _descriptorSetLayout;
+    private PipelineLayout _pipelineLayout;
+    private Pipeline _pipeline;
+    private CommandPool _commandPool;
+    private bool _initialized;
+    private bool _disposed;
+
+    private PhysicalDeviceMemoryProperties _memoryProperties;
+    private bool _memoryPropertiesLoaded;
+
+    private readonly Dictionary<int[], TermBuffer> _xorTermBuffers = new(ReferenceComparer.Instance);
+    private readonly Dictionary<int[], TermBuffer> _blockTermBuffers = new(ReferenceComparer.Instance);
+    private TermBuffer _placeholderTermBuffer;
+
+    private readonly Dictionary<(ulong Bucket, bool HostVisible), Stack<Allocation>> _bufferPool = new();
+    private readonly List<Allocation> _allAllocations = new();
+
+    private readonly Stack<DescriptorSet> _freeDescriptorSets = new();
+    private readonly List<DescriptorPool> _descriptorPools = new();
+    private const uint DescriptorSetsPerPool = 64;
+    private const ulong MinimumBufferBucket = 4096;
+
+    public VulkanDetilePass(
+        Vk vk,
+        Device device,
+        Queue queue,
+        PhysicalDevice physicalDevice,
+        uint queueFamilyIndex)
+    {
+        _vk = vk;
+        _device = device;
+        _queue = queue;
+        _physicalDevice = physicalDevice;
+        _queueFamilyIndex = queueFamilyIndex;
+    }
+
+    /// <summary>
+    /// The kernel handles the exact-XOR and block-table modes at 4/8/16
+    /// bytes-per-element (one, two, or four 32-bit words per element). 1/2 bpp are
+    /// sub-word and stay on the CPU.
+    /// </summary>
+    public static bool Supports(in DetileParams parameters) =>
+        (parameters.Equation == DetileEquation.ExactXor ||
+         parameters.Equation == DetileEquation.BlockTable) &&
+        parameters.BytesPerElement is 4 or 8 or 16;
+
+    /// <summary>Opaque handle to the pooled resources one recorded detile is using.
+    /// The caller hands it back to <see cref="Retire"/> once the command buffer they
+    /// were recorded into has completed; nothing is destroyed, the buffers and the
+    /// descriptor set return to this pass's free lists for the next texture.</summary>
+    public sealed class Transients
+    {
+        internal static readonly Transients Empty = new();
+
+        internal Allocation Tiled;
+        internal Allocation Output;
+        internal DescriptorSet Set;
+        internal bool Rented;
+    }
+
+    internal readonly record struct Allocation(
+        VkBuffer Buffer,
+        DeviceMemory Memory,
+        ulong Capacity,
+        bool HostVisible);
+
+    private readonly record struct TermBuffer(VkBuffer Buffer, ulong ByteSize);
+
+    private struct DetileResources
+    {
+        public Allocation Tiled;
+        public Allocation Output;
+        public DescriptorSet Set;
+        public ulong OutputBytes;
+        public uint SrcSliceElements;
+        public uint EquationValue;
+        public uint UintsPerElement;
+    }
+
+    private sealed class ReferenceComparer : IEqualityComparer<int[]>
+    {
+        public static readonly ReferenceComparer Instance = new();
+
+        public bool Equals(int[]? x, int[]? y) => ReferenceEquals(x, y);
+
+        public int GetHashCode(int[] obj) => RuntimeHelpers.GetHashCode(obj);
+    }
+
+    /// <summary>
+    /// Records the deswizzle of <paramref name="tiled"/> into <paramref name="image"/>
+    /// (<paramref name="texelWidth"/> x <paramref name="texelHeight"/> texels x
+    /// <paramref name="layers"/> array slices, currently in
+    /// <paramref name="currentLayout"/>) onto <paramref name="commandBuffer"/>,
+    /// leaving the image <see cref="ImageLayout.ShaderReadOnlyOptimal"/>. The kernel
+    /// iterates the element grid from <paramref name="parameters"/> (for
+    /// block-compressed formats a 4x4 block is one element, so the element grid is
+    /// smaller than the texel grid). The tiled buffer holds the array slices packed
+    /// contiguously (each an independently tiled 2D surface). Does not submit; the
+    /// caller retires <paramref name="transients"/> with the command buffer's fence.
+    /// Returns false (with empty transients) when unsupported.
+    /// </summary>
+    public bool RecordDetile(
+        CommandBuffer commandBuffer,
+        Image image,
+        ImageLayout currentLayout,
+        uint texelWidth,
+        uint texelHeight,
+        uint layers,
+        ReadOnlySpan<byte> tiled,
+        in DetileParams parameters,
+        out Transients transients)
+    {
+        transients = Transients.Empty;
+        if (_disposed || !Supports(parameters) || texelWidth == 0 || texelHeight == 0 || layers == 0 ||
+            tiled.IsEmpty || tiled.Length % (int)(layers * (uint)parameters.BytesPerElement) != 0)
+        {
+            return false;
+        }
+
+        EnsurePipeline();
+
+        var resources = default(DetileResources);
+        try
+        {
+            PrepareResources(tiled, parameters, layers, ref resources);
+            RecordCommands(commandBuffer, in resources, image, currentLayout, texelWidth, texelHeight, layers, in parameters);
+        }
+        catch
+        {
+            ReleaseResources(in resources);
+            throw;
+        }
+
+        transients = new Transients
+        {
+            Tiled = resources.Tiled,
+            Output = resources.Output,
+            Set = resources.Set,
+            Rented = true,
+        };
+        return true;
+    }
+
+    public void Retire(Transients transients)
+    {
+        if (_disposed || transients is null || !transients.Rented)
+        {
+            return;
+        }
+
+        transients.Rented = false;
+        ReturnBuffer(transients.Tiled);
+        ReturnBuffer(transients.Output);
+        _freeDescriptorSets.Push(transients.Set);
+    }
+
+    /// <summary>
+    /// One-shot variant used by the isolation self-test: records the detile onto a
+    /// private command buffer, submits, waits, and frees every transient. Never
+    /// call this on the render thread — its blocking wait would deadlock the
+    /// present pipeline; use <see cref="RecordDetile"/> there.
+    /// </summary>
+    public bool DetileIntoImage(
+        Image image,
+        ImageLayout currentLayout,
+        uint texelWidth,
+        uint texelHeight,
+        uint layers,
+        ReadOnlySpan<byte> tiled,
+        in DetileParams parameters)
+    {
+        if (_disposed || !Supports(parameters) || texelWidth == 0 || texelHeight == 0 || layers == 0 ||
+            tiled.IsEmpty || tiled.Length % (int)(layers * (uint)parameters.BytesPerElement) != 0)
+        {
+            return false;
+        }
+
+        EnsurePipeline();
+
+        var resources = default(DetileResources);
+        CommandBuffer commandBuffer = default;
+        Fence fence = default;
+        try
+        {
+            PrepareResources(tiled, parameters, layers, ref resources);
+
+            commandBuffer = AllocateCommandBuffer();
+            BeginCommandBuffer(commandBuffer);
+            RecordCommands(commandBuffer, in resources, image, currentLayout, texelWidth, texelHeight, layers, in parameters);
+            Check(_vk.EndCommandBuffer(commandBuffer), "vkEndCommandBuffer(detile)");
+
+            fence = CreateFence();
+            var submitInfo = new SubmitInfo
+            {
+                SType = StructureType.SubmitInfo,
+                CommandBufferCount = 1,
+                PCommandBuffers = &commandBuffer,
+            };
+            Check(_vk.QueueSubmit(_queue, 1, &submitInfo, fence), "vkQueueSubmit(detile)");
+            Check(_vk.WaitForFences(_device, 1, &fence, true, ulong.MaxValue), "vkWaitForFences(detile)");
+            return true;
+        }
+        finally
+        {
+            if (fence.Handle != 0)
+            {
+                _vk.DestroyFence(_device, fence, null);
+            }
+
+            if (commandBuffer.Handle != 0)
+            {
+                _vk.FreeCommandBuffers(_device, _commandPool, 1, &commandBuffer);
+            }
+
+            ReleaseResources(in resources);
+        }
+    }
+
+    private void PrepareResources(ReadOnlySpan<byte> tiled, in DetileParams parameters, uint layers, ref DetileResources resources)
+    {
+        // Binding 1 carries the within-block offset table, binding 2 the Y terms.
+        // ExactXor: xTerm/yTerm are byte offsets; the kernel indexes a uint[], so it
+        // wants element offsets — for a power-of-two element size the low
+        // log2(bpp) bits of every term are 0, so the right shift is exact.
+        // BlockTable: GetDetileParams' block table is already element offsets; it
+        // goes in binding 1 and binding 2 is an unused placeholder.
+        TermBuffer xTerm;
+        TermBuffer yTerm;
+        if (parameters.Equation == DetileEquation.BlockTable)
+        {
+            xTerm = GetTermBuffer(_blockTermBuffers, parameters.BlockTable, shift: 0);
+            yTerm = GetPlaceholderTermBuffer();
+            resources.EquationValue = 1;
+        }
+        else
+        {
+            var shift = BitOperations.TrailingZeroCount((uint)parameters.BytesPerElement);
+            xTerm = GetTermBuffer(_xorTermBuffers, parameters.XByteTerm, shift);
+            yTerm = GetTermBuffer(_xorTermBuffers, parameters.YByteTerm, shift);
+            resources.EquationValue = 0;
+        }
+
+        // The array slices are packed contiguously in the tiled buffer, so each
+        // slice's element stride is the whole tiled buffer split evenly by layer.
+        // Element sizes are in bytes-per-element; the kernel moves bpp/4 words each.
+        var bytesPerElement = (uint)parameters.BytesPerElement;
+        resources.UintsPerElement = bytesPerElement / sizeof(uint);
+        resources.SrcSliceElements = (uint)((ulong)tiled.Length / bytesPerElement / layers);
+        resources.OutputBytes =
+            (ulong)parameters.ElementsWide * (ulong)parameters.ElementsHigh * bytesPerElement * layers;
+
+        resources.Tiled = RentBuffer((ulong)tiled.Length, hostVisible: true);
+        UploadBytes(resources.Tiled.Memory, tiled);
+        resources.Output = RentBuffer(resources.OutputBytes, hostVisible: false);
+
+        resources.Set = RentDescriptorSet();
+        WriteDescriptors(
+            resources.Set,
+            (resources.Tiled.Buffer, (ulong)tiled.Length),
+            (xTerm.Buffer, xTerm.ByteSize),
+            (yTerm.Buffer, yTerm.ByteSize),
+            (resources.Output.Buffer, resources.OutputBytes));
+    }
+
+    private TermBuffer GetTermBuffer(Dictionary<int[], TermBuffer> cache, int[] table, int shift)
+    {
+        if (cache.TryGetValue(table, out var cached))
+        {
+            return cached;
+        }
+
+        var terms = ToElementTerms(table, shift);
+        var byteSize = (ulong)terms.Length * sizeof(uint);
+        var allocation = CreateBuffer(byteSize, hostVisible: true);
+        UploadUInts(allocation.Memory, terms);
+        var termBuffer = new TermBuffer(allocation.Buffer, byteSize);
+        cache[table] = termBuffer;
+        return termBuffer;
+    }
+
+    private TermBuffer GetPlaceholderTermBuffer()
+    {
+        if (_placeholderTermBuffer.Buffer.Handle != 0)
+        {
+            return _placeholderTermBuffer;
+        }
+
+        var allocation = CreateBuffer(sizeof(uint), hostVisible: true);
+        UploadUInts(allocation.Memory, [0u]);
+        _placeholderTermBuffer = new TermBuffer(allocation.Buffer, sizeof(uint));
+        return _placeholderTermBuffer;
+    }
+
+    private static ulong BucketFor(ulong size)
+    {
+        var bucket = MinimumBufferBucket;
+        while (bucket < size)
+        {
+            bucket <<= 1;
+        }
+
+        return bucket;
+    }
+
+    private Allocation RentBuffer(ulong size, bool hostVisible)
+    {
+        var bucket = BucketFor(size);
+        if (_bufferPool.TryGetValue((bucket, hostVisible), out var free) && free.Count > 0)
+        {
+            return free.Pop();
+        }
+
+        return CreateBuffer(bucket, hostVisible);
+    }
+
+    private void ReturnBuffer(Allocation allocation)
+    {
+        if (allocation.Buffer.Handle == 0)
+        {
+            return;
+        }
+
+        var key = (allocation.Capacity, allocation.HostVisible);
+        if (!_bufferPool.TryGetValue(key, out var free))
+        {
+            free = new Stack<Allocation>();
+            _bufferPool[key] = free;
+        }
+
+        free.Push(allocation);
+    }
+
+    private DescriptorSet RentDescriptorSet()
+    {
+        if (_freeDescriptorSets.Count > 0)
+        {
+            return _freeDescriptorSets.Pop();
+        }
+
+        var poolSize = new DescriptorPoolSize
+        {
+            Type = DescriptorType.StorageBuffer,
+            DescriptorCount = 4 * DescriptorSetsPerPool,
+        };
+        var poolInfo = new DescriptorPoolCreateInfo
+        {
+            SType = StructureType.DescriptorPoolCreateInfo,
+            MaxSets = DescriptorSetsPerPool,
+            PoolSizeCount = 1,
+            PPoolSizes = &poolSize,
+        };
+        Check(
+            _vk.CreateDescriptorPool(_device, &poolInfo, null, out var pool),
+            "vkCreateDescriptorPool(detile)");
+        _descriptorPools.Add(pool);
+
+        var layouts = stackalloc DescriptorSetLayout[(int)DescriptorSetsPerPool];
+        for (var index = 0; index < DescriptorSetsPerPool; index++)
+        {
+            layouts[index] = _descriptorSetLayout;
+        }
+
+        var sets = stackalloc DescriptorSet[(int)DescriptorSetsPerPool];
+        var allocateInfo = new DescriptorSetAllocateInfo
+        {
+            SType = StructureType.DescriptorSetAllocateInfo,
+            DescriptorPool = pool,
+            DescriptorSetCount = DescriptorSetsPerPool,
+            PSetLayouts = layouts,
+        };
+        Check(
+            _vk.AllocateDescriptorSets(_device, &allocateInfo, sets),
+            "vkAllocateDescriptorSets(detile)");
+        for (var index = 0; index < DescriptorSetsPerPool; index++)
+        {
+            _freeDescriptorSets.Push(sets[index]);
+        }
+
+        return _freeDescriptorSets.Pop();
+    }
+
+    private void ReleaseResources(in DetileResources resources)
+    {
+        ReturnBuffer(resources.Tiled);
+        ReturnBuffer(resources.Output);
+        if (resources.Set.Handle != 0)
+        {
+            _freeDescriptorSets.Push(resources.Set);
+        }
+    }
+
+    private void RecordCommands(
+        CommandBuffer commandBuffer,
+        in DetileResources resources,
+        Image image,
+        ImageLayout currentLayout,
+        uint texelWidth,
+        uint texelHeight,
+        uint layers,
+        in DetileParams parameters)
+    {
+        // The kernel iterates the element grid (smaller than the texel grid for
+        // block-compressed formats); the image copy below uses the texel grid.
+        var elementsWide = (uint)parameters.ElementsWide;
+        var elementsHigh = (uint)parameters.ElementsHigh;
+
+        var descriptorSet = resources.Set;
+        _vk.CmdBindPipeline(commandBuffer, PipelineBindPoint.Compute, _pipeline);
+        _vk.CmdBindDescriptorSets(
+            commandBuffer, PipelineBindPoint.Compute, _pipelineLayout, 0, 1, &descriptorSet, 0, null);
+
+        Span<uint> push =
+        [
+            elementsWide,
+            elementsHigh,
+            (uint)parameters.BlockWidth,
+            (uint)parameters.BlockHeight,
+            (uint)parameters.BlockElements,
+            (uint)parameters.BlocksPerRow,
+            (uint)parameters.XMask,
+            (uint)parameters.YMask,
+            resources.SrcSliceElements,
+            resources.EquationValue,
+            resources.UintsPerElement,
+        ];
+        fixed (uint* pushPointer = push)
+        {
+            _vk.CmdPushConstants(
+                commandBuffer, _pipelineLayout, ShaderStageFlags.ComputeBit, 0, PushConstantBytes, pushPointer);
+        }
+
+        // X is widened by uintsPerElement (each thread copies one word); one
+        // dispatch-Z layer per array slice.
+        _vk.CmdDispatch(
+            commandBuffer,
+            (elementsWide * resources.UintsPerElement + LocalSize - 1) / LocalSize,
+            (elementsHigh + LocalSize - 1) / LocalSize,
+            layers);
+
+        // Compute store -> transfer read on the linear output buffer.
+        var outputBarrier = new BufferMemoryBarrier
+        {
+            SType = StructureType.BufferMemoryBarrier,
+            SrcAccessMask = AccessFlags.ShaderWriteBit,
+            DstAccessMask = AccessFlags.TransferReadBit,
+            SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            Buffer = resources.Output.Buffer,
+            Offset = 0,
+            Size = resources.OutputBytes,
+        };
+        _vk.CmdPipelineBarrier(
+            commandBuffer,
+            PipelineStageFlags.ComputeShaderBit,
+            PipelineStageFlags.TransferBit,
+            0,
+            0,
+            null,
+            1,
+            &outputBarrier,
+            0,
+            null);
+
+        var initialized = currentLayout == ImageLayout.ShaderReadOnlyOptimal;
+        TransitionImage(
+            commandBuffer,
+            image,
+            currentLayout,
+            ImageLayout.TransferDstOptimal,
+            initialized ? AccessFlags.ShaderReadBit : 0,
+            AccessFlags.TransferWriteBit,
+            initialized ? PipelineStageFlags.FragmentShaderBit : PipelineStageFlags.TopOfPipeBit,
+            PipelineStageFlags.TransferBit,
+            layers);
+
+        // The output buffer is layer-major, tightly packed (BufferRowLength 0 =>
+        // one element-row per texel-row, which for compressed formats is the block
+        // row), so a single copy fills every array layer. Extent is in texels.
+        var copyRegion = new BufferImageCopy
+        {
+            BufferOffset = 0,
+            BufferRowLength = 0,
+            BufferImageHeight = 0,
+            ImageSubresource = new ImageSubresourceLayers(ImageAspectFlags.ColorBit, 0, 0, layers),
+            ImageOffset = default,
+            ImageExtent = new Extent3D(texelWidth, texelHeight, 1),
+        };
+        _vk.CmdCopyBufferToImage(
+            commandBuffer, resources.Output.Buffer, image, ImageLayout.TransferDstOptimal, 1, &copyRegion);
+
+        TransitionImage(
+            commandBuffer,
+            image,
+            ImageLayout.TransferDstOptimal,
+            ImageLayout.ShaderReadOnlyOptimal,
+            AccessFlags.TransferWriteBit,
+            AccessFlags.ShaderReadBit,
+            PipelineStageFlags.TransferBit,
+            PipelineStageFlags.FragmentShaderBit,
+            layers);
+    }
+
+    private void EnsurePipeline()
+    {
+        if (_initialized)
+        {
+            return;
+        }
+
+        var spirv = SpirvFixedShaders.CreateDetileCompute();
+        fixed (byte* code = spirv)
+        {
+            var moduleInfo = new ShaderModuleCreateInfo
+            {
+                SType = StructureType.ShaderModuleCreateInfo,
+                CodeSize = (nuint)spirv.Length,
+                PCode = (uint*)code,
+            };
+            Check(
+                _vk.CreateShaderModule(_device, &moduleInfo, null, out _shaderModule),
+                "vkCreateShaderModule(detile)");
+        }
+
+        var bindings = stackalloc DescriptorSetLayoutBinding[4];
+        for (uint index = 0; index < 4; index++)
+        {
+            bindings[index] = new DescriptorSetLayoutBinding
+            {
+                Binding = index,
+                DescriptorType = DescriptorType.StorageBuffer,
+                DescriptorCount = 1,
+                StageFlags = ShaderStageFlags.ComputeBit,
+            };
+        }
+
+        var layoutInfo = new DescriptorSetLayoutCreateInfo
+        {
+            SType = StructureType.DescriptorSetLayoutCreateInfo,
+            BindingCount = 4,
+            PBindings = bindings,
+        };
+        Check(
+            _vk.CreateDescriptorSetLayout(_device, &layoutInfo, null, out _descriptorSetLayout),
+            "vkCreateDescriptorSetLayout(detile)");
+
+        var pushRange = new PushConstantRange
+        {
+            StageFlags = ShaderStageFlags.ComputeBit,
+            Offset = 0,
+            Size = PushConstantBytes,
+        };
+        var setLayout = _descriptorSetLayout;
+        var pipelineLayoutInfo = new PipelineLayoutCreateInfo
+        {
+            SType = StructureType.PipelineLayoutCreateInfo,
+            SetLayoutCount = 1,
+            PSetLayouts = &setLayout,
+            PushConstantRangeCount = 1,
+            PPushConstantRanges = &pushRange,
+        };
+        Check(
+            _vk.CreatePipelineLayout(_device, &pipelineLayoutInfo, null, out _pipelineLayout),
+            "vkCreatePipelineLayout(detile)");
+
+        ReadOnlySpan<byte> entryPoint = "main\0"u8;
+        fixed (byte* entry = entryPoint)
+        {
+            var pipelineInfo = new ComputePipelineCreateInfo
+            {
+                SType = StructureType.ComputePipelineCreateInfo,
+                Layout = _pipelineLayout,
+                Stage = new PipelineShaderStageCreateInfo
+                {
+                    SType = StructureType.PipelineShaderStageCreateInfo,
+                    Stage = ShaderStageFlags.ComputeBit,
+                    Module = _shaderModule,
+                    PName = entry,
+                },
+            };
+            Check(
+                _vk.CreateComputePipelines(_device, default, 1, &pipelineInfo, null, out _pipeline),
+                "vkCreateComputePipelines(detile)");
+        }
+
+        var poolInfo = new CommandPoolCreateInfo
+        {
+            SType = StructureType.CommandPoolCreateInfo,
+            QueueFamilyIndex = _queueFamilyIndex,
+            Flags = CommandPoolCreateFlags.ResetCommandBufferBit,
+        };
+        Check(
+            _vk.CreateCommandPool(_device, &poolInfo, null, out _commandPool),
+            "vkCreateCommandPool(detile)");
+
+        _initialized = true;
+    }
+
+    private static uint[] ToElementTerms(int[] byteTerms, int shift)
+    {
+        var terms = new uint[byteTerms.Length];
+        for (var index = 0; index < byteTerms.Length; index++)
+        {
+            terms[index] = (uint)byteTerms[index] >> shift;
+        }
+
+        return terms;
+    }
+
+    private Allocation CreateBuffer(ulong size, bool hostVisible)
+    {
+        var bufferInfo = new BufferCreateInfo
+        {
+            SType = StructureType.BufferCreateInfo,
+            Size = size,
+            Usage = BufferUsageFlags.StorageBufferBit | BufferUsageFlags.TransferSrcBit,
+            SharingMode = SharingMode.Exclusive,
+        };
+        Check(_vk.CreateBuffer(_device, &bufferInfo, null, out var buffer), "vkCreateBuffer(detile)");
+
+        _vk.GetBufferMemoryRequirements(_device, buffer, out var requirements);
+        var required = hostVisible
+            ? MemoryPropertyFlags.HostVisibleBit | MemoryPropertyFlags.HostCoherentBit
+            : MemoryPropertyFlags.DeviceLocalBit;
+        var allocateInfo = new MemoryAllocateInfo
+        {
+            SType = StructureType.MemoryAllocateInfo,
+            AllocationSize = requirements.Size,
+            MemoryTypeIndex = FindMemoryType(requirements.MemoryTypeBits, required, hostVisible),
+        };
+        Check(_vk.AllocateMemory(_device, &allocateInfo, null, out var memory), "vkAllocateMemory(detile)");
+        Check(_vk.BindBufferMemory(_device, buffer, memory, 0), "vkBindBufferMemory(detile)");
+        var allocation = new Allocation(buffer, memory, size, hostVisible);
+        _allAllocations.Add(allocation);
+        return allocation;
+    }
+
+    private uint FindMemoryType(uint typeBits, MemoryPropertyFlags requiredFlags, bool hostVisible)
+    {
+        if (!_memoryPropertiesLoaded)
+        {
+            _vk.GetPhysicalDeviceMemoryProperties(_physicalDevice, out _memoryProperties);
+            _memoryPropertiesLoaded = true;
+        }
+
+        fixed (PhysicalDeviceMemoryProperties* properties = &_memoryProperties)
+        {
+            var memoryTypes = &properties->MemoryTypes.Element0;
+            for (uint index = 0; index < properties->MemoryTypeCount; index++)
+            {
+                if ((typeBits & (1u << (int)index)) != 0 &&
+                    (memoryTypes[index].PropertyFlags & requiredFlags) == requiredFlags)
+                {
+                    return index;
+                }
+            }
+
+            if (!hostVisible)
+            {
+                for (uint index = 0; index < properties->MemoryTypeCount; index++)
+                {
+                    if ((typeBits & (1u << (int)index)) != 0)
+                    {
+                        return index;
+                    }
+                }
+            }
+        }
+
+        throw new InvalidOperationException("No compatible Vulkan memory type for detile.");
+    }
+
+    private void UploadBytes(DeviceMemory memory, ReadOnlySpan<byte> data)
+    {
+        void* mapped;
+        Check(_vk.MapMemory(_device, memory, 0, (ulong)data.Length, 0, &mapped), "vkMapMemory(detile)");
+        data.CopyTo(new Span<byte>(mapped, data.Length));
+        _vk.UnmapMemory(_device, memory);
+    }
+
+    private void UploadUInts(DeviceMemory memory, uint[] data)
+    {
+        void* mapped;
+        var byteCount = (ulong)data.Length * sizeof(uint);
+        Check(_vk.MapMemory(_device, memory, 0, byteCount, 0, &mapped), "vkMapMemory(detile terms)");
+        data.AsSpan().CopyTo(new Span<uint>(mapped, data.Length));
+        _vk.UnmapMemory(_device, memory);
+    }
+
+    private void WriteDescriptors(
+        DescriptorSet descriptorSet,
+        (VkBuffer Buffer, ulong Size) binding0,
+        (VkBuffer Buffer, ulong Size) binding1,
+        (VkBuffer Buffer, ulong Size) binding2,
+        (VkBuffer Buffer, ulong Size) binding3)
+    {
+        var buffers = stackalloc DescriptorBufferInfo[4]
+        {
+            new DescriptorBufferInfo { Buffer = binding0.Buffer, Offset = 0, Range = binding0.Size },
+            new DescriptorBufferInfo { Buffer = binding1.Buffer, Offset = 0, Range = binding1.Size },
+            new DescriptorBufferInfo { Buffer = binding2.Buffer, Offset = 0, Range = binding2.Size },
+            new DescriptorBufferInfo { Buffer = binding3.Buffer, Offset = 0, Range = binding3.Size },
+        };
+
+        var writes = stackalloc WriteDescriptorSet[4];
+        for (uint index = 0; index < 4; index++)
+        {
+            writes[index] = new WriteDescriptorSet
+            {
+                SType = StructureType.WriteDescriptorSet,
+                DstSet = descriptorSet,
+                DstBinding = index,
+                DstArrayElement = 0,
+                DescriptorCount = 1,
+                DescriptorType = DescriptorType.StorageBuffer,
+                PBufferInfo = &buffers[index],
+            };
+        }
+
+        _vk.UpdateDescriptorSets(_device, 4, writes, 0, null);
+    }
+
+    private CommandBuffer AllocateCommandBuffer()
+    {
+        var allocateInfo = new CommandBufferAllocateInfo
+        {
+            SType = StructureType.CommandBufferAllocateInfo,
+            CommandPool = _commandPool,
+            Level = CommandBufferLevel.Primary,
+            CommandBufferCount = 1,
+        };
+        Check(
+            _vk.AllocateCommandBuffers(_device, &allocateInfo, out var commandBuffer),
+            "vkAllocateCommandBuffers(detile)");
+        return commandBuffer;
+    }
+
+    private void BeginCommandBuffer(CommandBuffer commandBuffer)
+    {
+        var beginInfo = new CommandBufferBeginInfo
+        {
+            SType = StructureType.CommandBufferBeginInfo,
+            Flags = CommandBufferUsageFlags.OneTimeSubmitBit,
+        };
+        Check(_vk.BeginCommandBuffer(commandBuffer, &beginInfo), "vkBeginCommandBuffer(detile)");
+    }
+
+    private void TransitionImage(
+        CommandBuffer commandBuffer,
+        Image image,
+        ImageLayout oldLayout,
+        ImageLayout newLayout,
+        AccessFlags srcAccess,
+        AccessFlags dstAccess,
+        PipelineStageFlags srcStage,
+        PipelineStageFlags dstStage,
+        uint layers)
+    {
+        var barrier = new ImageMemoryBarrier
+        {
+            SType = StructureType.ImageMemoryBarrier,
+            SrcAccessMask = srcAccess,
+            DstAccessMask = dstAccess,
+            OldLayout = oldLayout,
+            NewLayout = newLayout,
+            SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            Image = image,
+            SubresourceRange = new ImageSubresourceRange(ImageAspectFlags.ColorBit, 0, 1, 0, layers),
+        };
+        _vk.CmdPipelineBarrier(commandBuffer, srcStage, dstStage, 0, 0, null, 0, null, 1, &barrier);
+    }
+
+    private Fence CreateFence()
+    {
+        var fenceInfo = new FenceCreateInfo { SType = StructureType.FenceCreateInfo };
+        Check(_vk.CreateFence(_device, &fenceInfo, null, out var fence), "vkCreateFence(detile)");
+        return fence;
+    }
+
+    private void DestroyBuffer(VkBuffer buffer, DeviceMemory memory)
+    {
+        if (buffer.Handle != 0)
+        {
+            _vk.DestroyBuffer(_device, buffer, null);
+        }
+
+        if (memory.Handle != 0)
+        {
+            _vk.FreeMemory(_device, memory, null);
+        }
+    }
+
+    private void Check(Result result, string operation)
+    {
+        if (result != Result.Success)
+        {
+            throw new InvalidOperationException($"{operation} failed: {result}");
+        }
+    }
+
+    public void Dispose()
+    {
+        if (_disposed)
+        {
+            return;
+        }
+
+        _disposed = true;
+
+        foreach (var allocation in _allAllocations)
+        {
+            DestroyBuffer(allocation.Buffer, allocation.Memory);
+        }
+
+        _allAllocations.Clear();
+        _bufferPool.Clear();
+        _xorTermBuffers.Clear();
+        _blockTermBuffers.Clear();
+        _placeholderTermBuffer = default;
+        _freeDescriptorSets.Clear();
+        foreach (var pool in _descriptorPools)
+        {
+            _vk.DestroyDescriptorPool(_device, pool, null);
+        }
+
+        _descriptorPools.Clear();
+
+        if (_pipeline.Handle != 0)
+        {
+            _vk.DestroyPipeline(_device, _pipeline, null);
+        }
+
+        if (_pipelineLayout.Handle != 0)
+        {
+            _vk.DestroyPipelineLayout(_device, _pipelineLayout, null);
+        }
+
+        if (_descriptorSetLayout.Handle != 0)
+        {
+            _vk.DestroyDescriptorSetLayout(_device, _descriptorSetLayout, null);
+        }
+
+        if (_shaderModule.Handle != 0)
+        {
+            _vk.DestroyShaderModule(_device, _shaderModule, null);
+        }
+
+        if (_commandPool.Handle != 0)
+        {
+            _vk.DestroyCommandPool(_device, _commandPool, null);
+        }
+    }
+}

src/SharpEmu.Libs/VideoOut/VulkanDetileSelfTest.cs

@@ -0,0 +1,482 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Agc;
+using Silk.NET.Vulkan;
+using VkBuffer = Silk.NET.Vulkan.Buffer;
+
+namespace SharpEmu.Libs.VideoOut;
+
+/// <summary>
+/// Opt-in GPU equivalence check for <see cref="VulkanDetilePass"/>. When
+/// SHARPEMU_DETILE_SELFTEST=1 it builds a known tiled surface, deswizzles it on
+/// the GPU into a real image, reads the image back, and compares against the CPU
+/// <see cref="GnmTiling.TryDetile"/> — the same equivalence the unit test proves
+/// for the params, now end-to-end through the actual Vulkan pass. It logs
+/// [DETILE-SELFTEST] PASS/FAIL and never throws into startup (any failure is
+/// caught and logged), so it is safe to leave wired.
+/// </summary>
+internal static unsafe class VulkanDetileSelfTest
+{
+    private const uint Width = 256;
+    private const uint Height = 256;
+
+    // (swizzleMode, bytesPerElement, image format). Mode 27 is exact-XOR, mode 8
+    // (64 KiB Z) is block-table — both branches. bpp 4/8/16 exercises the
+    // one/two/four-words-per-element copy. These formats are non-block-compressed
+    // (element grid == texel grid), so Width/Height are both element and texel dims.
+    private static readonly (uint Mode, int Bpp, Format Format)[] Cases =
+    [
+        (27, 4, Format.R8G8B8A8Unorm),
+        (8, 4, Format.R8G8B8A8Unorm),
+        (27, 8, Format.R32G32Uint),
+        (27, 16, Format.R32G32B32A32Uint),
+    ];
+
+    public static void RunIfRequested(
+        Vk vk,
+        Device device,
+        Queue queue,
+        PhysicalDevice physicalDevice,
+        uint queueFamilyIndex)
+    {
+        if (Environment.GetEnvironmentVariable("SHARPEMU_DETILE_SELFTEST") != "1")
+        {
+            return;
+        }
+
+        try
+        {
+            Run(vk, device, queue, physicalDevice, queueFamilyIndex);
+        }
+        catch (Exception exception)
+        {
+            Console.Error.WriteLine($"[DETILE-SELFTEST] FAIL (exception): {exception.Message}");
+        }
+    }
+
+    private static void Run(
+        Vk vk,
+        Device device,
+        Queue queue,
+        PhysicalDevice physicalDevice,
+        uint queueFamilyIndex)
+    {
+        using var pass = new VulkanDetilePass(vk, device, queue, physicalDevice, queueFamilyIndex);
+        var commandPool = CreateCommandPool(vk, device, queueFamilyIndex);
+        try
+        {
+            foreach (var (mode, bpp, format) in Cases)
+            {
+                // A plain 2D texture (1 layer) and an array texture (2 layers) — the
+                // arrayed case exercises the kernel's dispatch-Z slice addressing.
+                RunCase(vk, device, physicalDevice, queue, commandPool, pass, mode, bpp, format, layers: 1);
+                RunCase(vk, device, physicalDevice, queue, commandPool, pass, mode, bpp, format, layers: 2);
+            }
+        }
+        finally
+        {
+            if (commandPool.Handle != 0)
+            {
+                vk.DestroyCommandPool(device, commandPool, null);
+            }
+        }
+    }
+
+    private static void RunCase(
+        Vk vk,
+        Device device,
+        PhysicalDevice physicalDevice,
+        Queue queue,
+        CommandPool commandPool,
+        VulkanDetilePass pass,
+        uint swizzleMode,
+        int bytesPerElement,
+        Format format,
+        uint layers)
+    {
+        var parameters = GnmTiling.GetDetileParams(swizzleMode, bytesPerElement, (int)Width, (int)Height);
+        if (!parameters.IsSupported || !VulkanDetilePass.Supports(parameters))
+        {
+            Console.Error.WriteLine(
+                $"[DETILE-SELFTEST] FAIL: mode {swizzleMode} bpp {bytesPerElement} not supported by the GPU pass.");
+            return;
+        }
+
+        // Whole-block tiled source with a per-layer-distinct deterministic pattern
+        // (so a slice mix-up is caught), the array slices packed contiguously.
+        var blocksHigh = ((int)Height + parameters.BlockHeight - 1) / parameters.BlockHeight;
+        var sliceTiledBytes = (int)((long)parameters.BlocksPerRow * blocksHigh * parameters.BlockBytes);
+        var sliceLinearBytes = (int)(Width * Height * bytesPerElement);
+        var tiled = new byte[sliceTiledBytes * layers];
+        var expected = new byte[sliceLinearBytes * layers];
+        for (var layer = 0; layer < layers; layer++)
+        {
+            for (var index = 0; index < sliceTiledBytes; index++)
+            {
+                tiled[layer * sliceTiledBytes + index] = (byte)((index * 31 + 7 + layer * 101) & 0xFF);
+            }
+
+            if (!GnmTiling.TryDetile(
+                    tiled.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
+                    expected.AsSpan(layer * sliceLinearBytes, sliceLinearBytes),
+                    swizzleMode, (int)Width, (int)Height, bytesPerElement))
+            {
+                Console.Error.WriteLine("[DETILE-SELFTEST] FAIL: CPU TryDetile declined.");
+                return;
+            }
+        }
+
+        var tiledBytes = tiled;
+        var label = $"mode{swizzleMode} {bytesPerElement}bpp x{layers}";
+
+        // Phase 1: the one-shot DetileIntoImage (submit + wait in place).
+        VerifyPhase(
+            vk, device, physicalDevice, queue, commandPool, expected, layers, format, $"DetileIntoImage {label}",
+            image => pass.DetileIntoImage(image, ImageLayout.Undefined, Width, Height, layers, tiledBytes, parameters));
+
+        // Phase 2: RecordDetile — the exact code path the render loop uses
+        // (record into a command buffer, submit, retire the transients).
+        VerifyPhase(
+            vk, device, physicalDevice, queue, commandPool, expected, layers, format, $"RecordDetile {label}",
+            image => RecordDetileAndSubmit(vk, device, queue, commandPool, pass, image, layers, tiledBytes, parameters));
+    }
+
+    private static void VerifyPhase(
+        Vk vk,
+        Device device,
+        PhysicalDevice physicalDevice,
+        Queue queue,
+        CommandPool commandPool,
+        byte[] expected,
+        uint layers,
+        Format format,
+        string label,
+        Func<Image, bool> detile)
+    {
+        var image = CreateImage(vk, device, physicalDevice, format, layers, out var imageMemory);
+        var readback = CreateHostBuffer(
+            vk, device, physicalDevice, (ulong)expected.Length, BufferUsageFlags.TransferDstBit, out var readbackMemory);
+        try
+        {
+            if (!detile(image))
+            {
+                Console.Error.WriteLine($"[DETILE-SELFTEST] {label} FAIL: declined.");
+                return;
+            }
+
+            CopyImageToBuffer(vk, device, queue, commandPool, image, readback, layers);
+
+            void* mapped;
+            Check(
+                vk.MapMemory(device, readbackMemory, 0, (ulong)expected.Length, 0, &mapped),
+                $"vkMapMemory(selftest {label})");
+            var actual = new Span<byte>(mapped, expected.Length);
+            var firstMismatch = -1;
+            for (var index = 0; index < expected.Length; index++)
+            {
+                if (actual[index] != expected[index])
+                {
+                    firstMismatch = index;
+                    break;
+                }
+            }
+
+            vk.UnmapMemory(device, readbackMemory);
+
+            Console.Error.WriteLine(firstMismatch < 0
+                ? $"[DETILE-SELFTEST] {label} PASS: {Width}x{Height}x{layers} matches CPU detile ({expected.Length} bytes)."
+                : $"[DETILE-SELFTEST] {label} FAIL: first mismatch at byte {firstMismatch}.");
+        }
+        finally
+        {
+            if (readback.Handle != 0)
+            {
+                vk.DestroyBuffer(device, readback, null);
+            }
+
+            if (readbackMemory.Handle != 0)
+            {
+                vk.FreeMemory(device, readbackMemory, null);
+            }
+
+            if (image.Handle != 0)
+            {
+                vk.DestroyImage(device, image, null);
+            }
+
+            if (imageMemory.Handle != 0)
+            {
+                vk.FreeMemory(device, imageMemory, null);
+            }
+        }
+    }
+
+    // Records the detile into a fresh command buffer, submits, waits, and retires
+    // the transients exactly as the presenter's batch does — verifying the render
+    // path's code (RecordDetile) without needing a game to trigger it.
+    private static bool RecordDetileAndSubmit(
+        Vk vk,
+        Device device,
+        Queue queue,
+        CommandPool commandPool,
+        VulkanDetilePass pass,
+        Image image,
+        uint layers,
+        ReadOnlySpan<byte> tiled,
+        in DetileParams parameters)
+    {
+        var commandBuffer = AllocateCommandBuffer(vk, device, commandPool);
+        var beginInfo = new CommandBufferBeginInfo
+        {
+            SType = StructureType.CommandBufferBeginInfo,
+            Flags = CommandBufferUsageFlags.OneTimeSubmitBit,
+        };
+        Check(vk.BeginCommandBuffer(commandBuffer, &beginInfo), "vkBeginCommandBuffer(selftest record)");
+
+        if (!pass.RecordDetile(
+                commandBuffer, image, ImageLayout.Undefined, Width, Height, layers, tiled, parameters, out var transients))
+        {
+            _ = vk.EndCommandBuffer(commandBuffer);
+            vk.FreeCommandBuffers(device, commandPool, 1, &commandBuffer);
+            return false;
+        }
+
+        Check(vk.EndCommandBuffer(commandBuffer), "vkEndCommandBuffer(selftest record)");
+
+        var fenceInfo = new FenceCreateInfo { SType = StructureType.FenceCreateInfo };
+        Check(vk.CreateFence(device, &fenceInfo, null, out var fence), "vkCreateFence(selftest record)");
+        try
+        {
+            var submitInfo = new SubmitInfo
+            {
+                SType = StructureType.SubmitInfo,
+                CommandBufferCount = 1,
+                PCommandBuffers = &commandBuffer,
+            };
+            Check(vk.QueueSubmit(queue, 1, &submitInfo, fence), "vkQueueSubmit(selftest record)");
+            Check(vk.WaitForFences(device, 1, &fence, true, ulong.MaxValue), "vkWaitForFences(selftest record)");
+        }
+        finally
+        {
+            vk.DestroyFence(device, fence, null);
+            vk.FreeCommandBuffers(device, commandPool, 1, &commandBuffer);
+            pass.Retire(transients);
+        }
+
+        return true;
+    }
+
+    private static Image CreateImage(
+        Vk vk,
+        Device device,
+        PhysicalDevice physicalDevice,
+        Format format,
+        uint layers,
+        out DeviceMemory memory)
+    {
+        var imageInfo = new ImageCreateInfo
+        {
+            SType = StructureType.ImageCreateInfo,
+            ImageType = ImageType.Type2D,
+            Format = format,
+            Extent = new Extent3D(Width, Height, 1),
+            MipLevels = 1,
+            ArrayLayers = layers,
+            Samples = SampleCountFlags.Count1Bit,
+            Tiling = ImageTiling.Optimal,
+            Usage = ImageUsageFlags.TransferDstBit | ImageUsageFlags.TransferSrcBit,
+            SharingMode = SharingMode.Exclusive,
+            InitialLayout = ImageLayout.Undefined,
+        };
+        Check(vk.CreateImage(device, &imageInfo, null, out var image), "vkCreateImage(selftest)");
+
+        vk.GetImageMemoryRequirements(device, image, out var requirements);
+        var allocateInfo = new MemoryAllocateInfo
+        {
+            SType = StructureType.MemoryAllocateInfo,
+            AllocationSize = requirements.Size,
+            MemoryTypeIndex = FindMemoryType(
+                vk,
+                physicalDevice,
+                requirements.MemoryTypeBits,
+                MemoryPropertyFlags.DeviceLocalBit),
+        };
+        Check(vk.AllocateMemory(device, &allocateInfo, null, out memory), "vkAllocateMemory(selftest image)");
+        Check(vk.BindImageMemory(device, image, memory, 0), "vkBindImageMemory(selftest)");
+        return image;
+    }
+
+    private static void CopyImageToBuffer(
+        Vk vk,
+        Device device,
+        Queue queue,
+        CommandPool commandPool,
+        Image image,
+        VkBuffer destination,
+        uint layers)
+    {
+        var commandBuffer = AllocateCommandBuffer(vk, device, commandPool);
+        var beginInfo = new CommandBufferBeginInfo
+        {
+            SType = StructureType.CommandBufferBeginInfo,
+            Flags = CommandBufferUsageFlags.OneTimeSubmitBit,
+        };
+        Check(vk.BeginCommandBuffer(commandBuffer, &beginInfo), "vkBeginCommandBuffer(selftest readback)");
+
+        // DetileIntoImage left the image ShaderReadOnly; move it to TransferSrc.
+        var toTransferSrc = new ImageMemoryBarrier
+        {
+            SType = StructureType.ImageMemoryBarrier,
+            SrcAccessMask = AccessFlags.ShaderReadBit,
+            DstAccessMask = AccessFlags.TransferReadBit,
+            OldLayout = ImageLayout.ShaderReadOnlyOptimal,
+            NewLayout = ImageLayout.TransferSrcOptimal,
+            SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+            Image = image,
+            SubresourceRange = new ImageSubresourceRange(ImageAspectFlags.ColorBit, 0, 1, 0, layers),
+        };
+        vk.CmdPipelineBarrier(
+            commandBuffer,
+            PipelineStageFlags.FragmentShaderBit,
+            PipelineStageFlags.TransferBit,
+            0,
+            0,
+            null,
+            0,
+            null,
+            1,
+            &toTransferSrc);
+
+        // Layer-major readback: one copy pulls every array slice back into the
+        // buffer contiguously, matching the packed `expected` layout.
+        var region = new BufferImageCopy
+        {
+            BufferOffset = 0,
+            BufferRowLength = 0,
+            BufferImageHeight = 0,
+            ImageSubresource = new ImageSubresourceLayers(ImageAspectFlags.ColorBit, 0, 0, layers),
+            ImageOffset = default,
+            ImageExtent = new Extent3D(Width, Height, 1),
+        };
+        vk.CmdCopyImageToBuffer(
+            commandBuffer,
+            image,
+            ImageLayout.TransferSrcOptimal,
+            destination,
+            1,
+            &region);
+
+        Check(vk.EndCommandBuffer(commandBuffer), "vkEndCommandBuffer(selftest readback)");
+
+        var fenceInfo = new FenceCreateInfo { SType = StructureType.FenceCreateInfo };
+        Check(vk.CreateFence(device, &fenceInfo, null, out var fence), "vkCreateFence(selftest)");
+        try
+        {
+            var submitInfo = new SubmitInfo
+            {
+                SType = StructureType.SubmitInfo,
+                CommandBufferCount = 1,
+                PCommandBuffers = &commandBuffer,
+            };
+            Check(vk.QueueSubmit(queue, 1, &submitInfo, fence), "vkQueueSubmit(selftest readback)");
+            Check(
+                vk.WaitForFences(device, 1, &fence, true, ulong.MaxValue),
+                "vkWaitForFences(selftest readback)");
+        }
+        finally
+        {
+            vk.DestroyFence(device, fence, null);
+            vk.FreeCommandBuffers(device, commandPool, 1, &commandBuffer);
+        }
+    }
+
+    private static CommandPool CreateCommandPool(Vk vk, Device device, uint queueFamilyIndex)
+    {
+        var poolInfo = new CommandPoolCreateInfo
+        {
+            SType = StructureType.CommandPoolCreateInfo,
+            QueueFamilyIndex = queueFamilyIndex,
+            Flags = CommandPoolCreateFlags.ResetCommandBufferBit,
+        };
+        Check(vk.CreateCommandPool(device, &poolInfo, null, out var pool), "vkCreateCommandPool(selftest)");
+        return pool;
+    }
+
+    private static CommandBuffer AllocateCommandBuffer(Vk vk, Device device, CommandPool commandPool)
+    {
+        var allocateInfo = new CommandBufferAllocateInfo
+        {
+            SType = StructureType.CommandBufferAllocateInfo,
+            CommandPool = commandPool,
+            Level = CommandBufferLevel.Primary,
+            CommandBufferCount = 1,
+        };
+        Check(
+            vk.AllocateCommandBuffers(device, &allocateInfo, out var commandBuffer),
+            "vkAllocateCommandBuffers(selftest)");
+        return commandBuffer;
+    }
+
+    private static VkBuffer CreateHostBuffer(
+        Vk vk,
+        Device device,
+        PhysicalDevice physicalDevice,
+        ulong size,
+        BufferUsageFlags usage,
+        out DeviceMemory memory)
+    {
+        var bufferInfo = new BufferCreateInfo
+        {
+            SType = StructureType.BufferCreateInfo,
+            Size = size,
+            Usage = usage,
+            SharingMode = SharingMode.Exclusive,
+        };
+        Check(vk.CreateBuffer(device, &bufferInfo, null, out var buffer), "vkCreateBuffer(selftest)");
+
+        vk.GetBufferMemoryRequirements(device, buffer, out var requirements);
+        var allocateInfo = new MemoryAllocateInfo
+        {
+            SType = StructureType.MemoryAllocateInfo,
+            AllocationSize = requirements.Size,
+            MemoryTypeIndex = FindMemoryType(
+                vk,
+                physicalDevice,
+                requirements.MemoryTypeBits,
+                MemoryPropertyFlags.HostVisibleBit | MemoryPropertyFlags.HostCoherentBit),
+        };
+        Check(vk.AllocateMemory(device, &allocateInfo, null, out memory), "vkAllocateMemory(selftest buffer)");
+        Check(vk.BindBufferMemory(device, buffer, memory, 0), "vkBindBufferMemory(selftest)");
+        return buffer;
+    }
+
+    private static uint FindMemoryType(
+        Vk vk,
+        PhysicalDevice physicalDevice,
+        uint typeBits,
+        MemoryPropertyFlags requiredFlags)
+    {
+        vk.GetPhysicalDeviceMemoryProperties(physicalDevice, out var properties);
+        var memoryTypes = &properties.MemoryTypes.Element0;
+        for (uint index = 0; index < properties.MemoryTypeCount; index++)
+        {
+            if ((typeBits & (1u << (int)index)) != 0 &&
+                (memoryTypes[index].PropertyFlags & requiredFlags) == requiredFlags)
+            {
+                return index;
+            }
+        }
+
+        throw new InvalidOperationException("No compatible Vulkan memory type for the detile self-test.");
+    }
+
+    private static void Check(Result result, string operation)
+    {
+        if (result != Result.Success)
+        {
+            throw new InvalidOperationException($"{operation} failed: {result}");
+        }
+    }
+}

src/SharpEmu.ShaderCompiler.Metal/MslFixedShaders.cs

@@ -76,6 +76,14 @@ public static class MslFixedShaders
     /// </summary>
     public static string CreateDepthOnlyFragment() => MslTemplates.Render("depth_only_fragment");
 
+    /// <summary>
+    /// Compute kernel that deswizzles one RDNA2 exact-XOR tiled surface (swizzle
+    /// modes 5/9/24/27) at 4 bytes/element into a linear output buffer — the MSL
+    /// twin of <c>SpirvFixedShaders.CreateDetileCompute</c>. Entry point
+    /// "detile_cs"; buffers 0=tiled, 1=xTerm, 2=yTerm, 3=out, 4=DetileParams.
+    /// </summary>
+    public static string CreateDetileCompute() => MslTemplates.Render("detile_compute");
+
     private static string Format(float value) =>
         value.ToString("0.0######", CultureInfo.InvariantCulture) + "f";
 }

src/SharpEmu.ShaderCompiler.Metal/Templates/detile_compute.msl

@@ -0,0 +1,70 @@
+#include <metal_stdlib>
+
+using namespace metal;
+
+// GPU deswizzle for RDNA2 tiled surfaces at 4/8/16 bytes/element — the MSL twin of
+// SpirvFixedShaders.CreateDetileCompute and a direct mirror of
+// GnmTiling.GetDetileParams. Handles both supported equation families and both
+// plain 2D textures and array textures (one grid-Z layer per slice; the caller
+// packs the tiled slices contiguously). width/height are ELEMENT dims (for
+// block-compressed formats a 4x4 block is one element); each element spans
+// uintsPerElement = bpp/4 words, and the X grid is widened by that factor so each
+// thread copies one word (elemX = gidX / upe, word = gidX % upe):
+//   inBlock = equation == 1  // BlockTable, modes 1/4/8
+//             ? blockTable[(y % blockHeight) * blockWidth + (elemX % blockWidth)]
+//             : xTerm[elemX & xMask] ^ yTerm[y & yMask];   // ExactXor 5/9/24/27
+//   srcElem = z * srcSliceElements
+//           + (y / blockHeight * blocksPerRow + elemX / blockWidth) * blockElements
+//           + inBlock;
+//   dstElem = z * width * height + y * width + elemX;
+//   out[dstElem * upe + word] = tiled[srcElem * upe + word];
+// Term tables hold ELEMENT offsets. buffer(1) carries xTerm (ExactXor) OR the
+// block table (BlockTable); the two equations index different-sized buffers, so
+// exactly one branch runs. 1/2 bpp are sub-word and stay on the CPU.
+struct DetileParams
+{
+    uint width;
+    uint height;
+    uint blockWidth;
+    uint blockHeight;
+    uint blockElements;
+    uint blocksPerRow;
+    uint xMask;
+    uint yMask;
+    uint srcSliceElements;
+    uint equation;
+    uint uintsPerElement;
+};
+
+kernel void detile_cs(
+    device const uint* tiled        [[buffer(0)]],
+    device const uint* xTermOrTable [[buffer(1)]],
+    device const uint* yTerm        [[buffer(2)]],
+    device uint* outLinear          [[buffer(3)]],
+    constant DetileParams& params   [[buffer(4)]],
+    uint3 gid                       [[thread_position_in_grid]])
+{
+    uint elemX = gid.x / params.uintsPerElement;
+    uint word = gid.x - elemX * params.uintsPerElement;
+    if (elemX >= params.width || gid.y >= params.height)
+    {
+        return;
+    }
+
+    uint blockIndex = (gid.y / params.blockHeight) * params.blocksPerRow + (elemX / params.blockWidth);
+    uint off;
+    if (params.equation != 0u)
+    {
+        uint inX = elemX - (elemX / params.blockWidth) * params.blockWidth;
+        uint inY = gid.y - (gid.y / params.blockHeight) * params.blockHeight;
+        off = xTermOrTable[inY * params.blockWidth + inX];
+    }
+    else
+    {
+        off = xTermOrTable[elemX & params.xMask] ^ yTerm[gid.y & params.yMask];
+    }
+
+    uint srcElem = gid.z * params.srcSliceElements + blockIndex * params.blockElements + off;
+    uint dstElem = gid.z * params.width * params.height + gid.y * params.width + elemX;
+    outLinear[dstElem * params.uintsPerElement + word] = tiled[srcElem * params.uintsPerElement + word];
+}

src/SharpEmu.ShaderCompiler.Vulkan/Gen5SpirvTranslator.cs

@@ -314,7 +314,8 @@ public static partial class Gen5SpirvTranslator
             uint VectorType,
             ImageComponentKind ComponentKind,
             bool IsStorage,
-            bool Arrayed);
+            bool Arrayed,
+            SpirvImageDim Dimension);
 
         private readonly record struct SpirvVertexInput(
             uint Variable,
@@ -1001,11 +1002,15 @@ public static partial class Gen5SpirvTranslator
                         SpirvCapability.StorageImageExtendedFormats);
                 }
 
-                var isArrayed = !isStorage &&
+                var dimension = binding.Control.Dimension == 2
+                    ? SpirvImageDim.Dim3D
+                    : SpirvImageDim.Dim2D;
+                var isArrayed = dimension != SpirvImageDim.Dim3D &&
+                    !isStorage &&
                     Gen5ShaderTranslator.IsArrayedImageBinding(binding);
                 var imageType = _module.TypeImage(
                     componentType,
-                    SpirvImageDim.Dim2D,
+                    dimension,
                     depth: false,
                     arrayed: isArrayed,
                     multisampled: false,
@@ -1035,7 +1040,8 @@ public static partial class Gen5SpirvTranslator
                         _module.TypeVector(componentType, 4),
                         componentKind,
                         isStorage,
-                        isArrayed));
+                        isArrayed,
+                        dimension));
                 _interfaces.Add(variable);
             }
         }
@@ -2259,20 +2265,39 @@ public static partial class Gen5SpirvTranslator
                 return false;
             }
 
+            var memoryOpcode = control.UsesFlatAddress
+                ? "Global" + instruction.Opcode["Flat".Length..]
+                : instruction.Opcode;
+            var vectorByteAddress = LoadV(control.VectorAddress);
+            if (control.UsesFlatAddress)
+            {
+                // FLAT instructions carry the complete 64-bit guest address
+                // in a VGPR pair. The scalar evaluator captures the buffer
+                // rooted at the inferred SGPR pair, so convert the low address
+                // dword back to a byte offset inside that binding. Subtraction
+                // modulo 2^32 also handles an address addition that carried
+                // into the high dword, while the bounded binding prevents an
+                // unrelated pointer from escaping into host storage.
+                vectorByteAddress = _module.AddInstruction(
+                    SpirvOp.ISub,
+                    _uintType,
+                    vectorByteAddress,
+                    LoadS(control.ScalarAddress));
+            }
             var byteAddress = IAdd(
-                LoadV(control.VectorAddress),
+                vectorByteAddress,
                 UInt(unchecked((uint)control.OffsetBytes)));
             byteAddress = ApplyGuestBufferByteBias(bindingIndex, byteAddress);
             var dwordAddress = ShiftRightLogical(byteAddress, UInt(2));
 
-            if (instruction.Opcode is "GlobalAtomicAdd" or "GlobalAtomicUMax")
+            if (memoryOpcode is "GlobalAtomicAdd" or "GlobalAtomicUMax")
             {
                 EmitExecConditional(() =>
                 {
                     EmitConditional(IsBufferWordInRange(bindingIndex, dwordAddress), () =>
                     {
                         var original = _module.AddInstruction(
-                            instruction.Opcode == "GlobalAtomicAdd"
+                            memoryOpcode == "GlobalAtomicAdd"
                                 ? SpirvOp.AtomicIAdd
                                 : SpirvOp.AtomicUMax,
                             _uintType,
@@ -2289,12 +2314,12 @@ public static partial class Gen5SpirvTranslator
                 return true;
             }
 
-            if (instruction.Opcode.StartsWith("GlobalStore", StringComparison.Ordinal))
+            if (memoryOpcode.StartsWith("GlobalStore", StringComparison.Ordinal))
             {
                 EmitExecConditional(() =>
                 {
                     if (TryGetSubdwordStoreInfo(
-                            instruction.Opcode,
+                            memoryOpcode,
                             out var byteCount,
                             out var sourceShift))
                     {
@@ -2324,7 +2349,7 @@ public static partial class Gen5SpirvTranslator
             }
 
             if (TryGetSubdwordLoadInfo(
-                    instruction.Opcode,
+                    memoryOpcode,
                     out var loadByteCount,
                     out var signExtend,
                     out var d16,
@@ -3268,6 +3293,7 @@ public static partial class Gen5SpirvTranslator
             var imageObject = Load(resource.ObjectType, resource.Variable);
             if (instruction.Opcode == "ImageGetResinfo")
             {
+                var sizeComponentCount = ImageCoordinateComponentCount(resource);
                 var queryImage = resource.IsStorage
                     ? imageObject
                     : _module.AddInstruction(
@@ -3278,7 +3304,7 @@ public static partial class Gen5SpirvTranslator
                     resource.IsStorage
                         ? SpirvOp.ImageQuerySize
                         : SpirvOp.ImageQuerySizeLod,
-                    _module.TypeVector(_intType, 2),
+                    _module.TypeVector(_intType, sizeComponentCount),
                     resource.IsStorage
                         ? [queryImage]
                         : [queryImage, UInt(0)]);
@@ -3291,7 +3317,7 @@ public static partial class Gen5SpirvTranslator
                     }
 
                     uint value;
-                    if (component < 2)
+                    if (component < sizeComponentCount)
                     {
                         var signedValue = _module.AddInstruction(
                             SpirvOp.CompositeExtract,
@@ -3319,7 +3345,12 @@ public static partial class Gen5SpirvTranslator
                     return false;
                 }
 
-                var coordinates = BuildIntegerCoordinates(image, 0);
+                var coordinateComponentCount =
+                    ImageCoordinateComponentCount(resource);
+                var coordinates = BuildIntegerCoordinates(
+                    image,
+                    0,
+                    coordinateComponentCount);
                 var components = new uint[4];
                 uint sourceIndex = 0;
                 for (var component = 0; component < components.Length; component++)
@@ -3355,13 +3386,14 @@ public static partial class Gen5SpirvTranslator
                     components);
                 var imageSize = _module.AddInstruction(
                     SpirvOp.ImageQuerySize,
-                    _module.TypeVector(_intType, 2),
+                    _module.TypeVector(_intType, coordinateComponentCount),
                     imageObject);
                 EmitBoundsCheckedImageWrite(
                     coordinates,
                     imageSize,
                     imageObject,
-                    texel);
+                    texel,
+                    coordinateComponentCount);
 
                 return true;
             }
@@ -3382,14 +3414,17 @@ public static partial class Gen5SpirvTranslator
                 }
 
                 var signed = resource.ComponentKind == ImageComponentKind.Sint;
+                var coordinateComponentCount =
+                    ImageCoordinateComponentCount(resource);
                 var atomicImageSize = _module.AddInstruction(
                     SpirvOp.ImageQuerySize,
-                    _module.TypeVector(_intType, 2),
+                    _module.TypeVector(_intType, coordinateComponentCount),
                     imageObject);
                 var coordinates = BuildClampedIntegerCoordinates(
                     image,
                     0,
-                    atomicImageSize);
+                    atomicImageSize,
+                    coordinateComponentCount);
                 EmitExecConditional(() =>
                 {
                     var pointer = _module.AddInstruction(
@@ -3433,14 +3468,17 @@ public static partial class Gen5SpirvTranslator
             {
                 if (resource.IsStorage)
                 {
+                    var coordinateComponentCount =
+                        ImageCoordinateComponentCount(resource);
                     var imageSize = _module.AddInstruction(
                         SpirvOp.ImageQuerySize,
-                        _module.TypeVector(_intType, 2),
+                        _module.TypeVector(_intType, coordinateComponentCount),
                         imageObject);
                     var coordinates = BuildClampedIntegerCoordinates(
                         image,
                         0,
-                        imageSize);
+                        imageSize,
+                        coordinateComponentCount);
                     sampled = _module.AddInstruction(
                         SpirvOp.ImageRead,
                         resource.VectorType,
@@ -3454,15 +3492,18 @@ public static partial class Gen5SpirvTranslator
                         SpirvOp.Image,
                         resource.ImageType,
                         imageObject);
+                    var coordinateComponentCount =
+                        ImageCoordinateComponentCount(resource);
                     var imageSize = _module.AddInstruction(
                         SpirvOp.ImageQuerySizeLod,
-                        _module.TypeVector(_intType, 2),
+                        _module.TypeVector(_intType, coordinateComponentCount),
                         fetchedImage,
                         UInt(mipLevel));
                     var coordinates = BuildClampedIntegerCoordinates(
                         image,
                         0,
-                        imageSize);
+                        imageSize,
+                        coordinateComponentCount);
                     sampled = _module.AddInstruction(
                         SpirvOp.ImageFetch,
                         resource.VectorType,
@@ -3494,12 +3535,19 @@ public static partial class Gen5SpirvTranslator
                 // lowering treated SAMPLE_D as body-first and consequently
                 // sampled gradients as coordinates in every captured
                 // derivative operation.
+                var spatialComponentCount =
+                    ImageSpatialComponentCount(resource);
+                var coordinateComponentCount =
+                    ImageCoordinateComponentCount(resource);
                 var addressCursor = 0;
                 var offset = 0u;
                 if (hasOffset)
                 {
                     addressCursor = AlignFullImageAddress(image, addressCursor);
-                    offset = BuildImageOffset(image, addressCursor);
+                    offset = BuildImageOffset(
+                        image,
+                        addressCursor,
+                        spatialComponentCount);
                     addressCursor += ImageFullAddressSlots(image);
                 }
 
@@ -3523,26 +3571,33 @@ public static partial class Gen5SpirvTranslator
                 }
 
                 var gradientX = hasGradients
-                    ? BuildFloatCoordinates(image, addressCursor)
+                    ? BuildFloatCoordinates(
+                        image,
+                        addressCursor,
+                        spatialComponentCount)
                     : 0u;
                 var gradientY = hasGradients
-                    ? BuildFloatCoordinates(image, addressCursor + 2)
+                    ? BuildFloatCoordinates(
+                        image,
+                        addressCursor + (int)spatialComponentCount,
+                        spatialComponentCount)
                     : 0u;
                 if (hasGradients)
                 {
-                    addressCursor += 4;
+                    addressCursor += checked((int)(spatialComponentCount * 2));
                 }
 
-                var coordinates = resource.Arrayed
-                    ? BuildFloatArrayCoordinates(image, addressCursor)
-                    : BuildFloatCoordinates(image, addressCursor);
+                var coordinates = BuildFloatCoordinates(
+                    image,
+                    addressCursor,
+                    coordinateComponentCount);
                 var explicitLod = hasGradients || hasZeroLod || hasLod;
                 var lod = hasZeroLod
                     ? Float(0)
                     : hasLod
                         ? LoadImageFloatAddress(
                             image,
-                            addressCursor + (resource.Arrayed ? 3 : 2))
+                            addressCursor + (int)coordinateComponentCount)
                         : lodOrBias;
                 if (hasOffset)
                 {
@@ -3608,11 +3663,18 @@ public static partial class Gen5SpirvTranslator
                     instruction.Opcode.EndsWith("O", StringComparison.Ordinal);
                 var hasCompare =
                     instruction.Opcode.Contains("Gather4C", StringComparison.Ordinal);
+                var spatialComponentCount =
+                    ImageSpatialComponentCount(resource);
+                var coordinateComponentCount =
+                    ImageCoordinateComponentCount(resource);
                 var addressCursor = 0;
                 var offset = 0u;
                 if (hasOffset)
                 {
-                    offset = BuildImageOffset(image, addressCursor);
+                    offset = BuildImageOffset(
+                        image,
+                        addressCursor,
+                        spatialComponentCount);
                     addressCursor += ImageFullAddressSlots(image);
                 }
 
@@ -3627,9 +3689,10 @@ public static partial class Gen5SpirvTranslator
                     addressCursor += ImageFullAddressSlots(image);
                 }
 
-                var coordinates = resource.Arrayed
-                    ? BuildFloatArrayCoordinates(image, addressCursor)
-                    : BuildFloatCoordinates(image, addressCursor);
+                var coordinates = BuildFloatCoordinates(
+                    image,
+                    addressCursor,
+                    coordinateComponentCount);
                 var operands = new List<uint>
                 {
                     imageObject,
@@ -3822,28 +3885,31 @@ public static partial class Gen5SpirvTranslator
                 });
         }
 
-        private uint BuildFloatCoordinates(Gen5ImageControl image, int start)
-        {
-            var x = LoadImageFloatAddress(image, start);
-            var y = LoadImageFloatAddress(image, start + 1);
-            return _module.AddInstruction(
-                SpirvOp.CompositeConstruct,
-                _vec2Type,
-                x,
-                y);
-        }
+        private static uint ImageSpatialComponentCount(
+            SpirvImageResource resource) =>
+            resource.Dimension == SpirvImageDim.Dim3D ? 3u : 2u;
 
-        private uint BuildFloatArrayCoordinates(Gen5ImageControl image, int start)
+        private static uint ImageCoordinateComponentCount(
+            SpirvImageResource resource) =>
+            resource.Arrayed ? 3u : ImageSpatialComponentCount(resource);
+
+        private uint BuildFloatCoordinates(
+            Gen5ImageControl image,
+            int start,
+            uint componentCount)
         {
-            var x = LoadImageFloatAddress(image, start);
-            var y = LoadImageFloatAddress(image, start + 1);
-            var slice = LoadImageFloatAddress(image, start + 2);
+            var components = new uint[checked((int)componentCount)];
+            for (var component = 0; component < components.Length; component++)
+            {
+                components[component] = LoadImageFloatAddress(
+                    image,
+                    start + component);
+            }
+
             return _module.AddInstruction(
                 SpirvOp.CompositeConstruct,
-                _vec3Type,
-                x,
-                y,
-                slice);
+                _module.TypeVector(_floatType, componentCount),
+                components);
         }
 
         private static int ImageAddressRegister(
@@ -3948,47 +4014,49 @@ public static partial class Gen5SpirvTranslator
                 ShiftLeftLogical(BitwiseAnd(high, UInt(0xFFFF)), UInt(16)));
         }
 
-        private uint BuildIntegerCoordinates(Gen5ImageControl image, int start)
+        private uint BuildIntegerCoordinates(
+            Gen5ImageControl image,
+            int start,
+            uint componentCount)
         {
-            var ivec2 = _module.TypeVector(_intType, 2);
-            var x = Bitcast(_intType, LoadImageIntegerAddress(image, start));
-            var y = Bitcast(_intType, LoadImageIntegerAddress(image, start + 1));
+            var components = new uint[checked((int)componentCount)];
+            for (var component = 0; component < components.Length; component++)
+            {
+                components[component] = Bitcast(
+                    _intType,
+                    LoadImageIntegerAddress(image, start + component));
+            }
+
             return _module.AddInstruction(
                 SpirvOp.CompositeConstruct,
-                ivec2,
-                x,
-                y);
+                _module.TypeVector(_intType, componentCount),
+                components);
         }
 
         private uint BuildClampedIntegerCoordinates(
             Gen5ImageControl image,
             int start,
-            uint imageSize)
+            uint imageSize,
+            uint componentCount)
         {
-            var ivec2 = _module.TypeVector(_intType, 2);
-            var x = ClampSignedCoordinate(
-                Bitcast(
-                    _intType,
-                    LoadImageIntegerAddress(image, start)),
-                _module.AddInstruction(
-                    SpirvOp.CompositeExtract,
-                    _intType,
-                    imageSize,
-                    0));
-            var y = ClampSignedCoordinate(
-                Bitcast(
-                    _intType,
-                    LoadImageIntegerAddress(image, start + 1)),
-                _module.AddInstruction(
-                    SpirvOp.CompositeExtract,
-                    _intType,
-                    imageSize,
-                    1));
+            var components = new uint[checked((int)componentCount)];
+            for (var component = 0; component < components.Length; component++)
+            {
+                components[component] = ClampSignedCoordinate(
+                    Bitcast(
+                        _intType,
+                        LoadImageIntegerAddress(image, start + component)),
+                    _module.AddInstruction(
+                        SpirvOp.CompositeExtract,
+                        _intType,
+                        imageSize,
+                        (uint)component));
+            }
+
             return _module.AddInstruction(
                 SpirvOp.CompositeConstruct,
-                ivec2,
-                x,
-                y);
+                _module.TypeVector(_intType, componentCount),
+                components);
         }
 
         private uint ClampSignedCoordinate(uint value, uint extent)
@@ -4027,69 +4095,47 @@ public static partial class Gen5SpirvTranslator
             uint coordinates,
             uint imageSize,
             uint imageObject,
-            uint texel)
+            uint texel,
+            uint coordinateComponentCount)
         {
-            var x = _module.AddInstruction(
-                SpirvOp.CompositeExtract,
-                _intType,
-                coordinates,
-                0);
-            var y = _module.AddInstruction(
-                SpirvOp.CompositeExtract,
-                _intType,
-                coordinates,
-                1);
-            var width = _module.AddInstruction(
-                SpirvOp.CompositeExtract,
-                _intType,
-                imageSize,
-                0);
-            var height = _module.AddInstruction(
-                SpirvOp.CompositeExtract,
-                _intType,
-                imageSize,
-                1);
             var zero = _module.Constant(_intType, 0);
-            var xNonNegative = _module.AddInstruction(
-                SpirvOp.SGreaterThanEqual,
-                _boolType,
-                x,
-                zero);
-            var yNonNegative = _module.AddInstruction(
-                SpirvOp.SGreaterThanEqual,
-                _boolType,
-                y,
-                zero);
-            var xInRange = _module.AddInstruction(
-                SpirvOp.SLessThan,
-                _boolType,
-                x,
-                width);
-            var yInRange = _module.AddInstruction(
-                SpirvOp.SLessThan,
-                _boolType,
-                y,
-                height);
-            var lowerInRange = _module.AddInstruction(
-                SpirvOp.LogicalAnd,
-                _boolType,
-                xNonNegative,
-                yNonNegative);
-            var upperInRange = _module.AddInstruction(
-                SpirvOp.LogicalAnd,
-                _boolType,
-                xInRange,
-                yInRange);
-            var inRange = _module.AddInstruction(
-                SpirvOp.LogicalAnd,
-                _boolType,
-                lowerInRange,
-                upperInRange);
-            inRange = _module.AddInstruction(
-                SpirvOp.LogicalAnd,
-                _boolType,
-                Load(_boolType, _exec),
-                inRange);
+            var inRange = Load(_boolType, _exec);
+            for (uint component = 0;
+                 component < coordinateComponentCount;
+                 component++)
+            {
+                var coordinate = _module.AddInstruction(
+                    SpirvOp.CompositeExtract,
+                    _intType,
+                    coordinates,
+                    component);
+                var extent = _module.AddInstruction(
+                    SpirvOp.CompositeExtract,
+                    _intType,
+                    imageSize,
+                    component);
+                var nonNegative = _module.AddInstruction(
+                    SpirvOp.SGreaterThanEqual,
+                    _boolType,
+                    coordinate,
+                    zero);
+                var belowExtent = _module.AddInstruction(
+                    SpirvOp.SLessThan,
+                    _boolType,
+                    coordinate,
+                    extent);
+                var componentInRange = _module.AddInstruction(
+                    SpirvOp.LogicalAnd,
+                    _boolType,
+                    nonNegative,
+                    belowExtent);
+                inRange = _module.AddInstruction(
+                    SpirvOp.LogicalAnd,
+                    _boolType,
+                    inRange,
+                    componentInRange);
+            }
+
             var writeLabel = _module.AllocateId();
             var mergeLabel = _module.AllocateId();
             _module.AddStatement(SpirvOp.SelectionMerge, mergeLabel, 0);
@@ -4108,30 +4154,30 @@ public static partial class Gen5SpirvTranslator
             _module.AddLabel(mergeLabel);
         }
 
-        private uint BuildImageOffset(Gen5ImageControl image, int component)
+        private uint BuildImageOffset(
+            Gen5ImageControl image,
+            int component,
+            uint componentCount)
         {
-            var ivec2 = _module.TypeVector(_intType, 2);
             var packed = Bitcast(
                 _intType,
                 LoadV(image.GetAddressRegister(
                     ImageAddressRegister(image, component))));
-            var x = _module.AddInstruction(
-                SpirvOp.BitFieldSExtract,
-                _intType,
-                packed,
-                UInt(0),
-                UInt(6));
-            var y = _module.AddInstruction(
-                SpirvOp.BitFieldSExtract,
-                _intType,
-                packed,
-                UInt(8),
-                UInt(6));
+            var components = new uint[checked((int)componentCount)];
+            for (var index = 0; index < components.Length; index++)
+            {
+                components[index] = _module.AddInstruction(
+                    SpirvOp.BitFieldSExtract,
+                    _intType,
+                    packed,
+                    UInt((uint)(index * 8)),
+                    UInt(6));
+            }
+
             return _module.AddInstruction(
                 SpirvOp.CompositeConstruct,
-                ivec2,
-                x,
-                y);
+                _module.TypeVector(_intType, componentCount),
+                components);
         }
 
         private uint ApplyDynamicSampleOffset(
@@ -4141,7 +4187,16 @@ public static partial class Gen5SpirvTranslator
             uint texelOffset,
             uint lod)
         {
-            var ivec2 = _module.TypeVector(_intType, 2);
+            var spatialComponentCount = ImageSpatialComponentCount(resource);
+            var coordinateComponentCount =
+                ImageCoordinateComponentCount(resource);
+            var spatialIntegerType =
+                _module.TypeVector(_intType, spatialComponentCount);
+            var spatialFloatType =
+                _module.TypeVector(_floatType, spatialComponentCount);
+            var queryComponentCount = resource.Arrayed
+                ? coordinateComponentCount
+                : spatialComponentCount;
             var image = _module.AddInstruction(
                 SpirvOp.Image,
                 resource.ImageType,
@@ -4163,14 +4218,14 @@ public static partial class Gen5SpirvTranslator
                 signedLod);
             var size = _module.AddInstruction(
                 SpirvOp.ImageQuerySizeLod,
-                resource.Arrayed ? _module.TypeVector(_intType, 3) : ivec2,
+                _module.TypeVector(_intType, queryComponentCount),
                 image,
                 clampedLod);
             if (resource.Arrayed)
             {
                 size = _module.AddInstruction(
                     SpirvOp.VectorShuffle,
-                    ivec2,
+                    spatialIntegerType,
                     size,
                     size,
                     0u,
@@ -4179,27 +4234,27 @@ public static partial class Gen5SpirvTranslator
 
             var sizeFloat = _module.AddInstruction(
                 SpirvOp.ConvertSToF,
-                _vec2Type,
+                spatialFloatType,
                 size);
             var offsetFloat = _module.AddInstruction(
                 SpirvOp.ConvertSToF,
-                _vec2Type,
+                spatialFloatType,
                 texelOffset);
             var normalizedOffset = _module.AddInstruction(
                 SpirvOp.FDiv,
-                _vec2Type,
+                spatialFloatType,
                 offsetFloat,
                 sizeFloat);
             if (!resource.Arrayed)
             {
                 return _module.AddInstruction(
                     SpirvOp.FAdd,
-                    _vec2Type,
+                    spatialFloatType,
                     coordinates,
                     normalizedOffset);
             }
 
-            var offsetVec3 = _module.AddInstruction(
+            var arrayOffset = _module.AddInstruction(
                 SpirvOp.CompositeConstruct,
                 _vec3Type,
                 _module.AddInstruction(
@@ -4217,7 +4272,7 @@ public static partial class Gen5SpirvTranslator
                 SpirvOp.FAdd,
                 _vec3Type,
                 coordinates,
-                offsetVec3);
+                arrayOffset);
         }
 
         private bool TryEmitExport(

src/SharpEmu.ShaderCompiler.Vulkan/SpirvFixedShaders.cs

@@ -260,4 +260,233 @@ public static class SpirvFixedShaders
         module.AddExecutionMode(main, SpirvExecutionMode.OriginUpperLeft);
         return module.Build();
     }
+
+    /// <summary>
+    /// Compute kernel that deswizzles RDNA2 tiled surfaces at 4 bytes/element into
+    /// a linear output buffer — one GPU thread per texel, one dispatch-Z layer per
+    /// array slice. Mirrors <c>GnmTiling.GetDetileParams</c> so it is bit-identical
+    /// to the CPU fallback for both supported equation families:
+    /// <code>
+    ///   z = layer;
+    ///   inBlock = equation == BlockTable            // modes 1/4/8
+    ///             ? blockTable[(y % blockHeight) * blockWidth + (x % blockWidth)]
+    ///             : xTerm[x &amp; xMask] ^ yTerm[y &amp; yMask];   // ExactXor 5/9/24/27
+    ///   src = z * srcSliceElements
+    ///         + (y / blockHeight * blocksPerRow + x / blockWidth) * blockElements
+    ///         + inBlock;
+    ///   out[z * width * height + y * width + x] = tiled[src];
+    /// </code>
+    /// Each array slice is an independently tiled 2D surface; the caller packs the
+    /// slices contiguously in the tiled buffer (stride <c>srcSliceElements</c>) and
+    /// the output ends up layer-major, matching a single multi-layer
+    /// buffer-&gt;image copy. For a non-arrayed texture the caller dispatches a
+    /// single Z layer with <c>srcSliceElements</c> unused (z == 0).
+    ///
+    /// The term tables hold ELEMENT offsets. For ExactXor the caller pre-shifts the
+    /// byte-unit GetDetileParams terms right by log2(bytesPerElement) (exact at 4bpp
+    /// since the equation's low two byte-offset bits are 0); for BlockTable the
+    /// GetDetileParams block table is already in element units. Binding 1 carries
+    /// xTerm (ExactXor) OR blockTable (BlockTable) — the two equations index
+    /// different-sized buffers, so the kernel branches and evaluates exactly one.
+    ///
+    /// width/height are ELEMENT dims (for block-compressed formats a 4x4 block is
+    /// one element). Each element spans uintsPerElement = bpp/4 words (4bpp -> 1,
+    /// 8bpp -> 2, 16bpp -> 4); the X dispatch is widened by that factor so each
+    /// thread copies one word (elemX = gidX / upe, word = gidX % upe). 1/2 bpp are
+    /// sub-word and stay on the CPU.
+    ///
+    /// Descriptor set 0: binding 0 = tiled uint[], 1 = xTerm/blockTable uint[],
+    /// 2 = yTerm uint[], 3 = out uint[]. Push constants (11 x uint, offset i*4):
+    /// width, height, blockWidth, blockHeight, blockElements, blocksPerRow,
+    /// xMask, yMask, srcSliceElements, equation (0 = ExactXor, 1 = BlockTable),
+    /// uintsPerElement. Local size 8x8x1; dispatch X = ceil(width*upe/8),
+    /// Y = ceil(height/8), Z = arrayLayers.
+    /// </summary>
+    public static byte[] CreateDetileCompute()
+    {
+        var module = new SpirvModuleBuilder();
+        module.AddCapability(SpirvCapability.Shader);
+
+        var voidType = module.TypeVoid();
+        var boolType = module.TypeBool();
+        var uintType = module.TypeInt(32, signed: false);
+        var uvec3Type = module.TypeVector(uintType, 3);
+
+        // One shared Block-decorated storage-buffer struct: struct { uint data[]; }.
+        var runtimeArray = module.TypeRuntimeArray(uintType);
+        module.AddDecoration(runtimeArray, SpirvDecoration.ArrayStride, 4);
+        var bufferStruct = module.TypeStruct(runtimeArray);
+        module.AddDecoration(bufferStruct, SpirvDecoration.Block);
+        module.AddMemberDecoration(bufferStruct, 0, SpirvDecoration.Offset, 0);
+        var bufferPtrType = module.TypePointer(SpirvStorageClass.StorageBuffer, bufferStruct);
+        var uintStoragePtr = module.TypePointer(SpirvStorageClass.StorageBuffer, uintType);
+
+        uint MakeBuffer(uint binding, string name)
+        {
+            var variable = module.AddGlobalVariable(bufferPtrType, SpirvStorageClass.StorageBuffer);
+            module.AddName(variable, name);
+            module.AddDecoration(variable, SpirvDecoration.DescriptorSet, 0);
+            module.AddDecoration(variable, SpirvDecoration.Binding, binding);
+            return variable;
+        }
+
+        var tiledVar = MakeBuffer(0, "tiled");
+        var xTermVar = MakeBuffer(1, "xTerm");
+        var yTermVar = MakeBuffer(2, "yTerm");
+        var outVar = MakeBuffer(3, "outLinear");
+
+        // Push constants: struct { uint p0..p10; }, each member at offset i*4.
+        var pushStruct = module.TypeStruct(
+            uintType, uintType, uintType, uintType, uintType, uintType,
+            uintType, uintType, uintType, uintType, uintType);
+        module.AddDecoration(pushStruct, SpirvDecoration.Block);
+        for (uint member = 0; member < 11; member++)
+        {
+            module.AddMemberDecoration(pushStruct, member, SpirvDecoration.Offset, member * 4);
+        }
+
+        var pushPtrType = module.TypePointer(SpirvStorageClass.PushConstant, pushStruct);
+        var pushMemberPtrType = module.TypePointer(SpirvStorageClass.PushConstant, uintType);
+        var pushVar = module.AddGlobalVariable(pushPtrType, SpirvStorageClass.PushConstant);
+        module.AddName(pushVar, "pc");
+
+        var inputUvec3Ptr = module.TypePointer(SpirvStorageClass.Input, uvec3Type);
+        var gidVar = module.AddGlobalVariable(inputUvec3Ptr, SpirvStorageClass.Input);
+        module.AddName(gidVar, "gid");
+        module.AddDecoration(gidVar, SpirvDecoration.BuiltIn, (uint)SpirvBuiltIn.GlobalInvocationId);
+
+        var uintConst = new uint[11];
+        for (uint value = 0; value < 11; value++)
+        {
+            uintConst[value] = module.Constant(uintType, value);
+        }
+
+        var functionType = module.TypeFunction(voidType);
+        var main = module.BeginFunction(voidType, functionType);
+        module.AddName(main, "main");
+        module.AddLabel();
+
+        var gid = module.AddInstruction(SpirvOp.Load, uvec3Type, gidVar);
+        var gidX = module.AddInstruction(SpirvOp.CompositeExtract, uintType, gid, 0);
+        var y = module.AddInstruction(SpirvOp.CompositeExtract, uintType, gid, 1);
+        var z = module.AddInstruction(SpirvOp.CompositeExtract, uintType, gid, 2);
+
+        uint PushField(uint index)
+        {
+            var pointer = module.AddInstruction(
+                SpirvOp.AccessChain, pushMemberPtrType, pushVar, uintConst[index]);
+            return module.AddInstruction(SpirvOp.Load, uintType, pointer);
+        }
+
+        // width/height are ELEMENT dims (for BC, a 4x4 block is one element). Each
+        // element spans uintsPerElement 32-bit words (bpp/4: 4bpp->1, 8bpp->2,
+        // 16bpp->4). The X dispatch is widened by uintsPerElement so each thread
+        // copies exactly one word: elemX = gidX / upe, wordIndex = gidX % upe.
+        var width = PushField(0);
+        var height = PushField(1);
+        var blockWidth = PushField(2);
+        var blockHeight = PushField(3);
+        var blockElements = PushField(4);
+        var blocksPerRow = PushField(5);
+        var xMask = PushField(6);
+        var yMask = PushField(7);
+        var srcSliceElements = PushField(8);
+        var equation = PushField(9);
+        var uintsPerElement = PushField(10);
+
+        var elemX = module.AddInstruction(SpirvOp.UDiv, uintType, gidX, uintsPerElement);
+        var elemXTimesUpe = module.AddInstruction(SpirvOp.IMul, uintType, elemX, uintsPerElement);
+        var wordIndex = module.AddInstruction(SpirvOp.ISub, uintType, gidX, elemXTimesUpe);
+
+        var xInRange = module.AddInstruction(SpirvOp.ULessThan, boolType, elemX, width);
+        var yInRange = module.AddInstruction(SpirvOp.ULessThan, boolType, y, height);
+        var inRange = module.AddInstruction(SpirvOp.LogicalAnd, boolType, xInRange, yInRange);
+
+        var bodyLabel = module.AllocateId();
+        var mergeLabel = module.AllocateId();
+        module.AddStatement(SpirvOp.SelectionMerge, mergeLabel, 0);
+        module.AddStatement(SpirvOp.BranchConditional, inRange, bodyLabel, mergeLabel);
+
+        module.AddLabel(bodyLabel);
+
+        // blockIdx = (y / blockHeight) * blocksPerRow + (elemX / blockWidth)
+        var yDiv = module.AddInstruction(SpirvOp.UDiv, uintType, y, blockHeight);
+        var blockRow = module.AddInstruction(SpirvOp.IMul, uintType, yDiv, blocksPerRow);
+        var xDiv = module.AddInstruction(SpirvOp.UDiv, uintType, elemX, blockWidth);
+        var blockIdx = module.AddInstruction(SpirvOp.IAdd, uintType, blockRow, xDiv);
+
+        // off (element offset within the block) = equation == BlockTable
+        //   ? blockTable[(y % blockHeight) * blockWidth + (elemX % blockWidth)]
+        //   : xTerm[elemX & xMask] ^ yTerm[y & yMask]
+        // Binding 1 (xTermVar) doubles as the block table; the two equations index
+        // different-sized buffers, so exactly one branch executes (no OOB read).
+        var isBlockTable = module.AddInstruction(SpirvOp.INotEqual, boolType, equation, uintConst[0]);
+        var xorLabel = module.AllocateId();
+        var tableLabel = module.AllocateId();
+        var offMergeLabel = module.AllocateId();
+        module.AddStatement(SpirvOp.SelectionMerge, offMergeLabel, 0);
+        module.AddStatement(SpirvOp.BranchConditional, isBlockTable, tableLabel, xorLabel);
+
+        // ExactXor: xTerm[elemX & xMask] ^ yTerm[y & yMask]
+        module.AddLabel(xorLabel);
+        var xIdx = module.AddInstruction(SpirvOp.BitwiseAnd, uintType, elemX, xMask);
+        var xPtr = module.AddInstruction(SpirvOp.AccessChain, uintStoragePtr, xTermVar, uintConst[0], xIdx);
+        var xTerm = module.AddInstruction(SpirvOp.Load, uintType, xPtr);
+        var yIdx = module.AddInstruction(SpirvOp.BitwiseAnd, uintType, y, yMask);
+        var yPtr = module.AddInstruction(SpirvOp.AccessChain, uintStoragePtr, yTermVar, uintConst[0], yIdx);
+        var yTerm = module.AddInstruction(SpirvOp.Load, uintType, yPtr);
+        var offXor = module.AddInstruction(SpirvOp.BitwiseXor, uintType, xTerm, yTerm);
+        module.AddStatement(SpirvOp.Branch, offMergeLabel);
+
+        // BlockTable: blockTable[inY * blockWidth + inX], inX/inY = position in block
+        module.AddLabel(tableLabel);
+        var blockXBase = module.AddInstruction(SpirvOp.IMul, uintType, xDiv, blockWidth);
+        var inX = module.AddInstruction(SpirvOp.ISub, uintType, elemX, blockXBase);
+        var blockYBase = module.AddInstruction(SpirvOp.IMul, uintType, yDiv, blockHeight);
+        var inY = module.AddInstruction(SpirvOp.ISub, uintType, y, blockYBase);
+        var rowInBlock = module.AddInstruction(SpirvOp.IMul, uintType, inY, blockWidth);
+        var tableIdx = module.AddInstruction(SpirvOp.IAdd, uintType, rowInBlock, inX);
+        var tablePtr = module.AddInstruction(SpirvOp.AccessChain, uintStoragePtr, xTermVar, uintConst[0], tableIdx);
+        var offTable = module.AddInstruction(SpirvOp.Load, uintType, tablePtr);
+        module.AddStatement(SpirvOp.Branch, offMergeLabel);
+
+        module.AddLabel(offMergeLabel);
+        var off = module.AddInstruction(SpirvOp.Phi, uintType, offXor, xorLabel, offTable, tableLabel);
+
+        // srcElem = z * srcSliceElements + blockIdx * blockElements + off  (in elements)
+        // srcWord = srcElem * uintsPerElement + wordIndex
+        var srcSliceBase = module.AddInstruction(SpirvOp.IMul, uintType, z, srcSliceElements);
+        var blockBase = module.AddInstruction(SpirvOp.IMul, uintType, blockIdx, blockElements);
+        var srcInSlice = module.AddInstruction(SpirvOp.IAdd, uintType, blockBase, off);
+        var srcElem = module.AddInstruction(SpirvOp.IAdd, uintType, srcSliceBase, srcInSlice);
+        var srcElemWords = module.AddInstruction(SpirvOp.IMul, uintType, srcElem, uintsPerElement);
+        var src = module.AddInstruction(SpirvOp.IAdd, uintType, srcElemWords, wordIndex);
+        var srcPtr = module.AddInstruction(SpirvOp.AccessChain, uintStoragePtr, tiledVar, uintConst[0], src);
+        var word = module.AddInstruction(SpirvOp.Load, uintType, srcPtr);
+
+        // dstElem = z * width * height + y * width + elemX  (in elements)
+        // dstWord = dstElem * uintsPerElement + wordIndex
+        var sliceElements = module.AddInstruction(SpirvOp.IMul, uintType, width, height);
+        var dstSliceBase = module.AddInstruction(SpirvOp.IMul, uintType, z, sliceElements);
+        var rowBase = module.AddInstruction(SpirvOp.IMul, uintType, y, width);
+        var dstRow = module.AddInstruction(SpirvOp.IAdd, uintType, rowBase, elemX);
+        var dstElem = module.AddInstruction(SpirvOp.IAdd, uintType, dstSliceBase, dstRow);
+        var dstElemWords = module.AddInstruction(SpirvOp.IMul, uintType, dstElem, uintsPerElement);
+        var dstIdx = module.AddInstruction(SpirvOp.IAdd, uintType, dstElemWords, wordIndex);
+        var dstPtr = module.AddInstruction(SpirvOp.AccessChain, uintStoragePtr, outVar, uintConst[0], dstIdx);
+        module.AddStatement(SpirvOp.Store, dstPtr, word);
+
+        module.AddStatement(SpirvOp.Branch, mergeLabel);
+        module.AddLabel(mergeLabel);
+        module.AddStatement(SpirvOp.Return);
+        module.EndFunction();
+
+        module.AddExecutionMode(main, SpirvExecutionMode.LocalSize, 8, 8, 1);
+        module.AddEntryPoint(
+            SpirvExecutionModel.GLCompute,
+            main,
+            "main",
+            [gidVar, tiledVar, xTermVar, yTermVar, outVar, pushVar]);
+        return module.Build();
+    }
 }

src/SharpEmu.ShaderCompiler/Gen5ShaderIr.cs

@@ -192,7 +192,8 @@ public sealed record Gen5GlobalMemoryControl(
     uint ScalarAddress,
     int OffsetBytes,
     bool Glc,
-    bool Slc) : Gen5InstructionControl;
+    bool Slc,
+    bool UsesFlatAddress = false) : Gen5InstructionControl;
 
 public sealed record Gen5BufferMemoryControl(
     uint DwordCount,

src/SharpEmu.ShaderCompiler/Gen5ShaderScalarEvaluator.cs

@@ -4,6 +4,7 @@
 using SharpEmu.HLE;
 using System.Buffers;
 using System.Buffers.Binary;
+using System.Collections.Concurrent;
 using System.Diagnostics;
 using System.Numerics;
 
@@ -35,6 +36,14 @@ public static class Gen5ShaderScalarEvaluator
             StringComparison.Ordinal);
     private static readonly object _scalarFallbackTraceGate = new();
     private static readonly HashSet<(ulong Shader, uint Pc)> _tracedScalarFallbacks = [];
+    // Shaders whose empty SRT/EUD caused a null-base scalar pointer load.
+    // Host submit of those translations has lost the Vulkan device; Agc skips
+    // them before QueueSubmit.
+    private static readonly ConcurrentDictionary<ulong, byte> _emptySrtScalarPointerFallbacks =
+        new();
+
+    public static bool WasEmptySrtScalarPointerFallback(ulong shaderAddress) =>
+        _emptySrtScalarPointerFallbacks.ContainsKey(shaderAddress);
 
     // Uniform forward branches select material/resource bodies that remain
     // statically present in the translated shader. Discover the skipped body's
@@ -349,7 +358,10 @@ public static class Gen5ShaderScalarEvaluator
                 if (globalMemory.ScalarAddress >= ScalarRegisterCount - 1)
                 {
                     error =
-                        $"global-address-register-range pc=0x{instruction.Pc:X} " +
+                        $"{(globalMemory.UsesFlatAddress
+                            ? "flat-address-base-unresolved"
+                            : "global-address-register-range")} " +
+                        $"pc=0x{instruction.Pc:X} " +
                         $"s{globalMemory.ScalarAddress}";
                     return false;
                 }
@@ -364,11 +376,18 @@ public static class Gen5ShaderScalarEvaluator
                 }
 
                 var key = (globalMemory.ScalarAddress, baseAddress);
-                var writable = instruction.Opcode.StartsWith(
+                var writable =
+                    instruction.Opcode.StartsWith(
                         "GlobalStore",
                         StringComparison.Ordinal) ||
                     instruction.Opcode.StartsWith(
                         "GlobalAtomic",
+                        StringComparison.Ordinal) ||
+                    instruction.Opcode.StartsWith(
+                        "FlatStore",
+                        StringComparison.Ordinal) ||
+                    instruction.Opcode.StartsWith(
+                        "FlatAtomic",
                         StringComparison.Ordinal);
                 if (globalMemoryByAddress.TryGetValue(key, out var existingBinding))
                 {
@@ -2101,6 +2120,15 @@ public static class Gen5ShaderScalarEvaluator
             $"dynamic={dynamicOffset} definitions=[{string.Join(';', definitions)}] " +
             $"user_data=[{userData}] metadata=" +
             $"{(state.Metadata is null ? "missing" : $"srt={state.Metadata.ShaderResourceTableSizeDwords},eud={state.Metadata.ExtendedUserDataSizeDwords}")}");
+        if (baseAddress == 0 &&
+            state.Metadata is
+            {
+                ShaderResourceTableSizeDwords: 0,
+                ExtendedUserDataSizeDwords: 0,
+            })
+        {
+            _emptySrtScalarPointerFallbacks.TryAdd(state.Program.Address, 0);
+        }
     }
 
     [Conditional("DEBUG")]

src/SharpEmu.ShaderCompiler/Gen5ShaderTranslator.cs

@@ -305,6 +305,23 @@ public static class Gen5ShaderTranslator
             count |= 0x20;
         }
 
+        // Primary SH defaults leave SPI_SHADER_PGM_RSRC2_PS at 0. Draws that
+        // still wrote USER_DATA_n via SetShReg would otherwise translate with
+        // an empty SRT window (Astro title PS → Address-0 descriptors →
+        // device lost). Recover the window from contiguous live registers.
+        if (count == 0 &&
+            userDataBaseRegister is not ComputeUserDataRegister)
+        {
+            var probed = 0;
+            while (probed < MaximumHardwareUserSgprs &&
+                   shaderRegisters.ContainsKey(userDataBaseRegister + (uint)probed))
+            {
+                probed++;
+            }
+
+            count = probed;
+        }
+
         if (userDataBaseRegister is not (PsUserDataRegister or
                 VsUserDataRegister or
                 GsUserDataRegister or
@@ -454,7 +471,15 @@ public static class Gen5ShaderTranslator
                 }
             }
 
-            instructions.Add(CreateInstruction(pc, encoding, name, words));
+            var instruction = CreateInstruction(pc, encoding, name, words);
+            if (instruction.Control is Gen5GlobalMemoryControl
+                {
+                    UsesFlatAddress: true,
+                })
+            {
+                instruction = ResolveFlatAddressBase(instructions, instruction);
+            }
+            instructions.Add(instruction);
             instructionCount++;
 
             pc += sizeDwords * sizeof(uint);
@@ -1378,35 +1403,42 @@ public static class Gen5ShaderTranslator
         var opcode = (word >> 18) & 0x7F;
         sizeDwords = 2;
         error = string.Empty;
-        name = segment == 0x2
-            ? opcode switch
-            {
-                0x08 => "GlobalLoadUbyte",
-                0x09 => "GlobalLoadSbyte",
-                0x0A => "GlobalLoadUshort",
-                0x0B => "GlobalLoadSshort",
-                0x0C => "GlobalLoadDword",
-                0x0D => "GlobalLoadDwordx2",
-                0x0E => "GlobalLoadDwordx4",
-                0x0F => "GlobalLoadDwordx3",
-                0x18 => "GlobalStoreByte",
-                0x19 => "GlobalStoreByteD16Hi",
-                0x1A => "GlobalStoreShort",
-                0x1B => "GlobalStoreShortD16Hi",
-                0x1C => "GlobalStoreDword",
-                0x1D => "GlobalStoreDwordx2",
-                0x1E => "GlobalStoreDwordx4",
-                0x1F => "GlobalStoreDwordx3",
-                0x20 => "GlobalLoadUbyteD16",
-                0x21 => "GlobalLoadUbyteD16Hi",
-                0x22 => "GlobalLoadSbyteD16",
-                0x23 => "GlobalLoadSbyteD16Hi",
-                0x24 => "GlobalLoadShortD16",
-                0x25 => "GlobalLoadShortD16Hi",
-                0x32 => "GlobalAtomicAdd",
-                0x38 => "GlobalAtomicUMax",
-                _ => string.Empty,
-            }
+        var prefix = segment switch
+        {
+            0x0 => "Flat",
+            0x2 => "Global",
+            _ => string.Empty,
+        };
+        var suffix = opcode switch
+        {
+            0x08 => "LoadUbyte",
+            0x09 => "LoadSbyte",
+            0x0A => "LoadUshort",
+            0x0B => "LoadSshort",
+            0x0C => "LoadDword",
+            0x0D => "LoadDwordx2",
+            0x0E => "LoadDwordx4",
+            0x0F => "LoadDwordx3",
+            0x18 => "StoreByte",
+            0x19 => "StoreByteD16Hi",
+            0x1A => "StoreShort",
+            0x1B => "StoreShortD16Hi",
+            0x1C => "StoreDword",
+            0x1D => "StoreDwordx2",
+            0x1E => "StoreDwordx4",
+            0x1F => "StoreDwordx3",
+            0x20 => "LoadUbyteD16",
+            0x21 => "LoadUbyteD16Hi",
+            0x22 => "LoadSbyteD16",
+            0x23 => "LoadSbyteD16Hi",
+            0x24 => "LoadShortD16",
+            0x25 => "LoadShortD16Hi",
+            0x32 => "AtomicAdd",
+            0x38 => "AtomicUMax",
+            _ => string.Empty,
+        };
+        name = prefix.Length != 0 && suffix.Length != 0
+            ? prefix + suffix
             : string.Empty;
 
         return FinishDecode(
@@ -1629,6 +1661,82 @@ public static class Gen5ShaderTranslator
         _ => false,
     };
 
+    private static Gen5ShaderInstruction ResolveFlatAddressBase(
+        IReadOnlyList<Gen5ShaderInstruction> precedingInstructions,
+        Gen5ShaderInstruction instruction)
+    {
+        if (instruction.Control is not Gen5GlobalMemoryControl
+            {
+                UsesFlatAddress: true,
+            } control ||
+            !TryFindVectorDefinition(
+                precedingInstructions,
+                control.VectorAddress,
+                out var lowDefinition) ||
+            !TryFindVectorDefinition(
+                precedingInstructions,
+                control.VectorAddress + 1,
+                out var highDefinition))
+        {
+            return instruction;
+        }
+
+        foreach (var lowSource in lowDefinition.Sources)
+        {
+            if (lowSource.Kind != Gen5OperandKind.ScalarRegister)
+            {
+                continue;
+            }
+
+            foreach (var highSource in highDefinition.Sources)
+            {
+                if (highSource.Kind != Gen5OperandKind.ScalarRegister ||
+                    highSource.Value != lowSource.Value + 1)
+                {
+                    continue;
+                }
+
+                return instruction with
+                {
+                    Sources =
+                    [
+                        .. instruction.Sources,
+                        Gen5Operand.Scalar(lowSource.Value),
+                    ],
+                    Control = control with
+                    {
+                        ScalarAddress = lowSource.Value,
+                    },
+                };
+            }
+        }
+
+        return instruction;
+    }
+
+    private static bool TryFindVectorDefinition(
+        IReadOnlyList<Gen5ShaderInstruction> instructions,
+        uint register,
+        out Gen5ShaderInstruction definition)
+    {
+        for (var index = instructions.Count - 1; index >= 0; index--)
+        {
+            var candidate = instructions[index];
+            foreach (var destination in candidate.Destinations)
+            {
+                if (destination.Kind == Gen5OperandKind.VectorRegister &&
+                    destination.Value == register)
+                {
+                    definition = candidate;
+                    return true;
+                }
+            }
+        }
+
+        definition = default!;
+        return false;
+    }
+
     private static Gen5ShaderInstruction CreateInstruction(
         uint pc,
         Gen5ShaderEncoding encoding,
@@ -2040,7 +2148,13 @@ public static class Gen5ShaderTranslator
                 var vectorAddress = extra & 0xFF;
                 var vectorData = (extra >> 8) & 0xFF;
                 var scalarAddress = (extra >> 16) & 0x7F;
-                var dwordCount = opcode switch
+                var usesFlatAddress = opcode.StartsWith(
+                    "Flat",
+                    StringComparison.Ordinal);
+                var memoryOpcode = usesFlatAddress
+                    ? "Global" + opcode["Flat".Length..]
+                    : opcode;
+                var dwordCount = memoryOpcode switch
                 {
                     "GlobalLoadUbyte" or
                     "GlobalLoadSbyte" or
@@ -2068,12 +2182,20 @@ public static class Gen5ShaderTranslator
                     "GlobalStoreDwordx4" => 4u,
                     _ => 0u,
                 };
-                sources =
-                [
-                    Gen5Operand.Vector(vectorAddress),
-                    Gen5Operand.Scalar(scalarAddress),
-                ];
-                destinations = opcode.StartsWith("GlobalLoad", StringComparison.Ordinal)
+                sources = usesFlatAddress
+                    ?
+                    [
+                        Gen5Operand.Vector(vectorAddress),
+                        Gen5Operand.Vector(vectorAddress + 1),
+                    ]
+                    :
+                    [
+                        Gen5Operand.Vector(vectorAddress),
+                        Gen5Operand.Scalar(scalarAddress),
+                    ];
+                destinations = memoryOpcode.StartsWith(
+                        "GlobalLoad",
+                        StringComparison.Ordinal)
                     ? Enumerable
                         .Range((int)vectorData, checked((int)dwordCount))
                         .Select(index => Gen5Operand.Vector((uint)index))
@@ -2083,10 +2205,11 @@ public static class Gen5ShaderTranslator
                     dwordCount,
                     vectorAddress,
                     vectorData,
-                    scalarAddress,
+                    usesFlatAddress ? uint.MaxValue : scalarAddress,
                     SignExtend(word & 0x1FFF, 13),
                     ((word >> 16) & 1) != 0,
-                    ((word >> 17) & 1) != 0);
+                    ((word >> 17) & 1) != 0,
+                    usesFlatAddress);
                 break;
             }
             case Gen5ShaderEncoding.Mubuf:

tests/SharpEmu.Libs.Tests/Agc/AgcTextureTransportTests.cs

@@ -0,0 +1,106 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.Libs.Agc;
+using SharpEmu.Libs.Gpu;
+using Xunit;
+
+namespace SharpEmu.Libs.Tests.Agc;
+
+public sealed class AgcTextureTransportTests
+{
+    [Theory]
+    [InlineData(10u, 4u, 4u)]
+    [InlineData(10u, 0u, 1u)]
+    [InlineData(9u, 4u, 1u)]
+    [InlineData(13u, 4u, 1u)]
+    public void GetTextureVolumeDepth_OnlyUsesDescriptorDepthFor3D(
+        uint type,
+        uint descriptorDepth,
+        uint expectedDepth)
+    {
+        Assert.Equal(
+            expectedDepth,
+            AgcExports.GetTextureVolumeDepth(type, descriptorDepth));
+    }
+
+    [Fact]
+    public void GetTextureByteCount_MultipliesUncompressedVolumeDepth()
+    {
+        Assert.Equal(
+            4UL * 8 * 6 * 5,
+            AgcExports.GetTextureByteCount(
+                format: 10,
+                width: 8,
+                height: 6,
+                depth: 5));
+    }
+
+    [Fact]
+    public void GetTextureByteCount_MultipliesBlockCompressedVolumeDepth()
+    {
+        // Format 169 uses one eight-byte BC block for each 4x4 texel block.
+        Assert.Equal(
+            2UL * 2 * 8 * 3,
+            AgcExports.GetTextureByteCount(
+                format: 169,
+                width: 7,
+                height: 5,
+                depth: 3));
+    }
+
+    [Fact]
+    public void GetTextureByteCount_LeavesTwoDimensionalSizingUnchanged()
+    {
+        Assert.Equal(
+            AgcExports.GetTextureByteCount(10, 8, 6),
+            AgcExports.GetTextureByteCount(10, 8, 6, depth: 1));
+        Assert.Equal(
+            AgcExports.GetTextureByteCount(10, 8, 6),
+            AgcExports.GetTextureByteCount(10, 8, 6, depth: 0));
+    }
+
+    [Fact]
+    public void GuestDrawTexture_CarriesRawTypeAndNormalizedDepth()
+    {
+        var texture = new GuestDrawTexture(
+            Address: 0x1234,
+            Width: 8,
+            Height: 6,
+            Format: 10,
+            NumberType: 0,
+            RgbaPixels: [],
+            IsFallback: false,
+            IsStorage: false,
+            Type: 10,
+            Depth: 5);
+
+        Assert.Equal(10u, texture.Type);
+        Assert.Equal(5u, texture.Depth);
+    }
+
+    [Fact]
+    public void TextureContentIdentity_DistinguishesTypeAndDepth()
+    {
+        var twoDimensional = CreateIdentity(type: 9, depth: 1);
+        var threeDimensional = CreateIdentity(type: 10, depth: 1);
+        var deeperThreeDimensional = CreateIdentity(type: 10, depth: 5);
+
+        Assert.NotEqual(twoDimensional, threeDimensional);
+        Assert.NotEqual(threeDimensional, deeperThreeDimensional);
+    }
+
+    private static TextureContentIdentity CreateIdentity(uint type, uint depth) =>
+        new(
+            Address: 0x1234,
+            Width: 8,
+            Height: 6,
+            Format: 10,
+            NumberType: 0,
+            DstSelect: 0xFAC,
+            TileMode: 0,
+            Pitch: 8,
+            Sampler: default,
+            Type: type,
+            Depth: depth);
+}

tests/SharpEmu.Libs.Tests/Agc/DetileComputeSpirvTests.cs

@@ -0,0 +1,102 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using SharpEmu.ShaderCompiler.Vulkan;
+using Xunit;
+
+namespace SharpEmu.Libs.Tests.Agc;
+
+// Structural validation of the GPU detile compute kernel. This cannot run the
+// shader without a Vulkan device, but it pins the SPIR-V is well-formed: the
+// header is correct, every instruction's word count sums to exactly the module
+// length (the classic hand-emit bug), and the compute-specific pieces are
+// present (a GLCompute entry point, a LocalSize execution mode, and a runtime
+// array for the storage buffers). Full pixel correctness is verified on a GPU.
+public sealed class DetileComputeSpirvTests
+{
+    private const uint SpirvMagic = 0x07230203;
+    private const uint SpirvVersion15 = 0x00010500;
+
+    private const ushort OpEntryPoint = 15;
+    private const ushort OpExecutionMode = 16;
+    private const ushort OpTypeRuntimeArray = 29;
+    private const ushort OpFunction = 54;
+    private const ushort OpFunctionEnd = 56;
+
+    private const uint ExecutionModelGLCompute = 5;
+    private const uint ExecutionModeLocalSize = 17;
+
+    [Fact]
+    public void CreateDetileCompute_EmitsWellFormedComputeModule()
+    {
+        var spirv = SpirvFixedShaders.CreateDetileCompute();
+
+        Assert.True(spirv.Length % sizeof(uint) == 0, "SPIR-V must be a whole number of words.");
+        var words = new uint[spirv.Length / sizeof(uint)];
+        for (var i = 0; i < words.Length; i++)
+        {
+            words[i] = BinaryPrimitives.ReadUInt32LittleEndian(spirv.AsSpan(i * sizeof(uint)));
+        }
+
+        Assert.True(words.Length > 5, "Module must have a header plus instructions.");
+        Assert.Equal(SpirvMagic, words[0]);
+        Assert.Equal(SpirvVersion15, words[1]);
+
+        var bound = words[3];
+        Assert.True(bound > 1, "Id bound must be set.");
+
+        var sawComputeEntry = false;
+        var sawLocalSize = false;
+        var sawRuntimeArray = false;
+        var functionCount = 0;
+        var functionEndCount = 0;
+
+        var offset = 5;
+        while (offset < words.Length)
+        {
+            var word = words[offset];
+            var wordCount = (int)(word >> 16);
+            var opcode = (ushort)(word & 0xFFFF);
+
+            Assert.True(wordCount >= 1, $"Instruction at {offset} has a zero word count.");
+            Assert.True(
+                offset + wordCount <= words.Length,
+                $"Instruction at {offset} (op {opcode}, wc {wordCount}) overruns the module.");
+
+            switch (opcode)
+            {
+                case OpEntryPoint when words[offset + 1] == ExecutionModelGLCompute:
+                    sawComputeEntry = true;
+                    break;
+                case OpExecutionMode
+                    when wordCount >= 6 &&
+                         words[offset + 2] == ExecutionModeLocalSize &&
+                         words[offset + 3] == 8 &&
+                         words[offset + 4] == 8 &&
+                         words[offset + 5] == 1:
+                    sawLocalSize = true;
+                    break;
+                case OpTypeRuntimeArray:
+                    sawRuntimeArray = true;
+                    break;
+                case OpFunction:
+                    functionCount++;
+                    break;
+                case OpFunctionEnd:
+                    functionEndCount++;
+                    break;
+            }
+
+            offset += wordCount;
+        }
+
+        // Word counts must tile the module exactly — a wrong length lands here.
+        Assert.Equal(words.Length, offset);
+        Assert.True(sawComputeEntry, "Missing a GLCompute OpEntryPoint.");
+        Assert.True(sawLocalSize, "Missing an 8x8x1 LocalSize execution mode.");
+        Assert.True(sawRuntimeArray, "Missing a runtime array (storage buffers).");
+        Assert.Equal(1, functionCount);
+        Assert.Equal(1, functionEndCount);
+    }
+}

tests/SharpEmu.Libs.Tests/Agc/GnmTilingDetileTests.cs

@@ -83,4 +83,152 @@ public sealed class GnmTilingDetileTests
             Assert.Equal((ushort)i, value);
         }
     }
+
+    // GetDetileParams must reproduce TryDetile bit-for-bit: the CPU fallback and
+    // the GPU compute kernel both consume these params, so a detile driven purely
+    // by DetileParams (the shared addressing formula the kernel runs) must equal
+    // the shipped CPU detile for every supported mode/bpp.
+    [Theory]
+    [InlineData(27u, 2, 384, 200)] // 64 KiB RB+ R_X (exact-XOR)
+    [InlineData(27u, 4, 256, 256)] // 64 KiB RB+ R_X (exact-XOR)
+    [InlineData(9u, 4, 300, 300)]  // 64 KiB standard (exact-XOR)
+    [InlineData(24u, 4, 128, 256)] // 64 KiB RB+ Z_X (exact-XOR)
+    [InlineData(5u, 4, 200, 120)]  // 4 KiB standard (exact-XOR)
+    [InlineData(8u, 4, 128, 128)]  // 64 KiB Z (block-table path)
+    [InlineData(1u, 4, 64, 64)]    // 256 B standard (block-table path)
+    public void GetDetileParams_ReproducesTryDetile(uint mode, int bpp, int w, int h)
+    {
+        var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
+        Assert.True(p.IsSupported);
+
+        // Whole-block tiled buffer (block addressing overshoots the linear extent),
+        // filled with a deterministic non-trivial pattern.
+        var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
+        var tiled = new byte[(long)p.BlocksPerRow * blocksHigh * p.BlockBytes];
+        for (var i = 0; i < tiled.Length; i++)
+        {
+            tiled[i] = (byte)((i * 31 + 7) & 0xFF);
+        }
+
+        var expected = new byte[w * h * bpp];
+        Assert.True(GnmTiling.TryDetile(tiled, expected, mode, w, h, bpp));
+
+        var actual = DetileViaParams(tiled, p, w, h, bpp);
+        Assert.Equal(expected, actual);
+    }
+
+    // The production GnmTiling.DetileWithParams (the active CPU fallback used by
+    // the Metal path under default-on GPU detile) must equal TryDetile for every
+    // supported mode/bpp — same DetileParams addressing, no re-derived swizzle.
+    [Theory]
+    [InlineData(27u, 2, 384, 200)]
+    [InlineData(27u, 4, 256, 256)]
+    [InlineData(9u, 4, 300, 300)]
+    [InlineData(24u, 4, 128, 256)]
+    [InlineData(5u, 4, 200, 120)]
+    [InlineData(8u, 4, 128, 128)]
+    [InlineData(1u, 4, 64, 64)]
+    [InlineData(27u, 8, 256, 256)]  // 8bpp (GPU: 2 words/element)
+    [InlineData(27u, 16, 128, 128)] // 16bpp (GPU: 4 words/element)
+    [InlineData(9u, 8, 128, 96)]
+    [InlineData(8u, 16, 64, 64)]    // block-table, 16bpp
+    public void DetileWithParams_MatchesTryDetile(uint mode, int bpp, int w, int h)
+    {
+        var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
+        Assert.True(p.IsSupported);
+
+        var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
+        var tiled = new byte[(long)p.BlocksPerRow * blocksHigh * p.BlockBytes];
+        for (var i = 0; i < tiled.Length; i++)
+        {
+            tiled[i] = (byte)((i * 31 + 7) & 0xFF);
+        }
+
+        var expected = new byte[w * h * bpp];
+        Assert.True(GnmTiling.TryDetile(tiled, expected, mode, w, h, bpp));
+
+        var actual = new byte[w * h * bpp];
+        Assert.True(GnmTiling.DetileWithParams(p, tiled, actual));
+        Assert.Equal(expected, actual);
+    }
+
+    // Array textures are packed as contiguous tiled slices and detiled one slice
+    // per layer (dispatch-Z on the GPU; a per-layer loop in the CPU fallbacks)
+    // into a layer-major linear buffer. This pins that packing: each slice must
+    // deswizzle into its own region and match a per-slice TryDetile, and a
+    // per-layer-distinct pattern catches any slice cross-talk.
+    [Theory]
+    [InlineData(27u, 4, 256, 256, 3)]
+    [InlineData(9u, 4, 128, 96, 2)]
+    [InlineData(24u, 4, 64, 128, 4)]
+    public void DetileWithParams_MultiLayer_MatchesPerSliceTryDetile(uint mode, int bpp, int w, int h, int layers)
+    {
+        var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
+        Assert.True(p.IsSupported);
+
+        var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
+        var sliceTiledBytes = (int)((long)p.BlocksPerRow * blocksHigh * p.BlockBytes);
+        var sliceLinearBytes = w * h * bpp;
+
+        var tiled = new byte[sliceTiledBytes * layers];
+        for (var layer = 0; layer < layers; layer++)
+        {
+            for (var i = 0; i < sliceTiledBytes; i++)
+            {
+                tiled[layer * sliceTiledBytes + i] = (byte)((i * 31 + 7 + layer * 101) & 0xFF);
+            }
+        }
+
+        // Expected: each slice detiled independently via the shipped CPU detile.
+        var expected = new byte[sliceLinearBytes * layers];
+        for (var layer = 0; layer < layers; layer++)
+        {
+            Assert.True(GnmTiling.TryDetile(
+                tiled.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
+                expected.AsSpan(layer * sliceLinearBytes, sliceLinearBytes),
+                mode, w, h, bpp));
+        }
+
+        // Actual: the layer-major loop the Vulkan/Metal CPU fallbacks run.
+        var actual = new byte[sliceLinearBytes * layers];
+        for (var layer = 0; layer < layers; layer++)
+        {
+            Assert.True(GnmTiling.DetileWithParams(
+                p,
+                tiled.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
+                actual.AsSpan(layer * sliceLinearBytes, sliceLinearBytes)));
+        }
+
+        Assert.Equal(expected, actual);
+    }
+
+    // Reference detile driven entirely by DetileParams — the single shared
+    // addressing formula the Vulkan/Metal compute kernel will run per texel.
+    private static byte[] DetileViaParams(byte[] tiled, DetileParams p, int w, int h, int bpp)
+    {
+        var linear = new byte[w * h * bpp];
+        for (var y = 0; y < h; y++)
+        {
+            for (var x = 0; x < w; x++)
+            {
+                var blockX = x / p.BlockWidth;
+                var blockY = y / p.BlockHeight;
+                var inX = x % p.BlockWidth;
+                var inY = y % p.BlockHeight;
+                var inBlockByte = p.Equation == DetileEquation.ExactXor
+                    ? p.XByteTerm[x & p.XMask] ^ p.YByteTerm[y & p.YMask]
+                    : p.BlockTable[inY * p.BlockWidth + inX] * p.BytesPerElement;
+                var srcByte = ((long)blockY * p.BlocksPerRow + blockX) * p.BlockBytes + inBlockByte;
+                var dstByte = ((long)y * w + x) * bpp;
+                if (srcByte < 0 || srcByte + bpp > tiled.Length)
+                {
+                    continue;
+                }
+
+                Array.Copy(tiled, srcByte, linear, dstByte, bpp);
+            }
+        }
+
+        return linear;
+    }
 }

tests/SharpEmu.Libs.Tests/Audio/AudioOutExportsTests.cs

@@ -0,0 +1,231 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using SharpEmu.HLE;
+using SharpEmu.HLE.Host;
+using SharpEmu.Libs.Audio;
+using Xunit;
+
+namespace SharpEmu.Libs.Tests.Audio;
+
+[CollectionDefinition("AudioOutState", DisableParallelization = true)]
+public sealed class AudioOutStateCollection
+{
+    public const string Name = "AudioOutState";
+}
+
+[Collection(AudioOutStateCollection.Name)]
+public sealed class AudioOutExportsTests : IDisposable
+{
+    private const ulong MemoryBase = 0x1_0000_0000;
+    private const int MemorySize = 0x4000;
+    private const ulong ParameterAddress = MemoryBase + 0x100;
+    private const ulong FirstSourceAddress = MemoryBase + 0x1000;
+    private const ulong SecondSourceAddress = MemoryBase + 0x2000;
+
+    private readonly FakeCpuMemory _memory = new(MemoryBase, MemorySize);
+    private readonly CpuContext _ctx;
+    private readonly List<RecordingAudioStream> _streams = [];
+
+    public AudioOutExportsTests()
+    {
+        AudioOutExports.ResetForTests();
+        AudioOutExports.SetStreamFactoryForTests(_ =>
+        {
+            var stream = new RecordingAudioStream();
+            _streams.Add(stream);
+            return stream;
+        });
+        _ctx = new CpuContext(_memory, Generation.Gen5);
+    }
+
+    [Fact]
+    public void Outputs_StagesAndSubmitsSingleStereoPort()
+    {
+        var handle = OpenPort(bufferLength: 2);
+        Span<byte> source = stackalloc byte[8];
+        BinaryPrimitives.WriteInt16LittleEndian(source, -32768);
+        BinaryPrimitives.WriteInt16LittleEndian(source[2..], 32767);
+        BinaryPrimitives.WriteInt16LittleEndian(source[4..], -1234);
+        BinaryPrimitives.WriteInt16LittleEndian(source[6..], 5678);
+        Assert.True(_memory.TryWrite(FirstSourceAddress, source));
+        WriteDescriptor(0, handle, FirstSourceAddress);
+
+        var result = Submit(outputCount: 1);
+
+        Assert.Equal(2, result);
+        Assert.Equal(2UL, _ctx[CpuRegister.Rax]);
+        Assert.Equal(source.ToArray(), Assert.Single(_streams[0].Submissions));
+    }
+
+    [Fact]
+    public void Outputs_SubmitsEveryPortInTheBatch()
+    {
+        var firstHandle = OpenPort(bufferLength: 2);
+        var secondHandle = OpenPort(bufferLength: 2);
+        byte[] firstSource = [1, 0, 2, 0, 3, 0, 4, 0];
+        byte[] secondSource = [5, 0, 6, 0, 7, 0, 8, 0];
+        Assert.True(_memory.TryWrite(FirstSourceAddress, firstSource));
+        Assert.True(_memory.TryWrite(SecondSourceAddress, secondSource));
+
+        // Reverse handle order to exercise canonical lock ordering without
+        // changing which guest buffer belongs to each port.
+        WriteDescriptor(0, secondHandle, FirstSourceAddress);
+        WriteDescriptor(1, firstHandle, SecondSourceAddress);
+
+        var result = Submit(outputCount: 2);
+
+        Assert.Equal(2, result);
+        Assert.Equal(secondSource, Assert.Single(_streams[0].Submissions));
+        Assert.Equal(firstSource, Assert.Single(_streams[1].Submissions));
+    }
+
+    [Fact]
+    public void Outputs_AcceptsNullBufferAsSynchronizationOnly()
+    {
+        var handle = OpenPort(bufferLength: 2);
+        WriteDescriptor(0, handle, sourceAddress: 0);
+
+        var result = Submit(outputCount: 1);
+
+        Assert.Equal(2, result);
+        Assert.Empty(_streams[0].Submissions);
+    }
+
+    [Fact]
+    public void Outputs_FaultInLaterBufferDoesNotPartiallySubmit()
+    {
+        var firstHandle = OpenPort(bufferLength: 2);
+        var secondHandle = OpenPort(bufferLength: 2);
+        Assert.True(_memory.TryWrite(FirstSourceAddress, new byte[8]));
+        WriteDescriptor(0, firstHandle, FirstSourceAddress);
+        WriteDescriptor(1, secondHandle, MemoryBase + MemorySize);
+
+        var result = Submit(outputCount: 2);
+
+        Assert.Equal(AudioOutExports.AudioOutErrorInvalidPointer, result);
+        Assert.All(_streams, stream => Assert.Empty(stream.Submissions));
+    }
+
+    [Fact]
+    public void Outputs_RejectsDuplicateHandlesBeforeSubmission()
+    {
+        var handle = OpenPort(bufferLength: 2);
+        Assert.True(_memory.TryWrite(FirstSourceAddress, new byte[8]));
+        Assert.True(_memory.TryWrite(SecondSourceAddress, new byte[8]));
+        WriteDescriptor(0, handle, FirstSourceAddress);
+        WriteDescriptor(1, handle, SecondSourceAddress);
+
+        var result = Submit(outputCount: 2);
+
+        Assert.Equal(AudioOutExports.AudioOutErrorInvalidPort, result);
+        Assert.Empty(_streams[0].Submissions);
+    }
+
+    [Fact]
+    public void Outputs_RejectsPortsWithDifferentBufferLengths()
+    {
+        var firstHandle = OpenPort(bufferLength: 2);
+        var secondHandle = OpenPort(bufferLength: 4);
+        WriteDescriptor(0, firstHandle, FirstSourceAddress);
+        WriteDescriptor(1, secondHandle, SecondSourceAddress);
+
+        var result = Submit(outputCount: 2);
+
+        Assert.Equal(AudioOutExports.AudioOutErrorInvalidSize, result);
+        Assert.All(_streams, stream => Assert.Empty(stream.Submissions));
+    }
+
+    [Fact]
+    public void Outputs_RejectsUnknownHandle()
+    {
+        WriteDescriptor(0, 99, FirstSourceAddress);
+
+        Assert.Equal(
+            AudioOutExports.AudioOutErrorInvalidPort,
+            Submit(outputCount: 1));
+    }
+
+    [Theory]
+    [InlineData(0u)]
+    [InlineData(26u)]
+    public void Outputs_RejectsInvalidOutputCount(uint outputCount)
+    {
+        Assert.Equal(
+            AudioOutExports.AudioOutErrorPortFull,
+            Submit(outputCount));
+    }
+
+    [Fact]
+    public void Outputs_RejectsNullOrUnreadableParameterArray()
+    {
+        Assert.Equal(
+            AudioOutExports.AudioOutErrorInvalidPointer,
+            Submit(outputCount: 1, parameterAddress: 0));
+        Assert.Equal(
+            AudioOutExports.AudioOutErrorInvalidPointer,
+            Submit(outputCount: 1, parameterAddress: MemoryBase + MemorySize));
+    }
+
+    [Fact]
+    public void OutputsExportRegistersForBothGenerations()
+    {
+        foreach (var generation in new[] { Generation.Gen4, Generation.Gen5 })
+        {
+            var manager = new ModuleManager();
+            manager.RegisterExports(
+                SharpEmu.Generated.SysAbiExportRegistry.CreateExports(generation));
+
+            Assert.True(manager.TryGetExport("w3PdaSTSwGE", out var export));
+            Assert.Equal("sceAudioOutOutputs", export.Name);
+            Assert.Equal("libSceAudioOut", export.LibraryName);
+        }
+    }
+
+    public void Dispose() => AudioOutExports.ResetForTests();
+
+    private int OpenPort(uint bufferLength)
+    {
+        _ctx[CpuRegister.Rdi] = 1;
+        _ctx[CpuRegister.Rsi] = 0;
+        _ctx[CpuRegister.Rdx] = 0;
+        _ctx[CpuRegister.Rcx] = bufferLength;
+        _ctx[CpuRegister.R8] = 48000;
+        _ctx[CpuRegister.R9] = 1;
+        return AudioOutExports.AudioOutOpen(_ctx);
+    }
+
+    private int Submit(uint outputCount, ulong parameterAddress = ParameterAddress)
+    {
+        _ctx[CpuRegister.Rdi] = parameterAddress;
+        _ctx[CpuRegister.Rsi] = outputCount;
+        return AudioOutExports.AudioOutOutputs(_ctx);
+    }
+
+    private void WriteDescriptor(int index, int handle, ulong sourceAddress)
+    {
+        Span<byte> descriptor = stackalloc byte[16];
+        descriptor.Clear();
+        BinaryPrimitives.WriteInt32LittleEndian(descriptor, handle);
+        BinaryPrimitives.WriteUInt64LittleEndian(descriptor[8..], sourceAddress);
+        Assert.True(_memory.TryWrite(
+            ParameterAddress + unchecked((ulong)(index * descriptor.Length)),
+            descriptor));
+    }
+
+    private sealed class RecordingAudioStream : IHostAudioStream
+    {
+        public List<byte[]> Submissions { get; } = [];
+
+        public bool Submit(ReadOnlySpan<byte> stereoPcm16)
+        {
+            Submissions.Add(stereoPcm16.ToArray());
+            return true;
+        }
+
+        public void Dispose()
+        {
+        }
+    }
+}

tests/SharpEmu.Libs.Tests/Bink/BinkFramePlaybackTests.cs

@@ -83,7 +83,9 @@ public sealed class BinkFramePlaybackTests
 
         public uint Height => 1;
 
-        public uint FramesPerSecondNumerator => 20;
+        // Keep frame boundaries far enough apart that a loaded CI runner cannot
+        // skip an expected frame between polling iterations.
+        public uint FramesPerSecondNumerator => 2;
 
         public uint FramesPerSecondDenominator => 1;
 

tests/SharpEmu.Libs.Tests/SystemService/SystemServiceExportsTests.cs

@@ -11,6 +11,11 @@ public sealed class SystemServiceExportsTests
 {
     private const ulong MemoryBase = 0x1_0000_0000;
 
+    public SystemServiceExportsTests()
+    {
+        SystemServiceExports.ResetForTests();
+    }
+
     [Fact]
     public void GetNoticeScreenSkipFlagWritesOneByteAtMemoryBoundary()
     {
@@ -26,4 +31,23 @@ public sealed class SystemServiceExportsTests
         Assert.True(memory.TryRead(MemoryBase, flag));
         Assert.Equal(0, flag[0]);
     }
+
+    [Fact]
+    public void SetNoticeScreenSkipFlagRoundTripsThroughGetter()
+    {
+        var memory = new FakeCpuMemory(MemoryBase, 2);
+        var context = new CpuContext(memory, Generation.Gen5)
+        {
+            [CpuRegister.Rdi] = 1,
+        };
+
+        Assert.Equal(0, SystemServiceExports.SystemServiceSetNoticeScreenSkipFlag(context));
+
+        context[CpuRegister.Rdi] = MemoryBase;
+        Assert.Equal(0, SystemServiceExports.SystemServiceGetNoticeScreenSkipFlag(context));
+
+        Span<byte> flag = stackalloc byte[1];
+        Assert.True(memory.TryRead(MemoryBase, flag));
+        Assert.Equal(1, flag[0]);
+    }
 }

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

@@ -36,4 +36,23 @@ public sealed class VulkanGuestImageByteCountTests
             expected,
             VulkanVideoPresenter.GetGuestImageByteCount(format, width, height));
     }
+
+    [Theory]
+    [InlineData(10u, 8u, 4u, 3u, 384UL)]
+    [InlineData(169u, 5u, 5u, 7u, 224UL)]
+    public void MultipliesSurfaceSizeByVolumeDepth(
+        uint format,
+        uint width,
+        uint height,
+        uint depth,
+        ulong expected)
+    {
+        Assert.Equal(
+            expected,
+            VulkanVideoPresenter.GetGuestImageByteCount(
+                format,
+                width,
+                height,
+                depth));
+    }
 }

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

@@ -0,0 +1,49 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using Silk.NET.Vulkan;
+using SharpEmu.Libs.VideoOut;
+using Xunit;
+
+namespace SharpEmu.Libs.Tests.VideoOut;
+
+public sealed class VulkanGuestImageTypeTests
+{
+    [Fact]
+    public void ThreeDimensionalDescriptorsMapToVolumeImageAndViewTypes()
+    {
+        Assert.Equal(
+            ImageType.Type3D,
+            VulkanVideoPresenter.GetGuestTextureImageType(
+                VulkanVideoPresenter.Gen5TextureType3D));
+        Assert.Equal(
+            ImageViewType.Type3D,
+            VulkanVideoPresenter.GetGuestTextureViewType(
+                VulkanVideoPresenter.Gen5TextureType3D,
+                arrayedView: true));
+        Assert.Equal(
+            7u,
+            VulkanVideoPresenter.GetGuestTextureDepth(
+                VulkanVideoPresenter.Gen5TextureType3D,
+                7));
+    }
+
+    [Fact]
+    public void TwoDimensionalArraysKeepLayersSeparateFromImageDepth()
+    {
+        Assert.Equal(
+            ImageType.Type2D,
+            VulkanVideoPresenter.GetGuestTextureImageType(
+                VulkanVideoPresenter.Gen5TextureType2D));
+        Assert.Equal(
+            ImageViewType.Type2DArray,
+            VulkanVideoPresenter.GetGuestTextureViewType(
+                VulkanVideoPresenter.Gen5TextureType2D,
+                arrayedView: true));
+        Assert.Equal(
+            1u,
+            VulkanVideoPresenter.GetGuestTextureDepth(
+                VulkanVideoPresenter.Gen5TextureType2D,
+                7));
+    }
+}

tests/SharpEmu.ShaderCompiler.Tests/Gen5FlatMemoryTests.cs

@@ -0,0 +1,179 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using SharpEmu.HLE;
+using SharpEmu.ShaderCompiler;
+using SharpEmu.ShaderCompiler.Vulkan;
+using Xunit;
+
+namespace SharpEmu.ShaderCompiler.Tests;
+
+public sealed class Gen5FlatMemoryTests
+{
+    private const ulong ShaderAddress = 0x1_0000_0000;
+    private const uint SEndpgm = 0xBF810000;
+
+    [Fact]
+    public void FlatLoadUbyteInfersScalarBaseAndCompiles()
+    {
+        var memory = new TestCpuMemory(ShaderAddress, 0x4000);
+        uint[] words =
+        [
+            // v_add_co_u32 v1, vcc_lo, s12, v6
+            0xD70F6A01,
+            0x00020C0C,
+            // v_add_co_ci_u32_sdwa v2, vcc_lo, 0, s13, vcc_lo
+            0x50041AF9,
+            0x86860680,
+            // flat_load_ubyte v0, v[1:2]
+            0xDC200000,
+            0x007D0001,
+            SEndpgm,
+        ];
+        var shader = new byte[words.Length * sizeof(uint)];
+        for (var index = 0; index < words.Length; index++)
+        {
+            BinaryPrimitives.WriteUInt32LittleEndian(
+                shader.AsSpan(index * sizeof(uint)),
+                words[index]);
+        }
+        Assert.True(memory.TryWrite(ShaderAddress, shader));
+
+        var ctx = new CpuContext(memory, Generation.Gen5);
+        Assert.True(
+            Gen5ShaderTranslator.TryDecodeProgram(
+                ctx,
+                ShaderAddress,
+                out var program,
+                out var decodeError),
+            decodeError);
+
+        var instruction = Assert.Single(
+            program.Instructions,
+            item => item.Opcode == "FlatLoadUbyte");
+        var control = Assert.IsType<Gen5GlobalMemoryControl>(
+            instruction.Control);
+        Assert.True(control.UsesFlatAddress);
+        Assert.Equal(1u, control.VectorAddress);
+        Assert.Equal(0u, control.VectorData);
+        Assert.Equal(12u, control.ScalarAddress);
+        Assert.Equal(
+            [
+                Gen5Operand.Vector(1),
+                Gen5Operand.Vector(2),
+                Gen5Operand.Scalar(12),
+            ],
+            instruction.Sources);
+
+        uint[] userData =
+        [
+            unchecked((uint)ShaderAddress),
+            unchecked((uint)(ShaderAddress >> 32)),
+        ];
+        var state = new Gen5ShaderState(
+            program,
+            userData,
+            null,
+            UserDataScalarRegisterBase: 12);
+        Assert.True(
+            Gen5ShaderScalarEvaluator.TryEvaluate(
+                ctx,
+                state,
+                out var evaluation,
+                out var evaluationError),
+            evaluationError);
+
+        var binding = Assert.Single(evaluation.GlobalMemoryBindings);
+        Assert.Equal(12u, binding.ScalarAddress);
+        Assert.Contains(instruction.Pc, binding.InstructionPcs);
+        Assert.True(
+            Gen5SpirvTranslator.TryCompileComputeShader(
+                state,
+                evaluation,
+                1,
+                1,
+                1,
+                out var compiled,
+                out var compileError),
+            compileError);
+        Assert.Contains(
+            (ushort)SpirvOp.ISub,
+            ReadSpirvOpcodes(compiled.Spirv));
+    }
+
+    private static IReadOnlyList<ushort> ReadSpirvOpcodes(byte[] spirv)
+    {
+        Assert.Equal(0, spirv.Length % sizeof(uint));
+        Assert.True(spirv.Length >= 5 * sizeof(uint));
+        Assert.Equal(
+            0x07230203u,
+            BinaryPrimitives.ReadUInt32LittleEndian(spirv));
+
+        var opcodes = new List<ushort>();
+        for (var offset = 5 * sizeof(uint); offset < spirv.Length;)
+        {
+            var instruction =
+                BinaryPrimitives.ReadUInt32LittleEndian(spirv.AsSpan(offset));
+            var wordCount = checked((int)(instruction >> 16));
+            Assert.InRange(
+                wordCount,
+                1,
+                (spirv.Length - offset) / sizeof(uint));
+            opcodes.Add((ushort)instruction);
+            offset += wordCount * sizeof(uint);
+        }
+
+        return opcodes;
+    }
+
+    private sealed class TestCpuMemory(ulong baseAddress, int size) : ICpuMemory
+    {
+        private readonly byte[] _storage = new byte[size];
+
+        public bool TryRead(ulong virtualAddress, Span<byte> destination)
+        {
+            if (!TryResolve(virtualAddress, destination.Length, out var offset))
+            {
+                return false;
+            }
+
+            _storage.AsSpan(offset, destination.Length).CopyTo(destination);
+            return true;
+        }
+
+        public bool TryWrite(
+            ulong virtualAddress,
+            ReadOnlySpan<byte> source)
+        {
+            if (!TryResolve(virtualAddress, source.Length, out var offset))
+            {
+                return false;
+            }
+
+            source.CopyTo(_storage.AsSpan(offset, source.Length));
+            return true;
+        }
+
+        private bool TryResolve(
+            ulong virtualAddress,
+            int length,
+            out int offset)
+        {
+            offset = 0;
+            if (virtualAddress < baseAddress)
+            {
+                return false;
+            }
+
+            var relative = virtualAddress - baseAddress;
+            if (relative + (ulong)length > (ulong)_storage.Length)
+            {
+                return false;
+            }
+
+            offset = (int)relative;
+            return true;
+        }
+    }
+}

tests/SharpEmu.ShaderCompiler.Tests/Gen5ImageTests.cs

@@ -0,0 +1,197 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using SharpEmu.ShaderCompiler;
+using SharpEmu.ShaderCompiler.Vulkan;
+using Xunit;
+
+namespace SharpEmu.ShaderCompiler.Tests;
+
+public sealed class Gen5ImageTests
+{
+    private const ulong ShaderAddress = 0x1_0000_C000;
+    private const uint SEndpgm = 0xBF810000;
+
+    [Theory]
+    [InlineData(1u, SpirvImageDim.Dim2D, 2u)]
+    [InlineData(2u, SpirvImageDim.Dim3D, 3u)]
+    public void ImageStoreDimensionControlsImageAndCoordinateTypes(
+        uint dimension,
+        SpirvImageDim expectedImageDimension,
+        uint expectedCoordinateComponents)
+    {
+        var instructions = ReadSpirvInstructions(
+            CompileImageOperation("ImageStore", dimension));
+        var imageType = Assert.Single(
+            instructions,
+            item => item.Opcode == SpirvOp.TypeImage);
+
+        Assert.Equal((uint)expectedImageDimension, imageType.Operands[2]);
+        Assert.Equal(2u, imageType.Operands[6]);
+        AssertCoordinateVectorWidth(
+            instructions,
+            SpirvOp.ImageWrite,
+            coordinateOperand: 1,
+            expectedComponents: expectedCoordinateComponents);
+
+        var sizeQuery = Assert.Single(
+            instructions,
+            item => item.Opcode == SpirvOp.ImageQuerySize);
+        AssertVectorTypeWidth(
+            instructions,
+            sizeQuery.Operands[0],
+            expectedCoordinateComponents);
+    }
+
+    [Fact]
+    public void ImageSampleDim3DUsesThreeComponentSampleCoordinates()
+    {
+        var instructions = ReadSpirvInstructions(
+            CompileImageOperation("ImageSampleLz", dimension: 2));
+        var imageType = Assert.Single(
+            instructions,
+            item => item.Opcode == SpirvOp.TypeImage);
+
+        Assert.Equal((uint)SpirvImageDim.Dim3D, imageType.Operands[2]);
+        Assert.Equal(1u, imageType.Operands[6]);
+        AssertCoordinateVectorWidth(
+            instructions,
+            SpirvOp.ImageSampleExplicitLod,
+            coordinateOperand: 3,
+            expectedComponents: 3);
+    }
+
+    private static byte[] CompileImageOperation(string opcode, uint dimension)
+    {
+        var addressRegisters = dimension == 2
+            ? new uint[] { 0, 1, 2 }
+            : [0, 1];
+        var control = new Gen5ImageControl(
+            Dmask: 0xF,
+            VectorAddress: 0,
+            AddressRegisters: addressRegisters,
+            VectorData: 4,
+            ScalarResource: 8,
+            ScalarSampler: 16,
+            Dimension: dimension,
+            IsArray: false,
+            Glc: false,
+            Slc: false,
+            A16: false,
+            D16: false);
+        var imageInstruction = new Gen5ShaderInstruction(
+            0,
+            Gen5ShaderEncoding.Mimg,
+            opcode,
+            [],
+            [],
+            [],
+            control);
+        var end = new Gen5ShaderInstruction(
+            8,
+            Gen5ShaderEncoding.Sopp,
+            "SEndpgm",
+            [SEndpgm],
+            [],
+            [],
+            null);
+        var state = new Gen5ShaderState(
+            new Gen5ShaderProgram(ShaderAddress, [imageInstruction, end]),
+            [],
+            null);
+        var scalarRegisters = new uint[256];
+        var descriptor = new uint[8];
+        descriptor[1] = 71u << 20; // FORMAT_16_16_16_16_FLOAT
+        descriptor[3] = (dimension == 2 ? 10u : 9u) << 28;
+        var evaluation = new Gen5ShaderEvaluation(
+            scalarRegisters,
+            scalarRegisters,
+            [
+                new Gen5ImageBinding(
+                    imageInstruction.Pc,
+                    imageInstruction.Opcode,
+                    control,
+                    descriptor,
+                    new uint[4],
+                    null),
+            ],
+            []);
+
+        Assert.True(
+            Gen5SpirvTranslator.TryCompileComputeShader(
+                state,
+                evaluation,
+                1,
+                1,
+                1,
+                out var shader,
+                out var error),
+            error);
+        return shader.Spirv;
+    }
+
+    private static IReadOnlyList<ParsedSpirvInstruction> ReadSpirvInstructions(
+        byte[] spirv)
+    {
+        var instructions = new List<ParsedSpirvInstruction>();
+        for (var offset = 5 * sizeof(uint); offset < spirv.Length;)
+        {
+            var instruction = BinaryPrimitives.ReadUInt32LittleEndian(
+                spirv.AsSpan(offset));
+            var wordCount = checked((int)(instruction >> 16));
+            Assert.InRange(wordCount, 1, (spirv.Length - offset) / sizeof(uint));
+            var operands = new uint[wordCount - 1];
+            for (var operand = 0; operand < operands.Length; operand++)
+            {
+                operands[operand] = BinaryPrimitives.ReadUInt32LittleEndian(
+                    spirv.AsSpan(offset + (operand + 1) * sizeof(uint)));
+            }
+
+            instructions.Add(
+                new ParsedSpirvInstruction((SpirvOp)(ushort)instruction, operands));
+            offset += wordCount * sizeof(uint);
+        }
+
+        return instructions;
+    }
+
+    private static void AssertCoordinateVectorWidth(
+        IReadOnlyList<ParsedSpirvInstruction> instructions,
+        SpirvOp operation,
+        int coordinateOperand,
+        uint expectedComponents)
+    {
+        var imageOperation = Assert.Single(
+            instructions,
+            item => item.Opcode == operation);
+        var coordinateId = imageOperation.Operands[coordinateOperand];
+        var coordinate = Assert.Single(
+            instructions,
+            item =>
+                item.Opcode == SpirvOp.CompositeConstruct &&
+                item.Operands.Length >= 2 &&
+                item.Operands[1] == coordinateId);
+        AssertVectorTypeWidth(
+            instructions,
+            coordinate.Operands[0],
+            expectedComponents);
+    }
+
+    private static void AssertVectorTypeWidth(
+        IReadOnlyList<ParsedSpirvInstruction> instructions,
+        uint vectorTypeId,
+        uint expectedComponents)
+    {
+        var vectorType = Assert.Single(
+            instructions,
+            item =>
+                item.Opcode == SpirvOp.TypeVector &&
+                item.Operands[0] == vectorTypeId);
+        Assert.Equal(expectedComponents, vectorType.Operands[2]);
+    }
+
+    private readonly record struct ParsedSpirvInstruction(
+        SpirvOp Opcode,
+        uint[] Operands);
+}

tests/SharpEmu.ShaderCompiler.Tests/SharpEmu.ShaderCompiler.Tests.csproj

@@ -0,0 +1,22 @@
+<!--
+Copyright (C) 2026 SharpEmu Emulator Project
+SPDX-License-Identifier: GPL-2.0-or-later
+-->
+
+<Project Sdk="Microsoft.NET.Sdk">
+  <PropertyGroup>
+    <IsPackable>false</IsPackable>
+    <GenerateDocumentationFile>false</GenerateDocumentationFile>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="..\..\src\SharpEmu.ShaderCompiler\SharpEmu.ShaderCompiler.csproj" />
+    <ProjectReference Include="..\..\src\SharpEmu.ShaderCompiler.Vulkan\SharpEmu.ShaderCompiler.Vulkan.csproj" />
+  </ItemGroup>
+
+  <ItemGroup>
+    <PackageReference Include="Microsoft.NET.Test.Sdk" />
+    <PackageReference Include="xunit" />
+    <PackageReference Include="xunit.runner.visualstudio" />
+  </ItemGroup>
+</Project>