[GPU] Host cached guest buffer (#649)

* fix(memory): reserve only large regions

* Potential fix for pull request finding

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/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";
+        var lazyPrimeState = reservedOnly ? ReserveRegion(actualAddress, alignedSize) : "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";
-            }
-        }
 
         _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.Libs/Agc/AgcExports.cs

@@ -8263,6 +8263,9 @@ public static partial class AgcExports
     private static bool IsGpuDetileBytesPerElement(int bytesPerElement) =>
         bytesPerElement is 4 or 8 or 16;
 
+    private static bool IsGpuDetileTextureType(uint type) =>
+        type != Gen5TextureType3D;
+
     private static bool TryGetTextureElementLayout(
         TextureDescriptor descriptor,
         uint sourceWidth,
@@ -8746,6 +8749,7 @@ public static partial class AgcExports
             // take this path.
             if (_gpuDetileEnabled && hasElementLayout && !baseMipInTail &&
                 IsGpuDetileBytesPerElement(bytesPerElement) &&
+                IsGpuDetileTextureType(descriptor.Type) &&
                 (long)physicalSourceByteCount * arrayLayers <= int.MaxValue)
             {
                 var gpuArrayParams = GnmTiling.GetDetileParams(
@@ -8789,6 +8793,12 @@ public static partial class AgcExports
                             WriteGeneration: hasWriteGeneration ? writeGeneration : -1,
                             ArrayedView: true,
                             ArrayLayers: arrayLayers,
+                            // Must match the identity the CPU path below ships, or
+                            // the presenter caches this texture under a different
+                            // key than IsTextureContentCached queries above and the
+                            // texel-copy skip never hits for non-2D descriptors.
+                            Type: descriptor.Type,
+                            Depth: textureDepth,
                             TiledSource: tiledLayers,
                             Detile: gpuArrayParams);
                         return true;
@@ -8920,7 +8930,8 @@ public static partial class AgcExports
         // Arrayed textures are handled by the arrayed branch above (they package
         // every layer's tiled slice); this branch is the single-layer case.
         if (_gpuDetileEnabled && hasElementLayout && !baseMipInTail &&
-            IsGpuDetileBytesPerElement(bytesPerElement) && !isArrayed)
+            IsGpuDetileBytesPerElement(bytesPerElement) && !isArrayed &&
+            IsGpuDetileTextureType(descriptor.Type))
         {
             var gpuDetileParams = GnmTiling.GetDetileParams(
                 descriptor.TileMode, bytesPerElement, elementsWide, elementsHigh);
@@ -8946,6 +8957,8 @@ public static partial class AgcExports
                     Sampler: ToGuestSampler(samplerDescriptor),
                     WriteGeneration: hasWriteGeneration ? writeGeneration : -1,
                     ArrayedView: isArrayed,
+                    Type: descriptor.Type,
+                    Depth: textureDepth,
                     TiledSource: source,
                     Detile: gpuDetileParams);
                 return true;

src/SharpEmu.Libs/VideoOut/VulkanDetilePass.cs

@@ -2,6 +2,7 @@
 // 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;
@@ -45,6 +46,21 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
     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,
@@ -69,23 +85,32 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
          parameters.Equation == DetileEquation.BlockTable) &&
         parameters.BytesPerElement is 4 or 8 or 16;
 
-    /// <summary>Transient per-detile resources the caller must retire once the
+    /// <summary>Opaque handle to the pooled resources one recorded detile is using.
-    /// command buffer they were recorded into has completed.</summary>
+    /// The caller hands it back to <see cref="Retire"/> once the command buffer they
-    public readonly record struct Transients(
+    /// were recorded into has completed; nothing is destroyed, the buffers and the
-        (VkBuffer Buffer, DeviceMemory Memory)[] Buffers,
+    /// descriptor set return to this pass's free lists for the next texture.</summary>
-        DescriptorPool DescriptorPool);
+    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 VkBuffer Tiled;
+        public Allocation Tiled;
-        public DeviceMemory TiledMemory;
+        public Allocation Output;
-        public VkBuffer XTerm;
-        public DeviceMemory XMemory;
-        public VkBuffer YTerm;
-        public DeviceMemory YMemory;
-        public VkBuffer Output;
-        public DeviceMemory OutputMemory;
-        public DescriptorPool Pool;
         public DescriptorSet Set;
         public ulong OutputBytes;
         public uint SrcSliceElements;
@@ -93,6 +118,15 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         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
@@ -117,7 +151,7 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         in DetileParams parameters,
         out Transients transients)
     {
-        transients = new Transients([], default);
+        transients = Transients.Empty;
         if (_disposed || !Supports(parameters) || texelWidth == 0 || texelHeight == 0 || layers == 0 ||
             tiled.IsEmpty || tiled.Length % (int)(layers * (uint)parameters.BytesPerElement) != 0)
         {
@@ -134,21 +168,33 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         }
         catch
         {
-            DestroyResources(in resources);
+            ReleaseResources(in resources);
             throw;
         }
 
-        transients = new Transients(
+        transients = new Transients
-            [
+        {
-                (resources.Tiled, resources.TiledMemory),
+            Tiled = resources.Tiled,
-                (resources.XTerm, resources.XMemory),
+            Output = resources.Output,
-                (resources.YTerm, resources.YMemory),
+            Set = resources.Set,
-                (resources.Output, resources.OutputMemory),
+            Rented = true,
-            ],
+        };
-            resources.Pool);
         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
@@ -207,7 +253,7 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
                 _vk.FreeCommandBuffers(_device, _commandPool, 1, &commandBuffer);
             }
 
-            DestroyResources(in resources);
+            ReleaseResources(in resources);
         }
     }
 
@@ -219,24 +265,19 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         // 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.
-        uint[] xTerm;
+        TermBuffer xTerm;
-        uint[] yTerm;
+        TermBuffer yTerm;
         if (parameters.Equation == DetileEquation.BlockTable)
         {
-            xTerm = new uint[parameters.BlockTable.Length];
+            xTerm = GetTermBuffer(_blockTermBuffers, parameters.BlockTable, shift: 0);
-            for (var index = 0; index < xTerm.Length; index++)
+            yTerm = GetPlaceholderTermBuffer();
-            {
-                xTerm[index] = (uint)parameters.BlockTable[index];
-            }
-
-            yTerm = [0];
             resources.EquationValue = 1;
         }
         else
         {
             var shift = BitOperations.TrailingZeroCount((uint)parameters.BytesPerElement);
-            xTerm = ToElementTerms(parameters.XByteTerm, shift);
+            xTerm = GetTermBuffer(_xorTermBuffers, parameters.XByteTerm, shift);
-            yTerm = ToElementTerms(parameters.YByteTerm, shift);
+            yTerm = GetTermBuffer(_xorTermBuffers, parameters.YByteTerm, shift);
             resources.EquationValue = 0;
         }
 
@@ -249,25 +290,144 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         resources.OutputBytes =
             (ulong)parameters.ElementsWide * (ulong)parameters.ElementsHigh * bytesPerElement * layers;
 
-        resources.Tiled = CreateHostBuffer((ulong)tiled.Length, BufferUsageFlags.StorageBufferBit, out resources.TiledMemory);
+        resources.Tiled = RentBuffer((ulong)tiled.Length, hostVisible: true);
-        UploadBytes(resources.TiledMemory, tiled);
+        UploadBytes(resources.Tiled.Memory, tiled);
-        resources.XTerm = CreateHostBuffer((ulong)xTerm.Length * sizeof(uint), BufferUsageFlags.StorageBufferBit, out resources.XMemory);
+        resources.Output = RentBuffer(resources.OutputBytes, hostVisible: false);
-        UploadUInts(resources.XMemory, xTerm);
-        resources.YTerm = CreateHostBuffer((ulong)yTerm.Length * sizeof(uint), BufferUsageFlags.StorageBufferBit, out resources.YMemory);
-        UploadUInts(resources.YMemory, yTerm);
-        resources.Output = CreateHostBuffer(
-            resources.OutputBytes,
-            BufferUsageFlags.StorageBufferBit | BufferUsageFlags.TransferSrcBit,
-            out resources.OutputMemory);
 
-        resources.Pool = CreateDescriptorPool();
+        resources.Set = RentDescriptorSet();
-        resources.Set = AllocateDescriptorSet(resources.Pool);
         WriteDescriptors(
             resources.Set,
-            (resources.Tiled, (ulong)tiled.Length),
+            (resources.Tiled.Buffer, (ulong)tiled.Length),
-            (resources.XTerm, (ulong)xTerm.Length * sizeof(uint)),
+            (xTerm.Buffer, xTerm.ByteSize),
-            (resources.YTerm, (ulong)yTerm.Length * sizeof(uint)),
+            (yTerm.Buffer, yTerm.ByteSize),
-            (resources.Output, resources.OutputBytes));
+            (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(
@@ -326,7 +486,7 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
             DstAccessMask = AccessFlags.TransferReadBit,
             SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
             DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
-            Buffer = resources.Output,
+            Buffer = resources.Output.Buffer,
             Offset = 0,
             Size = resources.OutputBytes,
         };
@@ -367,7 +527,7 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
             ImageExtent = new Extent3D(texelWidth, texelHeight, 1),
         };
         _vk.CmdCopyBufferToImage(
-            commandBuffer, resources.Output, image, ImageLayout.TransferDstOptimal, 1, &copyRegion);
+            commandBuffer, resources.Output.Buffer, image, ImageLayout.TransferDstOptimal, 1, &copyRegion);
 
         TransitionImage(
             commandBuffer,
@@ -381,19 +541,6 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
             layers);
     }
 
-    private void DestroyResources(in DetileResources resources)
-    {
-        if (resources.Pool.Handle != 0)
-        {
-            _vk.DestroyDescriptorPool(_device, resources.Pool, null);
-        }
-
-        DestroyBuffer(resources.Output, resources.OutputMemory);
-        DestroyBuffer(resources.YTerm, resources.YMemory);
-        DestroyBuffer(resources.XTerm, resources.XMemory);
-        DestroyBuffer(resources.Tiled, resources.TiledMemory);
-    }
-
     private void EnsurePipeline()
     {
         if (_initialized)
@@ -500,45 +647,67 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         return terms;
     }
 
-    private VkBuffer CreateHostBuffer(ulong size, BufferUsageFlags usage, out DeviceMemory memory)
+    private Allocation CreateBuffer(ulong size, bool hostVisible)
     {
         var bufferInfo = new BufferCreateInfo
         {
             SType = StructureType.BufferCreateInfo,
             Size = size,
-            Usage = usage,
+            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(
+            MemoryTypeIndex = FindMemoryType(requirements.MemoryTypeBits, required, hostVisible),
-                requirements.MemoryTypeBits,
-                MemoryPropertyFlags.HostVisibleBit | MemoryPropertyFlags.HostCoherentBit),
         };
-        Check(_vk.AllocateMemory(_device, &allocateInfo, null, out memory), "vkAllocateMemory(detile)");
+        Check(_vk.AllocateMemory(_device, &allocateInfo, null, out var memory), "vkAllocateMemory(detile)");
         Check(_vk.BindBufferMemory(_device, buffer, memory, 0), "vkBindBufferMemory(detile)");
-        return buffer;
+        var allocation = new Allocation(buffer, memory, size, hostVisible);
+        _allAllocations.Add(allocation);
+        return allocation;
     }
 
-    private uint FindMemoryType(uint typeBits, MemoryPropertyFlags requiredFlags)
+    private uint FindMemoryType(uint typeBits, MemoryPropertyFlags requiredFlags, bool hostVisible)
     {
-        _vk.GetPhysicalDeviceMemoryProperties(_physicalDevice, out var properties);
+        if (!_memoryPropertiesLoaded)
-        var memoryTypes = &properties.MemoryTypes.Element0;
-        for (uint index = 0; index < properties.MemoryTypeCount; index++)
         {
-            if ((typeBits & (1u << (int)index)) != 0 &&
+            _vk.GetPhysicalDeviceMemoryProperties(_physicalDevice, out _memoryProperties);
-                (memoryTypes[index].PropertyFlags & requiredFlags) == requiredFlags)
+            _memoryPropertiesLoaded = true;
+        }
+
+        fixed (PhysicalDeviceMemoryProperties* properties = &_memoryProperties)
+        {
+            var memoryTypes = &properties->MemoryTypes.Element0;
+            for (uint index = 0; index < properties->MemoryTypeCount; index++)
             {
-                return 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 host-visible memory type for detile.");
+        throw new InvalidOperationException("No compatible Vulkan memory type for detile.");
     }
 
     private void UploadBytes(DeviceMemory memory, ReadOnlySpan<byte> data)
@@ -558,42 +727,6 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         _vk.UnmapMemory(_device, memory);
     }
 
-    private DescriptorPool CreateDescriptorPool()
-    {
-        var poolSize = new DescriptorPoolSize
-        {
-            Type = DescriptorType.StorageBuffer,
-            DescriptorCount = 4,
-        };
-        var poolInfo = new DescriptorPoolCreateInfo
-        {
-            SType = StructureType.DescriptorPoolCreateInfo,
-            MaxSets = 1,
-            PoolSizeCount = 1,
-            PPoolSizes = &poolSize,
-        };
-        Check(
-            _vk.CreateDescriptorPool(_device, &poolInfo, null, out var pool),
-            "vkCreateDescriptorPool(detile)");
-        return pool;
-    }
-
-    private DescriptorSet AllocateDescriptorSet(DescriptorPool pool)
-    {
-        var setLayout = _descriptorSetLayout;
-        var allocateInfo = new DescriptorSetAllocateInfo
-        {
-            SType = StructureType.DescriptorSetAllocateInfo,
-            DescriptorPool = pool,
-            DescriptorSetCount = 1,
-            PSetLayouts = &setLayout,
-        };
-        Check(
-            _vk.AllocateDescriptorSets(_device, &allocateInfo, out var descriptorSet),
-            "vkAllocateDescriptorSets(detile)");
-        return descriptorSet;
-    }
-
     private void WriteDescriptors(
         DescriptorSet descriptorSet,
         (VkBuffer Buffer, ulong Size) binding0,
@@ -714,6 +847,25 @@ internal sealed unsafe class VulkanDetilePass : IDisposable
         }
 
         _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);

src/SharpEmu.Libs/VideoOut/VulkanDetileSelfTest.cs

@@ -261,23 +261,7 @@ internal static unsafe class VulkanDetileSelfTest
         {
             vk.DestroyFence(device, fence, null);
             vk.FreeCommandBuffers(device, commandPool, 1, &commandBuffer);
-            foreach (var (buffer, memory) in transients.Buffers)
+            pass.Retire(transients);
-            {
-                if (buffer.Handle != 0)
-                {
-                    vk.DestroyBuffer(device, buffer, null);
-                }
-
-                if (memory.Handle != 0)
-                {
-                    vk.FreeMemory(device, memory, null);
-                }
-            }
-
-            if (transients.DescriptorPool.Handle != 0)
-            {
-                vk.DestroyDescriptorPool(device, transients.DescriptorPool, null);
-            }
         }
 
         return true;

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.cs

@@ -2914,9 +2914,7 @@ internal static unsafe class VulkanVideoPresenter
         private readonly Stack<Fence> _recycledGuestFences = new();
         private readonly Stack<CommandBuffer> _recycledGuestCommandBuffers = new();
         private readonly List<(VkBuffer Buffer, DeviceMemory Memory)> _batchRetireBuffers = new();
-        // Descriptor pools from GPU-detile passes recorded into the batch; retired
+        private readonly List<VulkanDetilePass.Transients> _batchRetireDetile = new();
-        // with the batch's fence, alongside _batchRetireBuffers.
-        private readonly List<DescriptorPool> _batchRetireDescriptorPools = new();
         private const int MaxRecycledGuestFences = 32;
         private const int MaxRecycledGuestCommandBuffers = 32;
         private VkBuffer _stagingBuffer;
@@ -3288,7 +3286,7 @@ internal static unsafe class VulkanVideoPresenter
             IReadOnlyList<TranslatedDrawResources> Resources,
             IReadOnlyList<GuestImageResource> TraceImages,
             IReadOnlyList<(VkBuffer Buffer, DeviceMemory Memory)> RetireBuffers,
-            IReadOnlyList<DescriptorPool> RetirePools,
+            IReadOnlyList<VulkanDetilePass.Transients> RetireDetile,
             ulong Timeline,
             string DebugName,
             VulkanGuestQueueIdentity Queue,
@@ -4963,8 +4961,8 @@ internal static unsafe class VulkanVideoPresenter
                     _batchResources.ToArray(),
                     _batchTraceImages.ToArray(),
                     _batchRetireBuffers.Count > 0 ? _batchRetireBuffers.ToArray() : [],
-                    retirePools: _batchRetireDescriptorPools.Count > 0
+                    retireDetile: _batchRetireDetile.Count > 0
-                        ? _batchRetireDescriptorPools.ToArray()
+                        ? _batchRetireDetile.ToArray()
                         : []);
             }
             catch
@@ -4983,9 +4981,9 @@ internal static unsafe class VulkanVideoPresenter
                     _vk.FreeMemory(_device, memory, null);
                 }
 
-                foreach (var pool in _batchRetireDescriptorPools)
+                foreach (var transients in _batchRetireDetile)
                 {
-                    _vk.DestroyDescriptorPool(_device, pool, null);
+                    _detilePass?.Retire(transients);
                 }
 
                 ReleaseGuestCommandBuffer(_batchCommandBuffer);
@@ -4996,7 +4994,7 @@ internal static unsafe class VulkanVideoPresenter
                 _batchResources.Clear();
                 _batchTraceImages.Clear();
                 _batchRetireBuffers.Clear();
-                _batchRetireDescriptorPools.Clear();
+                _batchRetireDetile.Clear();
                 _batchCommandBuffer = default;
             }
         }
@@ -5007,7 +5005,7 @@ internal static unsafe class VulkanVideoPresenter
             IReadOnlyList<GuestImageResource> traceImages,
             IReadOnlyList<(VkBuffer Buffer, DeviceMemory Memory)>? retireBuffers = null,
             IReadOnlyList<TranslatedDrawResources>? referencedResources = null,
-            IReadOnlyList<DescriptorPool>? retirePools = null)
+            IReadOnlyList<VulkanDetilePass.Transients>? retireDetile = null)
         {
             var fence = AcquireGuestFence();
             try
@@ -5065,7 +5063,7 @@ internal static unsafe class VulkanVideoPresenter
                     resources,
                     traceImages,
                     retireBuffers ?? [],
-                    retirePools ?? [],
+                    retireDetile ?? [],
                     _submitTimeline,
                     resources.Count > 0 ? resources[0].DebugName : "batch",
                     _activeGuestQueue,
@@ -5237,9 +5235,11 @@ internal static unsafe class VulkanVideoPresenter
                     _vk.FreeMemory(_device, memory, null);
                 }
 
-                foreach (var pool in submission.RetirePools)
+                // The fence has signalled, so the detile dispatch that used these
+                // is done reading them; hand them back for the next texture.
+                foreach (var transients in submission.RetireDetile)
                 {
-                    _vk.DestroyDescriptorPool(_device, pool, null);
+                    _detilePass?.Retire(transients);
                 }
 
                 ReleaseGuestCommandBuffer(submission.CommandBuffer);
@@ -7855,7 +7855,8 @@ internal static unsafe class VulkanVideoPresenter
             // copy because this identity was marked cached; a miss here is
             // an invalidation race (eviction, cache clear). Self-heal by
             // reading the texels directly rather than rendering a fallback.
-            if (texture.RgbaPixels.Length == 0)
+            if (texture.RgbaPixels.Length == 0 &&
+                texture.TiledSource is not { Length: > 0 })
             {
                 var refreshed = TryReadGuestTexturePixels(texture);
                 if (refreshed is null)
@@ -8300,6 +8301,7 @@ internal static unsafe class VulkanVideoPresenter
                 texture.Detile is { } detileCandidate &&
                 texture.TiledSource is { Length: > 0 } tiledCandidate &&
                 VulkanDetilePass.Supports(detileCandidate) &&
+                !IsGuestTexture3D(texture.Type) &&
                 detileCandidate.ElementsWide > 0 &&
                 detileCandidate.ElementsHigh > 0 &&
                 (long)tiledCandidate.Length >=
@@ -8458,8 +8460,7 @@ internal static unsafe class VulkanVideoPresenter
                             detileParameters,
                             out var detileTransients))
                     {
-                        _batchRetireBuffers.AddRange(detileTransients.Buffers);
+                        _batchRetireDetile.Add(detileTransients);
-                        _batchRetireDescriptorPools.Add(detileTransients.DescriptorPool);
                         gpuDetiled = true;
                     }
                 }
@@ -9444,7 +9445,8 @@ internal static unsafe class VulkanVideoPresenter
                 size,
                 BufferUsageFlags.StorageBufferBit,
                 MemoryPropertyFlags.HostVisibleBit | MemoryPropertyFlags.HostCoherentBit,
-                out var memory);
+                out var memory,
+                preferredMemoryFlags: MemoryPropertyFlags.HostCachedBit);
             void* mapped;
             Check(_vk.MapMemory(_device, memory, 0, size, 0, &mapped), "vkMapMemory(guest buffer)");
             var shadow = new byte[checked((int)size)];
@@ -10253,7 +10255,8 @@ internal static unsafe class VulkanVideoPresenter
             ulong size,
             BufferUsageFlags usage,
             MemoryPropertyFlags memoryFlags,
-            out DeviceMemory memory)
+            out DeviceMemory memory,
+            MemoryPropertyFlags preferredMemoryFlags = 0)
         {
             var bufferInfo = new BufferCreateInfo
             {
@@ -10269,7 +10272,10 @@ internal static unsafe class VulkanVideoPresenter
             {
                 SType = StructureType.MemoryAllocateInfo,
                 AllocationSize = requirements.Size,
-                MemoryTypeIndex = FindMemoryType(requirements.MemoryTypeBits, memoryFlags),
+                MemoryTypeIndex = FindMemoryType(
+                    requirements.MemoryTypeBits,
+                    memoryFlags,
+                    preferredMemoryFlags),
             };
             Check(_vk.AllocateMemory(_device, &memoryInfo, null, out memory), "vkAllocateMemory");
             Check(_vk.BindBufferMemory(_device, buffer, memory, 0), "vkBindBufferMemory");
@@ -10306,16 +10312,24 @@ internal static unsafe class VulkanVideoPresenter
             }
         }
 
-        private uint FindMemoryType(uint typeBits, MemoryPropertyFlags requiredFlags)
+        private uint FindMemoryType(
+            uint typeBits,
+            MemoryPropertyFlags requiredFlags,
+            MemoryPropertyFlags preferredFlags = 0)
         {
             _vk.GetPhysicalDeviceMemoryProperties(_physicalDevice, out var properties);
             var memoryTypes = &properties.MemoryTypes.Element0;
-            for (uint index = 0; index < properties.MemoryTypeCount; index++)
+
+            for (var pass = preferredFlags != 0 ? 0 : 1; pass < 2; pass++)
             {
-                if ((typeBits & (1u << (int)index)) != 0 &&
+                var wanted = pass == 0 ? requiredFlags | preferredFlags : requiredFlags;
-                    (memoryTypes[index].PropertyFlags & requiredFlags) == requiredFlags)
+                for (uint index = 0; index < properties.MemoryTypeCount; index++)
                 {
-                    return index;
+                    if ((typeBits & (1u << (int)index)) != 0 &&
+                        (memoryTypes[index].PropertyFlags & wanted) == wanted)
+                    {
+                        return index;
+                    }
                 }
             }