Use draw clear on Adreno, instead of vkCmdClearAttachments (#7013)

* Use draw clear on Adreno, instead of vkCmdClearAttachments

* Fix GTX TITAN detection

src/Ryujinx.Graphics.Vulkan/BackgroundResources.cs

@@ -29,7 +29,14 @@ namespace Ryujinx.Graphics.Vulkan
 
                 lock (queueLock)
                 {
-                    _pool = new CommandBufferPool(_gd.Api, _device, queue, queueLock, _gd.QueueFamilyIndex, isLight: true);
+                    _pool = new CommandBufferPool(
+                        _gd.Api,
+                        _device,
+                        queue,
+                        queueLock,
+                        _gd.QueueFamilyIndex,
+                        _gd.IsQualcommProprietary,
+                        isLight: true);
                 }
             }
 

src/Ryujinx.Graphics.Vulkan/CommandBufferPool.cs

@@ -18,6 +18,7 @@ namespace Ryujinx.Graphics.Vulkan
         private readonly Device _device;
         private readonly Queue _queue;
         private readonly object _queueLock;
+        private readonly bool _concurrentFenceWaitUnsupported;
         private readonly CommandPool _pool;
         private readonly Thread _owner;
 
@@ -61,12 +62,20 @@ namespace Ryujinx.Graphics.Vulkan
         private int _queuedCount;
         private int _inUseCount;
 
-        public unsafe CommandBufferPool(Vk api, Device device, Queue queue, object queueLock, uint queueFamilyIndex, bool isLight = false)
+        public unsafe CommandBufferPool(
+            Vk api,
+            Device device,
+            Queue queue,
+            object queueLock,
+            uint queueFamilyIndex,
+            bool concurrentFenceWaitUnsupported,
+            bool isLight = false)
         {
             _api = api;
             _device = device;
             _queue = queue;
             _queueLock = queueLock;
+            _concurrentFenceWaitUnsupported = concurrentFenceWaitUnsupported;
             _owner = Thread.CurrentThread;
 
             var commandPoolCreateInfo = new CommandPoolCreateInfo
@@ -357,7 +366,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             if (refreshFence)
             {
-                entry.Fence = new FenceHolder(_api, _device);
+                entry.Fence = new FenceHolder(_api, _device, _concurrentFenceWaitUnsupported);
             }
             else
             {

src/Ryujinx.Graphics.Vulkan/FenceHolder.cs

@@ -10,12 +10,15 @@ namespace Ryujinx.Graphics.Vulkan
         private readonly Device _device;
         private Fence _fence;
         private int _referenceCount;
+        private int _lock;
+        private readonly bool _concurrentWaitUnsupported;
         private bool _disposed;
 
-        public unsafe FenceHolder(Vk api, Device device)
+        public unsafe FenceHolder(Vk api, Device device, bool concurrentWaitUnsupported)
         {
             _api = api;
             _device = device;
+            _concurrentWaitUnsupported = concurrentWaitUnsupported;
 
             var fenceCreateInfo = new FenceCreateInfo
             {
@@ -47,6 +50,11 @@ namespace Ryujinx.Graphics.Vulkan
             }
             while (Interlocked.CompareExchange(ref _referenceCount, lastValue + 1, lastValue) != lastValue);
 
+            if (_concurrentWaitUnsupported)
+            {
+                AcquireLock();
+            }
+
             fence = _fence;
             return true;
         }
@@ -57,6 +65,16 @@ namespace Ryujinx.Graphics.Vulkan
             return _fence;
         }
 
+        public void PutLock()
+        {
+            Put();
+
+            if (_concurrentWaitUnsupported)
+            {
+                ReleaseLock();
+            }
+        }
+
         public void Put()
         {
             if (Interlocked.Decrement(ref _referenceCount) == 0)
@@ -66,24 +84,67 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        private void AcquireLock()
+        {
+            while (!TryAcquireLock())
+            {
+                Thread.SpinWait(32);
+            }
+        }
+
+        private bool TryAcquireLock()
+        {
+            return Interlocked.Exchange(ref _lock, 1) == 0;
+        }
+
+        private void ReleaseLock()
+        {
+            Interlocked.Exchange(ref _lock, 0);
+        }
+
         public void Wait()
         {
-            Span<Fence> fences = stackalloc Fence[]
+            if (_concurrentWaitUnsupported)
             {
-                _fence,
-            };
+                AcquireLock();
 
-            FenceHelper.WaitAllIndefinitely(_api, _device, fences);
+                try
+                {
+                    FenceHelper.WaitAllIndefinitely(_api, _device, stackalloc Fence[] { _fence });
+                }
+                finally
+                {
+                    ReleaseLock();
+                }
+            }
+            else
+            {
+                FenceHelper.WaitAllIndefinitely(_api, _device, stackalloc Fence[] { _fence });
+            }
         }
 
         public bool IsSignaled()
         {
-            Span<Fence> fences = stackalloc Fence[]
+            if (_concurrentWaitUnsupported)
             {
-                _fence,
-            };
+                if (!TryAcquireLock())
+                {
+                    return false;
+                }
 
-            return FenceHelper.AllSignaled(_api, _device, fences);
+                try
+                {
+                    return FenceHelper.AllSignaled(_api, _device, stackalloc Fence[] { _fence });
+                }
+                finally
+                {
+                    ReleaseLock();
+                }
+            }
+            else
+            {
+                return FenceHelper.AllSignaled(_api, _device, stackalloc Fence[] { _fence });
+            }
         }
 
         public void Dispose()

src/Ryujinx.Graphics.Vulkan/MultiFenceHolder.cs

@@ -196,18 +196,23 @@ namespace Ryujinx.Graphics.Vulkan
 
             bool signaled = true;
 
-            if (hasTimeout)
+            try
             {
-                signaled = FenceHelper.AllSignaled(api, device, fences[..fenceCount], timeout);
+                if (hasTimeout)
+                {
+                    signaled = FenceHelper.AllSignaled(api, device, fences[..fenceCount], timeout);
+                }
+                else
+                {
+                    FenceHelper.WaitAllIndefinitely(api, device, fences[..fenceCount]);
+                }
             }
-            else
+            finally
             {
-                FenceHelper.WaitAllIndefinitely(api, device, fences[..fenceCount]);
-            }
-
-            for (int i = 0; i < fenceCount; i++)
-            {
-                fenceHolders[i].Put();
+                for (int i = 0; i < fenceCount; i++)
+                {
+                    fenceHolders[i].PutLock();
+                }
             }
 
             return signaled;

src/Ryujinx.Graphics.Vulkan/PipelineBase.cs

@@ -1020,6 +1020,13 @@ namespace Ryujinx.Graphics.Vulkan
         {
             _newState.RasterizerDiscardEnable = discard;
             SignalStateChange();
+
+            if (!discard && Gd.IsQualcommProprietary)
+            {
+                // On Adreno, enabling rasterizer discard somehow corrupts the viewport state.
+                // Force it to be updated on next use to work around this bug.
+                DynamicState.ForceAllDirty();
+            }
         }
 
         public void SetRenderTargetColorMasks(ReadOnlySpan<uint> componentMask)

src/Ryujinx.Graphics.Vulkan/PipelineFull.cs

@@ -47,10 +47,11 @@ namespace Ryujinx.Graphics.Vulkan
                 return;
             }
 
-            if (componentMask != 0xf)
+            if (componentMask != 0xf || Gd.IsQualcommProprietary)
             {
                 // We can't use CmdClearAttachments if not writing all components,
                 // because on Vulkan, the pipeline state does not affect clears.
+                // On proprietary Adreno drivers, CmdClearAttachments appears to execute out of order, so it's better to not use it at all.
                 var dstTexture = FramebufferParams.GetColorView(index);
                 if (dstTexture == null)
                 {
@@ -87,10 +88,11 @@ namespace Ryujinx.Graphics.Vulkan
                 return;
             }
 
-            if (stencilMask != 0 && stencilMask != 0xff)
+            if ((stencilMask != 0 && stencilMask != 0xff) || Gd.IsQualcommProprietary)
             {
                 // We can't use CmdClearAttachments if not clearing all (mask is all ones, 0xFF) or none (mask is 0) of the stencil bits,
                 // because on Vulkan, the pipeline state does not affect clears.
+                // On proprietary Adreno drivers, CmdClearAttachments appears to execute out of order, so it's better to not use it at all.
                 var dstTexture = FramebufferParams.GetDepthStencilView();
                 if (dstTexture == null)
                 {

src/Ryujinx.Graphics.Vulkan/ShaderCollection.cs

@@ -133,7 +133,7 @@ namespace Ryujinx.Graphics.Vulkan
             Templates = BuildTemplates(usePushDescriptors);
 
             // Updating buffer texture bindings using template updates crashes the Adreno driver on Windows.
-            UpdateTexturesWithoutTemplate = gd.Vendor == Vendor.Qualcomm && usesBufferTextures;
+            UpdateTexturesWithoutTemplate = gd.IsQualcommProprietary && usesBufferTextures;
 
             _compileTask = Task.CompletedTask;
             _firstBackgroundUse = false;

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -87,9 +87,11 @@ namespace Ryujinx.Graphics.Vulkan
         internal bool IsAmdGcn { get; private set; }
         internal bool IsNvidiaPreTuring { get; private set; }
         internal bool IsIntelArc { get; private set; }
+        internal bool IsQualcommProprietary { get; private set; }
         internal bool IsMoltenVk { get; private set; }
         internal bool IsTBDR { get; private set; }
         internal bool IsSharedMemory { get; private set; }
+
         public string GpuVendor { get; private set; }
         public string GpuDriver { get; private set; }
         public string GpuRenderer { get; private set; }
@@ -344,7 +346,7 @@ namespace Ryujinx.Graphics.Vulkan
                 {
                     IsNvidiaPreTuring = gpuNumber < 2000;
                 }
-                else if (GpuDriver.Contains("TITAN") && !GpuDriver.Contains("RTX"))
+                else if (GpuRenderer.Contains("TITAN") && !GpuRenderer.Contains("RTX"))
                 {
                     IsNvidiaPreTuring = true;
                 }
@@ -354,6 +356,8 @@ namespace Ryujinx.Graphics.Vulkan
                 IsIntelArc = GpuRenderer.StartsWith("Intel(R) Arc(TM)");
             }
 
+            IsQualcommProprietary = hasDriverProperties && driverProperties.DriverID == DriverId.QualcommProprietary;
+
             ulong minResourceAlignment = Math.Max(
                 Math.Max(
                     properties.Limits.MinStorageBufferOffsetAlignment,
@@ -411,7 +415,7 @@ namespace Ryujinx.Graphics.Vulkan
             Api.TryGetDeviceExtension(_instance.Instance, _device, out ExtExternalMemoryHost hostMemoryApi);
             HostMemoryAllocator = new HostMemoryAllocator(MemoryAllocator, Api, hostMemoryApi, _device);
 
-            CommandBufferPool = new CommandBufferPool(Api, _device, Queue, QueueLock, queueFamilyIndex);
+            CommandBufferPool = new CommandBufferPool(Api, _device, Queue, QueueLock, queueFamilyIndex, IsQualcommProprietary);
 
             PipelineLayoutCache = new PipelineLayoutCache();
 
@@ -688,7 +692,7 @@ namespace Ryujinx.Graphics.Vulkan
                 GpuVendor,
                 memoryType: memoryType,
                 hasFrontFacingBug: IsIntelWindows,
-                hasVectorIndexingBug: Vendor == Vendor.Qualcomm,
+                hasVectorIndexingBug: IsQualcommProprietary,
                 needsFragmentOutputSpecialization: IsMoltenVk,
                 reduceShaderPrecision: IsMoltenVk,
                 supportsAstcCompression: features2.Features.TextureCompressionAstcLdr && supportsAstcFormats,