Add spin lock to prevent waiting for fences on multiple threads at once on Adreno

Support ballot operations with divergent control flow on Adreno

Extend Adreno binding workaround to buffer textures

src/Ryujinx.Graphics.GAL/Capabilities.cs

@@ -40,6 +40,7 @@ namespace Ryujinx.Graphics.GAL
         public readonly bool SupportsQuads;
         public readonly bool SupportsSeparateSampler;
         public readonly bool SupportsShaderBallot;
+        public readonly bool SupportsShaderBallotDivergence;
         public readonly bool SupportsShaderBarrierDivergence;
         public readonly bool SupportsShaderFloat64;
         public readonly bool SupportsTextureGatherOffsets;
@@ -106,6 +107,7 @@ namespace Ryujinx.Graphics.GAL
             bool supportsQuads,
             bool supportsSeparateSampler,
             bool supportsShaderBallot,
+            bool supportsShaderBallotDivergence,
             bool supportsShaderBarrierDivergence,
             bool supportsShaderFloat64,
             bool supportsTextureGatherOffsets,
@@ -167,6 +169,7 @@ namespace Ryujinx.Graphics.GAL
             SupportsQuads = supportsQuads;
             SupportsSeparateSampler = supportsSeparateSampler;
             SupportsShaderBallot = supportsShaderBallot;
+            SupportsShaderBallotDivergence = supportsShaderBallotDivergence;
             SupportsShaderBarrierDivergence = supportsShaderBarrierDivergence;
             SupportsShaderFloat64 = supportsShaderFloat64;
             SupportsTextureGatherOffsets = supportsTextureGatherOffsets;

src/Ryujinx.Graphics.Gpu/Shader/GpuAccessorBase.cs

@@ -227,6 +227,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
         public bool QueryHostSupportsShaderBallot() => _context.Capabilities.SupportsShaderBallot;
 
+        public bool QueryHostSupportsShaderBallotDivergence() => _context.Capabilities.SupportsShaderBallotDivergence;
+
         public bool QueryHostSupportsShaderBarrierDivergence() => _context.Capabilities.SupportsShaderBarrierDivergence;
 
         public bool QueryHostSupportsShaderFloat64() => _context.Capabilities.SupportsShaderFloat64;

src/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -177,6 +177,7 @@ namespace Ryujinx.Graphics.OpenGL
                 supportsQuads: HwCapabilities.SupportsQuads,
                 supportsSeparateSampler: false,
                 supportsShaderBallot: HwCapabilities.SupportsShaderBallot,
+                supportsShaderBallotDivergence: true,
                 supportsShaderBarrierDivergence: !(intelWindows || intelUnix),
                 supportsShaderFloat64: true,
                 supportsTextureGatherOffsets: true,

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs

@@ -227,14 +227,44 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
         private static OperationResult GenerateBallot(CodeGenContext context, AstOperation operation)
         {
             var source = operation.GetSource(0);
+            var predicate = context.Get(AggregateType.Bool, source);
 
-            var uvec4Type = context.TypeVector(context.TypeU32(), 4);
-            var execution = context.Constant(context.TypeU32(), Scope.Subgroup);
+            if (!context.HostCapabilities.SupportsShaderBallotDivergence &&
+                (context.CurrentBlock.Type != AstBlockType.Main || context.MayHaveReturned || !context.IsMainFunction))
+            {
+                // If divergent ballot is not supported, we can emulate it with a subgroupAdd operation,
+                // where we add a bit mask with a unique bit set for each subgroup invocation.
 
-            var maskVector = context.GroupNonUniformBallot(uvec4Type, execution, context.Get(AggregateType.Bool, source));
-            var mask = context.CompositeExtract(context.TypeU32(), maskVector, (SpvLiteralInteger)operation.Index);
+                var bit = context.Select(
+                    context.TypeU32(),
+                    predicate,
+                    context.Constant(context.TypeU32(), 1),
+                    context.Constant(context.TypeU32(), 0));
 
-            return new OperationResult(AggregateType.U32, mask);
+                var threadId = GetScalarInput(context, IoVariable.SubgroupLaneId);
+                var threadIdLow = context.BitwiseAnd(context.TypeU32(), threadId, context.Constant(context.TypeU32(), 0x1f));
+                var threadIdHigh = context.ShiftRightLogical(context.TypeU32(), threadId, context.Constant(context.TypeU32(), 5));
+                var bitMask = context.ShiftLeftLogical(context.TypeU32(), bit, threadIdLow);
+                var isGroup = context.IEqual(context.TypeBool(), threadIdHigh, context.Constant(context.TypeU32(), operation.Index));
+                bitMask = context.Select(context.TypeU32(), isGroup, bitMask, context.Constant(context.TypeU32(), 0));
+                var mask = context.GroupNonUniformIAdd(
+                    context.TypeU32(),
+                    context.Constant(context.TypeU32(), Scope.Subgroup),
+                    GroupOperation.Reduce,
+                    bitMask);
+
+                return new OperationResult(AggregateType.U32, mask);
+            }
+            else
+            {
+                var uvec4Type = context.TypeVector(context.TypeU32(), 4);
+                var execution = context.Constant(context.TypeU32(), Scope.Subgroup);
+
+                var maskVector = context.GroupNonUniformBallot(uvec4Type, execution, predicate);
+                var mask = context.CompositeExtract(context.TypeU32(), maskVector, (SpvLiteralInteger)operation.Index);
+
+                return new OperationResult(AggregateType.U32, mask);
+            }
         }
 
         private static OperationResult GenerateBarrier(CodeGenContext context, AstOperation operation)

src/Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvGenerator.cs

@@ -28,7 +28,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             _poolLock = new object();
         }
 
-        private const HelperFunctionsMask NeedsInvocationIdMask = HelperFunctionsMask.SwizzleAdd;
+        private const HelperFunctionsMask NeedsInvocationIdMask = HelperFunctionsMask.SwizzleAdd | HelperFunctionsMask.Ballot;
 
         public static byte[] Generate(StructuredProgramInfo info, CodeGenParameters parameters)
         {
@@ -60,6 +60,12 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                 context.AddCapability(Capability.Float64);
             }
 
+            if (info.HelperFunctionsMask.HasFlag(HelperFunctionsMask.Ballot) && !context.HostCapabilities.SupportsShaderBallotDivergence)
+            {
+                // Ballots might be emulated with subgroupAdd in those cases.
+                context.AddCapability(Capability.GroupNonUniformArithmetic);
+            }
+
             if (parameters.Definitions.TransformFeedbackEnabled && parameters.Definitions.LastInVertexPipeline)
             {
                 context.AddCapability(Capability.TransformFeedback);

src/Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -319,6 +319,15 @@ namespace Ryujinx.Graphics.Shader
             return true;
         }
 
+        /// <summary>
+        /// Queries host GPU shader support for ballot instructions on divergent control flow paths.
+        /// </summary>
+        /// <returns>True if the GPU supports ballot instructions on divergent control flow paths, false otherwise</returns>
+        bool QueryHostSupportsShaderBallotDivergence()
+        {
+            return true;
+        }
+
         /// <summary>
         /// Queries host GPU shader support for barrier instructions on divergent control flow paths.
         /// </summary>

src/Ryujinx.Graphics.Shader/StructuredIr/HelperFunctionsMask.cs

@@ -9,5 +9,6 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
         MultiplyHighU32 = 1 << 3,
         SwizzleAdd = 1 << 10,
         FSI = 1 << 11,
+        Ballot = 1 << 12,
     }
 }

src/Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs

@@ -338,6 +338,9 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
                 case Instruction.FSIEnd:
                     context.Info.HelperFunctionsMask |= HelperFunctionsMask.FSI;
                     break;
+                case Instruction.Ballot:
+                    context.Info.HelperFunctionsMask |= HelperFunctionsMask.Ballot;
+                    break;
             }
         }
 

src/Ryujinx.Graphics.Shader/Translation/HostCapabilities.cs

@@ -7,6 +7,7 @@ namespace Ryujinx.Graphics.Shader.Translation
         public readonly bool SupportsFragmentShaderOrderingIntel;
         public readonly bool SupportsGeometryShaderPassthrough;
         public readonly bool SupportsShaderBallot;
+        public readonly bool SupportsShaderBallotDivergence;
         public readonly bool SupportsShaderBarrierDivergence;
         public readonly bool SupportsShaderFloat64;
         public readonly bool SupportsTextureShadowLod;
@@ -18,6 +19,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             bool supportsFragmentShaderOrderingIntel,
             bool supportsGeometryShaderPassthrough,
             bool supportsShaderBallot,
+            bool supportsShaderBallotDivergence,
             bool supportsShaderBarrierDivergence,
             bool supportsShaderFloat64,
             bool supportsTextureShadowLod,
@@ -28,6 +30,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             SupportsFragmentShaderOrderingIntel = supportsFragmentShaderOrderingIntel;
             SupportsGeometryShaderPassthrough = supportsGeometryShaderPassthrough;
             SupportsShaderBallot = supportsShaderBallot;
+            SupportsShaderBallotDivergence = supportsShaderBallotDivergence;
             SupportsShaderBarrierDivergence = supportsShaderBarrierDivergence;
             SupportsShaderFloat64 = supportsShaderFloat64;
             SupportsTextureShadowLod = supportsTextureShadowLod;

src/Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs

@@ -362,6 +362,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                 GpuAccessor.QueryHostSupportsFragmentShaderOrderingIntel(),
                 GpuAccessor.QueryHostSupportsGeometryShaderPassthrough(),
                 GpuAccessor.QueryHostSupportsShaderBallot(),
+                GpuAccessor.QueryHostSupportsShaderBallotDivergence(),
                 GpuAccessor.QueryHostSupportsShaderBarrierDivergence(),
                 GpuAccessor.QueryHostSupportsShaderFloat64(),
                 GpuAccessor.QueryHostSupportsTextureShadowLod(),

src/Ryujinx.Graphics.Vulkan/ShaderCollection.cs

@@ -1,4 +1,4 @@
-using Ryujinx.Common.Logging;
+using Ryujinx.Common.Logging;
 using Ryujinx.Graphics.GAL;
 using Silk.NET.Vulkan;
 using System;
@@ -290,7 +290,7 @@ namespace Ryujinx.Graphics.Vulkan
             return segments;
         }
 
-        private static ResourceBindingSegment[][] BuildBindingSegments(ReadOnlyCollection<ResourceUsageCollection> setUsages, bool hasBatchedTextureSamplerBug, out bool usesBufferTextures)
+        private static ResourceBindingSegment[][] BuildBindingSegments(ReadOnlyCollection<ResourceUsageCollection> setUsages, bool hasBatchedTextureBug, out bool usesBufferTextures)
         {
             usesBufferTextures = false;
 
@@ -314,7 +314,7 @@ namespace Ryujinx.Graphics.Vulkan
 
                     if (currentUsage.Binding + currentCount != usage.Binding ||
                         currentUsage.Type != usage.Type ||
-                        (currentUsage.Type == ResourceType.TextureAndSampler && hasBatchedTextureSamplerBug) ||
+                        (IsReadOnlyTexture(currentUsage.Type) && hasBatchedTextureBug) ||
                         currentUsage.Stages != usage.Stages ||
                         currentUsage.ArrayLength > 1 ||
                         usage.ArrayLength > 1)
@@ -450,6 +450,12 @@ namespace Ryujinx.Graphics.Vulkan
             return (buffer, texture);
         }
 
+        private static bool IsReadOnlyTexture(ResourceType resourceType)
+        {
+            return resourceType == ResourceType.TextureAndSampler || resourceType == ResourceType.BufferTexture;
+
+        }
+
         private async Task BackgroundCompilation()
         {
             await Task.WhenAll(_shaders.Select(shader => shader.CompileTask));

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -758,6 +758,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportsQuads: false,
                 supportsSeparateSampler: true,
                 supportsShaderBallot: false,
+                supportsShaderBallotDivergence: Vendor != Vendor.Qualcomm,
                 supportsShaderBarrierDivergence: Vendor != Vendor.Intel,
                 supportsShaderFloat64: Capabilities.SupportsShaderFloat64,
                 supportsTextureGatherOffsets: features2.Features.ShaderImageGatherExtended && !IsMoltenVk,