Ensure shader local and shared memory sizes are not zero (#5321)

src/Ryujinx.Graphics.Shader/Decoders/Decoder.cs

@@ -247,6 +247,17 @@ namespace Ryujinx.Graphics.Shader.Decoders
                 {
                     block.AddPushOp(op);
                 }
+                else if (op.Name == InstName.Ldl || op.Name == InstName.Stl)
+                {
+                    config.SetUsedFeature(FeatureFlags.LocalMemory);
+                }
+                else if (op.Name == InstName.Atoms ||
+                         op.Name == InstName.AtomsCas ||
+                         op.Name == InstName.Lds ||
+                         op.Name == InstName.Sts)
+                {
+                    config.SetUsedFeature(FeatureFlags.SharedMemory);
+                }
 
                 block.OpCodes.Add(op);
 

src/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs

@@ -27,6 +27,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         public static void Atoms(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Atoms instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstAtoms op = context.GetOp<InstAtoms>();
 
             Operand offset = context.ShiftRightU32(GetSrcReg(context, op.SrcA), Const(2));
@@ -114,6 +120,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         public static void Lds(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Lds instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstLds op = context.GetOp<InstLds>();
 
             EmitLoad(context, StorageKind.SharedMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));
@@ -144,6 +156,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         public static void Sts(EmitterContext context)
         {
+            if (context.Config.Stage != ShaderStage.Compute)
+            {
+                context.Config.GpuAccessor.Log($"Sts instruction is not valid on \"{context.Config.Stage}\" stage.");
+                return;
+            }
+
             InstSts op = context.GetOp<InstSts>();
 
             EmitStore(context, StorageKind.SharedMemory, op.LsSize, GetSrcReg(context, op.SrcA), op.Dest, Imm24ToSInt(op.Imm24));

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

@@ -21,6 +21,8 @@ namespace Ryujinx.Graphics.Shader.Translation
         RtLayer = 1 << 5,
         IaIndexing = 1 << 7,
         OaIndexing = 1 << 8,
-        FixedFuncAttr = 1 << 9
+        FixedFuncAttr = 1 << 9,
+        LocalMemory = 1 << 10,
+        SharedMemory = 1 << 11
     }
 }

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

@@ -8,6 +8,11 @@ namespace Ryujinx.Graphics.Shader.Translation
 {
     class ResourceManager
     {
+        // Those values are used if the shader as local or shared memory access,
+        // but for some reason the supplied size was 0.
+        private const int DefaultLocalMemorySize = 128;
+        private const int DefaultSharedMemorySize = 4096;
+
         private static readonly string[] _stagePrefixes = new string[] { "cp", "vp", "tcp", "tep", "gp", "fp" };
 
         private readonly IGpuAccessor _gpuAccessor;
@@ -23,12 +28,12 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         private readonly HashSet<int> _usedConstantBufferBindings;
 
-        public int LocalMemoryId { get; }
+        public int LocalMemoryId { get; private set; }
-        public int SharedMemoryId { get; }
+        public int SharedMemoryId { get; private set; }
 
         public ShaderProperties Properties => _properties;
 
-        public ResourceManager(ShaderStage stage, IGpuAccessor gpuAccessor, ShaderProperties properties, int localMemorySize)
+        public ResourceManager(ShaderStage stage, IGpuAccessor gpuAccessor, ShaderProperties properties)
         {
             _gpuAccessor = gpuAccessor;
             _properties = properties;
@@ -48,21 +53,43 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             LocalMemoryId = -1;
             SharedMemoryId = -1;
+        }
 
-            if (localMemorySize != 0)
+        public void SetCurrentLocalMemory(int size, bool isUsed)
+        {
+            if (isUsed)
             {
-                var lmem = new MemoryDefinition("local_memory", AggregateType.Array | AggregateType.U32, BitUtils.DivRoundUp(localMemorySize, sizeof(uint)));
+                if (size <= 0)
+                {
+                    size = DefaultLocalMemorySize;
+                }
 
-                LocalMemoryId = properties.AddLocalMemory(lmem);
+                var lmem = new MemoryDefinition("local_memory", AggregateType.Array | AggregateType.U32, BitUtils.DivRoundUp(size, sizeof(uint)));
+
+                LocalMemoryId = Properties.AddLocalMemory(lmem);
             }
-
+            else
-            int sharedMemorySize = stage == ShaderStage.Compute ? gpuAccessor.QueryComputeSharedMemorySize() : 0;
-
-            if (sharedMemorySize != 0)
             {
-                var smem = new MemoryDefinition("shared_memory", AggregateType.Array | AggregateType.U32,  BitUtils.DivRoundUp(sharedMemorySize, sizeof(uint)));
+                LocalMemoryId = -1;
+            }
+        }
 
-                SharedMemoryId = properties.AddSharedMemory(smem);
+        public void SetCurrentSharedMemory(int size, bool isUsed)
+        {
+            if (isUsed)
+            {
+                if (size <= 0)
+                {
+                    size = DefaultSharedMemorySize;
+                }
+
+                var smem = new MemoryDefinition("shared_memory", AggregateType.Array | AggregateType.U32,  BitUtils.DivRoundUp(size, sizeof(uint)));
+
+                SharedMemoryId = Properties.AddSharedMemory(smem);
+            }
+            else
+            {
+                SharedMemoryId = -1;
             }
         }
 

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

@@ -126,9 +126,10 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public ShaderConfig(ShaderStage stage, IGpuAccessor gpuAccessor, TranslationOptions options, int localMemorySize)
         {
-            Stage       = stage;
+            Stage           = stage;
-            GpuAccessor = gpuAccessor;
+            GpuAccessor     = gpuAccessor;
-            Options     = options;
+            Options         = options;
+            LocalMemorySize = localMemorySize;
 
             _transformFeedbackDefinitions = new Dictionary<TransformFeedbackVariable, TransformFeedbackOutput>();
 
@@ -143,7 +144,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             _usedTextures = new Dictionary<TextureInfo, TextureMeta>();
             _usedImages   = new Dictionary<TextureInfo, TextureMeta>();
 
-            ResourceManager = new ResourceManager(stage, gpuAccessor, new ShaderProperties(), localMemorySize);
+            ResourceManager = new ResourceManager(stage, gpuAccessor, new ShaderProperties());
 
             if (!gpuAccessor.QueryHostSupportsTransformFeedback() && gpuAccessor.QueryTransformFeedbackEnabled())
             {
@@ -192,7 +193,6 @@ namespace Ryujinx.Graphics.Shader.Translation
             ThreadsPerInputPrimitive = header.ThreadsPerInputPrimitive;
             OutputTopology           = header.OutputTopology;
             MaxOutputVertices        = header.MaxOutputVertexCount;
-            LocalMemorySize          = header.ShaderLocalMemoryLowSize + header.ShaderLocalMemoryHighSize + (header.ShaderLocalMemoryCrsSize / ThreadsPerWarp);
             ImapTypes                = header.ImapTypes;
             OmapTargets              = header.OmapTargets;
             OmapSampleMask           = header.OmapSampleMask;

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

@@ -149,6 +149,17 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public ShaderProgram Translate(TranslatorContext other = null)
         {
+            bool usesLocalMemory = _config.UsedFeatures.HasFlag(FeatureFlags.LocalMemory);
+
+            _config.ResourceManager.SetCurrentLocalMemory(_config.LocalMemorySize, usesLocalMemory);
+
+            if (_config.Stage == ShaderStage.Compute)
+            {
+                bool usesSharedMemory = _config.UsedFeatures.HasFlag(FeatureFlags.SharedMemory);
+
+                _config.ResourceManager.SetCurrentSharedMemory(GpuAccessor.QueryComputeSharedMemorySize(), usesSharedMemory);
+            }
+
             FunctionCode[] code = EmitShader(_program, _config, initializeOutputs: other == null, out _);
 
             if (other != null)
@@ -157,6 +168,7 @@ namespace Ryujinx.Graphics.Shader.Translation
 
                 // We need to share the resource manager since both shaders accesses the same constant buffers.
                 other._config.ResourceManager = _config.ResourceManager;
+                other._config.ResourceManager.SetCurrentLocalMemory(other._config.LocalMemorySize, other._config.UsedFeatures.HasFlag(FeatureFlags.LocalMemory));
 
                 FunctionCode[] otherCode = EmitShader(other._program, other._config, initializeOutputs: true, out int aStart);