Fix shader 8-bit and 16-bit STS/STG (#2741)

* Fix 8 and 16-bit STG

* Fix 8 and 16-bit STS

* Shader cache version bump

9 files changed, 145 insertions, 4 deletions

diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
index 3e2e51b0..85288afa 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
@@ -270,6 +270,16 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/ShuffleXor.glsl");
             }
 
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.StoreSharedSmallInt) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl");
+            }
+
+            if ((info.HelperFunctionsMask & HelperFunctionsMask.StoreStorageSmallInt) != 0)
+            {
+                AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl");
+            }
+
             if ((info.HelperFunctionsMask & HelperFunctionsMask.SwizzleAdd) != 0)
             {
                 AppendHelperFunction(context, "Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/SwizzleAdd.glsl");
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl
index 9f8c641d..82b76bcc 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Shared.glsl
@@ -5,7 +5,7 @@ int Helper_AtomicMaxS32(int offset, int value)
     {
         oldValue = $SHARED_MEM$[offset];
         newValue = uint(max(int(oldValue), value));
-    } while (atomicCompSwap($SHARED_MEM$[offset], newValue, oldValue) != oldValue);
+    } while (atomicCompSwap($SHARED_MEM$[offset], oldValue, newValue) != oldValue);
     return int(oldValue);
 }
 
@@ -16,6 +16,6 @@ int Helper_AtomicMinS32(int offset, int value)
     {
         oldValue = $SHARED_MEM$[offset];
         newValue = uint(min(int(oldValue), value));
-    } while (atomicCompSwap($SHARED_MEM$[offset], newValue, oldValue) != oldValue);
+    } while (atomicCompSwap($SHARED_MEM$[offset], oldValue, newValue) != oldValue);
     return int(oldValue);
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl
index fc3af6a7..0862a71b 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/AtomicMinMaxS32Storage.glsl
@@ -5,7 +5,7 @@ int Helper_AtomicMaxS32(int index, int offset, int value)
     {
         oldValue = $STORAGE_MEM$[index].data[offset];
         newValue = uint(max(int(oldValue), value));
-    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], newValue, oldValue) != oldValue);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], oldValue, newValue) != oldValue);
     return int(oldValue);
 }
 
@@ -16,6 +16,6 @@ int Helper_AtomicMinS32(int index, int offset, int value)
     {
         oldValue = $STORAGE_MEM$[index].data[offset];
         newValue = uint(min(int(oldValue), value));
-    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], newValue, oldValue) != oldValue);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[offset], oldValue, newValue) != oldValue);
     return int(oldValue);
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/HelperFunctionNames.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/HelperFunctionNames.cs
index 1ff127bb..54f35b15 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/HelperFunctionNames.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/HelperFunctionNames.cs
@@ -13,5 +13,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
         public static string ShuffleUp   = "Helper_ShuffleUp";
         public static string ShuffleXor  = "Helper_ShuffleXor";
         public static string SwizzleAdd  = "Helper_SwizzleAdd";
+
+        public static string StoreShared16  = "Helper_StoreShared16";
+        public static string StoreShared8   = "Helper_StoreShared8";
+        public static string StoreStorage16 = "Helper_StoreStorage16";
+        public static string StoreStorage8  = "Helper_StoreStorage8";
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl
new file mode 100644
index 00000000..2f57b5ff
--- /dev/null
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreSharedSmallInt.glsl
@@ -0,0 +1,23 @@
+void Helper_StoreShared16(int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 16);
+    } while (atomicCompSwap($SHARED_MEM$[wordOffset], oldValue, newValue) != oldValue);
+}
+
+void Helper_StoreShared8(int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $SHARED_MEM$[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 8);
+    } while (atomicCompSwap($SHARED_MEM$[wordOffset], oldValue, newValue) != oldValue);
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl
new file mode 100644
index 00000000..f2253a79
--- /dev/null
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/StoreStorageSmallInt.glsl
@@ -0,0 +1,23 @@
+void Helper_StoreStorage16(int index, int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 16);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[wordOffset], oldValue, newValue) != oldValue);
+}
+
+void Helper_StoreStorage8(int index, int offset, uint value)
+{
+    int wordOffset = offset >> 2;
+    int bitOffset = (offset & 3) * 8;
+    uint oldValue, newValue;
+    do
+    {
+        oldValue = $STORAGE_MEM$[index].data[wordOffset];
+        newValue = bitfieldInsert(oldValue, value, bitOffset, 8);
+    } while (atomicCompSwap($STORAGE_MEM$[index].data[wordOffset], oldValue, newValue) != oldValue);
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
index 4c2582f0..4f059149 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
@@ -185,9 +185,21 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                     case Instruction.StoreShared:
                         return StoreShared(context, operation);
 
+                    case Instruction.StoreShared16:
+                        return StoreShared16(context, operation);
+
+                    case Instruction.StoreShared8:
+                        return StoreShared8(context, operation);
+
                     case Instruction.StoreStorage:
                         return StoreStorage(context, operation);
 
+                    case Instruction.StoreStorage16:
+                        return StoreStorage16(context, operation);
+
+                    case Instruction.StoreStorage8:
+                        return StoreStorage8(context, operation);
+
                     case Instruction.TextureSample:
                         return TextureSample(context, operation);
 
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
index 022cd746..c2d59ff7 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
@@ -113,7 +113,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             Add(Instruction.StoreAttribute,           InstType.Special);
             Add(Instruction.StoreLocal,               InstType.Special);
             Add(Instruction.StoreShared,              InstType.Special);
+            Add(Instruction.StoreShared16,            InstType.Special);
+            Add(Instruction.StoreShared8,             InstType.Special);
             Add(Instruction.StoreStorage,             InstType.Special);
+            Add(Instruction.StoreStorage16,           InstType.Special);
+            Add(Instruction.StoreStorage8,            InstType.Special);
             Add(Instruction.Subtract,                 InstType.OpBinary,       "-",               2);
             Add(Instruction.SwizzleAdd,               InstType.CallTernary,    HelperFunctionNames.SwizzleAdd);
             Add(Instruction.TextureSample,            InstType.Special);
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
index 5e46bb46..f976ec5e 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
@@ -365,6 +365,34 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             return $"{arrayName}[{offsetExpr}] = {src}";
         }
 
+        public static string StoreShared16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            VariableType srcType = OperandManager.GetNodeDestType(context, src2);
+
+            string src = TypeConversion.ReinterpretCast(context, src2, srcType, VariableType.U32);
+
+            return $"{HelperFunctionNames.StoreShared16}({offsetExpr}, {src})";
+        }
+
+        public static string StoreShared8(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+
+            string offsetExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+
+            VariableType srcType = OperandManager.GetNodeDestType(context, src2);
+
+            string src = TypeConversion.ReinterpretCast(context, src2, srcType, VariableType.U32);
+
+            return $"{HelperFunctionNames.StoreShared8}({offsetExpr}, {src})";
+        }
+
         public static string StoreStorage(CodeGenContext context, AstOperation operation)
         {
             IAstNode src1 = operation.GetSource(0);
@@ -383,6 +411,42 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             return $"{sb} = {src}";
         }
 
+        public static string StoreStorage16(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+            IAstNode src3 = operation.GetSource(2);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            VariableType srcType = OperandManager.GetNodeDestType(context, src3);
+
+            string src = TypeConversion.ReinterpretCast(context, src3, srcType, VariableType.U32);
+
+            string sb = GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+
+            return $"{HelperFunctionNames.StoreStorage16}({indexExpr}, {offsetExpr}, {src})";
+        }
+
+        public static string StoreStorage8(CodeGenContext context, AstOperation operation)
+        {
+            IAstNode src1 = operation.GetSource(0);
+            IAstNode src2 = operation.GetSource(1);
+            IAstNode src3 = operation.GetSource(2);
+
+            string indexExpr  = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
+            string offsetExpr = GetSoureExpr(context, src2, GetSrcVarType(operation.Inst, 1));
+
+            VariableType srcType = OperandManager.GetNodeDestType(context, src3);
+
+            string src = TypeConversion.ReinterpretCast(context, src3, srcType, VariableType.U32);
+
+            string sb = GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
+
+            return $"{HelperFunctionNames.StoreStorage8}({indexExpr}, {offsetExpr}, {src})";
+        }
+
         public static string TextureSample(CodeGenContext context, AstOperation operation)
         {
             AstTextureOperation texOp = (AstTextureOperation)operation;