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

4 files changed, 63 insertions, 12 deletions

diff --git a/Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs b/Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs
index 113ece99..a8fef95b 100644
--- a/Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs
+++ b/Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs
@@ -627,6 +627,16 @@ namespace Ryujinx.Graphics.Shader.Translation
             return context.Add(Instruction.StoreGlobal, null, a, b, c);
         }
 
+        public static Operand StoreGlobal16(this EmitterContext context, Operand a, Operand b, Operand c)
+        {
+            return context.Add(Instruction.StoreGlobal16, null, a, b, c);
+        }
+
+        public static Operand StoreGlobal8(this EmitterContext context, Operand a, Operand b, Operand c)
+        {
+            return context.Add(Instruction.StoreGlobal8, null, a, b, c);
+        }
+
         public static Operand StoreLocal(this EmitterContext context, Operand a, Operand b)
         {
             return context.Add(Instruction.StoreLocal, null, a, b);
@@ -637,6 +647,16 @@ namespace Ryujinx.Graphics.Shader.Translation
             return context.Add(Instruction.StoreShared, null, a, b);
         }
 
+        public static Operand StoreShared16(this EmitterContext context, Operand a, Operand b)
+        {
+            return context.Add(Instruction.StoreShared16, null, a, b);
+        }
+
+        public static Operand StoreShared8(this EmitterContext context, Operand a, Operand b)
+        {
+            return context.Add(Instruction.StoreShared8, null, a, b);
+        }
+
         public static Operand UnpackDouble2x32High(this EmitterContext context, Operand a)
         {
             return UnpackDouble2x32(context, a, 1);
diff --git a/Ryujinx.Graphics.Shader/Translation/GlobalMemory.cs b/Ryujinx.Graphics.Shader/Translation/GlobalMemory.cs
index 75bd9ddf..1be63868 100644
--- a/Ryujinx.Graphics.Shader/Translation/GlobalMemory.cs
+++ b/Ryujinx.Graphics.Shader/Translation/GlobalMemory.cs
@@ -20,7 +20,9 @@ namespace Ryujinx.Graphics.Shader.Translation
         {
             return (inst.IsAtomic() && IsGlobalMr(inst)) ||
                     inst == Instruction.LoadGlobal ||
-                    inst == Instruction.StoreGlobal;
+                    inst == Instruction.StoreGlobal ||
+                    inst == Instruction.StoreGlobal16 ||
+                    inst == Instruction.StoreGlobal8;
         }
 
         private static bool IsGlobalMr(Instruction inst)
diff --git a/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs b/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
index cc57102c..1cf43e5d 100644
--- a/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
+++ b/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
@@ -59,7 +59,8 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             Operation operation = (Operation)node.Value;
 
             bool isAtomic = operation.Inst.IsAtomic();
-            bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal;
+            bool isStg16Or8 = operation.Inst == Instruction.StoreGlobal16 || operation.Inst == Instruction.StoreGlobal8;
+            bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal || isStg16Or8;
 
             config.SetUsedStorageBuffer(storageIndex, isWrite);
 
@@ -78,12 +79,18 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                 node.List.AddBefore(node, andOp);
 
                 Operand byteOffset = Local();
-                Operand wordOffset = Local();
+                Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
 
-                Operation subOp = new Operation(Instruction.Subtract,      byteOffset, addrLow, baseAddrTrunc);
+                node.List.AddBefore(node, subOp);
+
+                if (isStg16Or8)
+                {
+                    return byteOffset;
+                }
+
+                Operand wordOffset = Local();
                 Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
 
-                node.List.AddBefore(node, subOp);
                 node.List.AddBefore(node, shrOp);
 
                 return wordOffset;
@@ -113,7 +120,14 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             }
             else
             {
-                storageOp = new Operation(Instruction.StoreStorage, null, sources);
+                Instruction storeInst = operation.Inst switch
+                {
+                    Instruction.StoreGlobal16 => Instruction.StoreStorage16,
+                    Instruction.StoreGlobal8 => Instruction.StoreStorage8,
+                    _ => Instruction.StoreStorage
+                };
+
+                storageOp = new Operation(storeInst, null, sources);
             }
 
             for (int index = 0; index < operation.SourcesCount; index++)
diff --git a/Ryujinx.Graphics.Shader/Translation/Rewriter.cs b/Ryujinx.Graphics.Shader/Translation/Rewriter.cs
index 47428520..02a0feda 100644
--- a/Ryujinx.Graphics.Shader/Translation/Rewriter.cs
+++ b/Ryujinx.Graphics.Shader/Translation/Rewriter.cs
@@ -49,7 +49,8 @@ namespace Ryujinx.Graphics.Shader.Translation
             Operation operation = (Operation)node.Value;
 
             bool isAtomic = operation.Inst.IsAtomic();
-            bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal;
+            bool isStg16Or8 = operation.Inst == Instruction.StoreGlobal16 || operation.Inst == Instruction.StoreGlobal8;
+            bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal || isStg16Or8;
 
             Operation storageOp;
 
@@ -95,14 +96,21 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
 
-            Operand baseAddrTrunc = PrependOperation(Instruction.BitwiseAnd,    sbBaseAddrLow, alignMask);
-            Operand byteOffset    = PrependOperation(Instruction.Subtract,      addrLow, baseAddrTrunc);
-            Operand wordOffset    = PrependOperation(Instruction.ShiftRightU32, byteOffset, Const(2));
+            Operand baseAddrTrunc = PrependOperation(Instruction.BitwiseAnd, sbBaseAddrLow, alignMask);
+            Operand byteOffset    = PrependOperation(Instruction.Subtract, addrLow, baseAddrTrunc);
 
             Operand[] sources = new Operand[operation.SourcesCount];
 
             sources[0] = sbSlot;
-            sources[1] = wordOffset;
+
+            if (isStg16Or8)
+            {
+                sources[1] = byteOffset;
+            }
+            else
+            {
+                sources[1] = PrependOperation(Instruction.ShiftRightU32, byteOffset, Const(2));
+            }
 
             for (int index = 2; index < operation.SourcesCount; index++)
             {
@@ -121,7 +129,14 @@ namespace Ryujinx.Graphics.Shader.Translation
             }
             else
             {
-                storageOp = new Operation(Instruction.StoreStorage, null, sources);
+                Instruction storeInst = operation.Inst switch
+                {
+                    Instruction.StoreGlobal16 => Instruction.StoreStorage16,
+                    Instruction.StoreGlobal8 => Instruction.StoreStorage8,
+                    _ => Instruction.StoreStorage
+                };
+
+                storageOp = new Operation(storeInst, null, sources);
             }
 
             for (int index = 0; index < operation.SourcesCount; index++)