Implement IMUL, PCNT and CONT shader instructions, fix FFMA32I and HFMA32I (#2972)

* Implement IMUL shader instruction

* Implement PCNT/CONT instruction and fix FFMA32I

* Add HFMA232I to the table

* Shader cache version bump

* No Rc on Ffma32i

4 files changed, 69 insertions, 54 deletions

diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs
index b0cb7028..c242963a 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs
@@ -54,18 +54,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Config.GpuAccessor.Log("Shader instruction Cctlt is not implemented.");
         }
 
-        public static void Cont(EmitterContext context)
-        {
-            InstCont op = context.GetOp<InstCont>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ContUnsup is not implemented.");
-        }
-
         public static void Cset(EmitterContext context)
         {
             InstCset op = context.GetOp<InstCset>();
 
-            context.Config.GpuAccessor.Log("Shader instruction CsetUnsup is not implemented.");
+            context.Config.GpuAccessor.Log("Shader instruction Cset is not implemented.");
         }
 
         public static void Cs2r(EmitterContext context)
@@ -159,34 +152,6 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Config.GpuAccessor.Log("Shader instruction ImadspRc is not implemented.");
         }
 
-        public static void ImulR(EmitterContext context)
-        {
-            InstImulR op = context.GetOp<InstImulR>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ImulR is not implemented.");
-        }
-
-        public static void ImulI(EmitterContext context)
-        {
-            InstImulI op = context.GetOp<InstImulI>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ImulI is not implemented.");
-        }
-
-        public static void ImulC(EmitterContext context)
-        {
-            InstImulC op = context.GetOp<InstImulC>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ImulC is not implemented.");
-        }
-
-        public static void Imul32i(EmitterContext context)
-        {
-            InstImul32i op = context.GetOp<InstImul32i>();
-
-            context.Config.GpuAccessor.Log("Shader instruction Imul32i is not implemented.");
-        }
-
         public static void Jcal(EmitterContext context)
         {
             InstJcal op = context.GetOp<InstJcal>();
@@ -250,13 +215,6 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Config.GpuAccessor.Log("Shader instruction P2rC is not implemented.");
         }
 
-        public static void Pcnt(EmitterContext context)
-        {
-            InstPcnt op = context.GetOp<InstPcnt>();
-
-            context.Config.GpuAccessor.Log("Shader instruction Pcnt is not implemented.");
-        }
-
         public static void Pexit(EmitterContext context)
         {
             InstPexit op = context.GetOp<InstPexit>();
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatArithmetic.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatArithmetic.cs
index 11d724c4..29803c31 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatArithmetic.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitFloatArithmetic.cs
@@ -204,7 +204,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             var srcA = GetSrcReg(context, op.SrcA);
             var srcB = GetSrcImm(context, op.Imm32);
-            var srcC = GetSrcReg(context, op.SrcC);
+            var srcC = GetSrcReg(context, op.Dest);
 
             EmitFfma(context, Instruction.FP32, srcA, srcB, srcC, op.Dest, op.NegA, op.NegC, op.Sat, op.WriteCC);
         }
@@ -333,13 +333,13 @@ namespace Ryujinx.Graphics.Shader.Instructions
             EmitHfma2(context, op.OFmt, srcA, srcB, srcC, op.Dest, op.Sat);
         }
 
-        public static void Hfma232iI(EmitterContext context)
+        public static void Hfma232i(EmitterContext context)
         {
             InstHfma232i op = context.GetOp<InstHfma232i>();
 
             var srcA = GetHalfSrc(context, op.ASwizzle, op.SrcA, false, false);
             var srcB = GetHalfSrc(context, op.Imm);
-            var srcC = GetHalfSrc(context, HalfSwizzle.F16, op.SrcC, op.NegC, false);
+            var srcC = GetHalfSrc(context, HalfSwizzle.F16, op.Dest, op.NegC, false);
 
             EmitHfma2(context, OFmt.F16, srcA, srcB, srcC, op.Dest, saturate: false);
         }
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs
index da34c1be..3cb8fe72 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs
@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstBrk op = context.GetOp<InstBrk>();
 
-            EmitBrkOrSync(context);
+            EmitBrkContSync(context);
         }
 
         public static void Brx(EmitterContext context)
@@ -87,6 +87,13 @@ namespace Ryujinx.Graphics.Shader.Instructions
             }
         }
 
+        public static void Cont(EmitterContext context)
+        {
+            InstCont op = context.GetOp<InstCont>();
+
+            EmitBrkContSync(context);
+        }
+
         public static void Exit(EmitterContext context)
         {
             InstExit op = context.GetOp<InstExit>();
@@ -116,7 +123,14 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstPbk op = context.GetOp<InstPbk>();
 
-            EmitPbkOrSsy(context);
+            EmitPbkPcntSsy(context);
+        }
+
+        public static void Pcnt(EmitterContext context)
+        {
+            InstPcnt op = context.GetOp<InstPcnt>();
+
+            EmitPbkPcntSsy(context);
         }
 
         public static void Ret(EmitterContext context)
@@ -137,17 +151,17 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstSsy op = context.GetOp<InstSsy>();
 
-            EmitPbkOrSsy(context);
+            EmitPbkPcntSsy(context);
         }
 
         public static void Sync(EmitterContext context)
         {
             InstSync op = context.GetOp<InstSync>();
 
-            EmitBrkOrSync(context);
+            EmitBrkContSync(context);
         }
 
-        private static void EmitPbkOrSsy(EmitterContext context)
+        private static void EmitPbkPcntSsy(EmitterContext context)
         {
             var consumers = context.CurrBlock.PushOpCodes.First(x => x.Op.Address == context.CurrOp.Address).Consumers;
 
@@ -162,7 +176,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
             }
         }
 
-        private static void EmitBrkOrSync(EmitterContext context)
+        private static void EmitBrkContSync(EmitterContext context)
         {
             var targets = context.CurrBlock.SyncTargets;
 
diff --git a/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerArithmetic.cs b/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerArithmetic.cs
index ac8cca1b..374e3d61 100644
--- a/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerArithmetic.cs
+++ b/Ryujinx.Graphics.Shader/Instructions/InstEmitIntegerArithmetic.cs
@@ -138,6 +138,46 @@ namespace Ryujinx.Graphics.Shader.Instructions
             EmitImad(context, srcA, srcB, srcC, op.Dest, op.AvgMode, op.ASigned, op.BSigned, op.Hilo);
         }
 
+        public static void ImulR(EmitterContext context)
+        {
+            InstImulR op = context.GetOp<InstImulR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImulI(EmitterContext context)
+        {
+            InstImulI op = context.GetOp<InstImulI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, Imm20ToSInt(op.Imm20));
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void ImulC(EmitterContext context)
+        {
+            InstImulC op = context.GetOp<InstImulC>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
+        public static void Imul32i(EmitterContext context)
+        {
+            InstImul32i op = context.GetOp<InstImul32i>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcImm(context, op.Imm32);
+
+            EmitImad(context, srcA, srcB, Const(0), op.Dest, AvgMode.NoNeg, op.ASigned, op.BSigned, op.Hilo);
+        }
+
         public static void IscaddR(EmitterContext context)
         {
             InstIscaddR op = context.GetOp<InstIscaddR>();
@@ -366,7 +406,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
             // TODO: CC, X, corner cases.
         }
 
-        public static void EmitImad(
+        private static void EmitImad(
             EmitterContext context,
             Operand srcA,
             Operand srcB,
@@ -407,7 +447,10 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 res = context.IMultiply(srcA, srcB);
             }
 
-            res = context.IAdd(res, srcC);
+            if (srcC.Type != OperandType.Constant || srcC.Value != 0)
+            {
+                res = context.IAdd(res, srcC);
+            }
 
             // TODO: CC, X, SAT, and more?