diff --git a/Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs b/Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs
index 0a3c0da98a..d4a76bc936 100644
--- a/Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs
+++ b/Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs
@@ -585,6 +585,7 @@ namespace Ryujinx.Graphics.Gal.Shader
             bool AbsA = OpCode.Read(46);
             bool NegA = OpCode.Read(48);
             bool AbsB = OpCode.Read(49);
+            bool Sat  = OpCode.Read(50);
 
             ShaderIrNode OperA = OpCode.Gpr8(), OperB;
 
@@ -603,12 +604,13 @@ namespace Ryujinx.Graphics.Gal.Shader
 
             ShaderIrNode Op = new ShaderIrOp(ShaderIrInst.Fadd, OperA, OperB);
 
-            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), Op)));
+            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), GetAluFsat(Op, Sat))));
         }
 
         private static void EmitFmul(ShaderIrBlock Block, long OpCode, ShaderOper Oper)
         {
             bool NegB = OpCode.Read(48);
+            bool Sat  = OpCode.Read(50);
 
             ShaderIrNode OperA = OpCode.Gpr8(), OperB;
 
@@ -625,13 +627,14 @@ namespace Ryujinx.Graphics.Gal.Shader
 
             ShaderIrNode Op = new ShaderIrOp(ShaderIrInst.Fmul, OperA, OperB);
 
-            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), Op)));
+            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), GetAluFsat(Op, Sat))));
         }
 
         private static void EmitFfma(ShaderIrBlock Block, long OpCode, ShaderOper Oper)
         {
             bool NegB = OpCode.Read(48);
             bool NegC = OpCode.Read(49);
+            bool Sat  = OpCode.Read(50);
 
             ShaderIrNode OperA = OpCode.Gpr8(), OperB, OperC;
 
@@ -658,7 +661,7 @@ namespace Ryujinx.Graphics.Gal.Shader
 
             ShaderIrOp Op = new ShaderIrOp(ShaderIrInst.Ffma, OperA, OperB, OperC);
 
-            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), Op)));
+            Block.AddNode(OpCode.PredNode(new ShaderIrAsg(OpCode.Gpr0(), GetAluFsat(Op, Sat))));
         }
 
         private static void EmitIadd(ShaderIrBlock Block, long OpCode, ShaderOper Oper)
diff --git a/Ryujinx.Graphics/Gal/Shader/ShaderDecodeHelper.cs b/Ryujinx.Graphics/Gal/Shader/ShaderDecodeHelper.cs
index ebacd53ab4..d07bcd9171 100644
--- a/Ryujinx.Graphics/Gal/Shader/ShaderDecodeHelper.cs
+++ b/Ryujinx.Graphics/Gal/Shader/ShaderDecodeHelper.cs
@@ -2,6 +2,9 @@ namespace Ryujinx.Graphics.Gal.Shader
 {
     static class ShaderDecodeHelper
     {
+        private static readonly ShaderIrOperImmf ImmfZero = new ShaderIrOperImmf(0);
+        private static readonly ShaderIrOperImmf ImmfOne  = new ShaderIrOperImmf(1);
+
         public static ShaderIrNode GetAluFabsFneg(ShaderIrNode Node, bool Abs, bool Neg)
         {
             return GetAluFneg(GetAluFabs(Node, Abs), Neg);
@@ -17,6 +20,11 @@ namespace Ryujinx.Graphics.Gal.Shader
             return Neg ? new ShaderIrOp(ShaderIrInst.Fneg, Node) : Node;
         }
 
+        public static ShaderIrNode GetAluFsat(ShaderIrNode Node, bool Sat)
+        {
+            return Sat ? new ShaderIrOp(ShaderIrInst.Fclamp, Node, ImmfZero, ImmfOne) : Node;
+        }
+
         public static ShaderIrNode GetAluIabsIneg(ShaderIrNode Node, bool Abs, bool Neg)
         {
             return GetAluIneg(GetAluIabs(Node, Abs), Neg);
diff --git a/Ryujinx.Graphics/Texture/ImageUtils.cs b/Ryujinx.Graphics/Texture/ImageUtils.cs
index 1b043245ef..e1f370cda9 100644
--- a/Ryujinx.Graphics/Texture/ImageUtils.cs
+++ b/Ryujinx.Graphics/Texture/ImageUtils.cs
@@ -235,18 +235,23 @@ namespace Ryujinx.Graphics.Texture
 
             int BytesPerPixel = Desc.BytesPerPixel;
 
-            int OutOffs = 0;
+            //Note: Each row of the texture needs to be aligned to 4 bytes.
+            int Pitch = (Width * BytesPerPixel + 3) & ~3;
 
-            byte[] Data = new byte[Width * Height * BytesPerPixel];
+            byte[] Data = new byte[Height * Pitch];
 
             for (int Y = 0; Y < Height; Y++)
-            for (int X = 0; X < Width;  X++)
             {
-                long Offset = (uint)Swizzle.GetSwizzleOffset(X, Y);
+                int OutOffs = Y * Pitch;
 
-                CpuMemory.ReadBytes(Position + Offset, Data, OutOffs, BytesPerPixel);
+                for (int X = 0; X < Width;  X++)
+                {
+                    long Offset = (uint)Swizzle.GetSwizzleOffset(X, Y);
 
-                OutOffs += BytesPerPixel;
+                    CpuMemory.ReadBytes(Position + Offset, Data, OutOffs, BytesPerPixel);
+
+                    OutOffs += BytesPerPixel;
+                }
             }
 
             return Data;