From 4e991c3e6f016b133295005654782d12426efb31 Mon Sep 17 00:00:00 2001
From: Kieron Morris <kjm@kieronmorris.me>
Date: Sun, 30 Jan 2022 00:17:50 +0000
Subject: [PATCH] Double Buffer with SSE 128 copy working

---
 src/driver/video/doublebuffer.pas | 15 +++++++--------
 src/include/util.pas              |  6 +++---
 src/kernel.pas                    | 29 +++++++++++++++++++++++++++++
 3 files changed, 39 insertions(+), 11 deletions(-)

diff --git a/src/driver/video/doublebuffer.pas b/src/driver/video/doublebuffer.pas
index dfff1206..36f1ecd7 100644
--- a/src/driver/video/doublebuffer.pas
+++ b/src/driver/video/doublebuffer.pas
@@ -55,25 +55,24 @@ end;
 procedure Flush(FrontBuffer : PVideoBuffer; BackBuffer : PVideoBuffer);
 var
     idx : uint32;
-    Back,Front : PuInt64;
+    Back,Front : uint32;
     BufferSize : uint32;
 
 const
-    COPY_WIDTH = 64;
+    //COPY_WIDTH = 64; //Use this for 64bit copies
+    COPY_WIDTH = 128; //Use this for SSE copies
 
 begin
     //tracer.push_trace('doublebuffer.Flush.enter');
     if not(BackBuffer^.Initialized) then exit;
     if ((FrontBuffer^.Width > BackBuffer^.Width) or (FrontBuffer^.Height > BackBuffer^.Height)) then exit;
-    Back:= PUint64(BackBuffer^.Location);
-    Front:= PuInt64(FrontBuffer^.Location);
+    Back:= BackBuffer^.Location;
+    Front:= FrontBuffer^.Location;
     BufferSize:= ( ( BackBuffer^.Width * BackBuffer^.Height * BackBuffer^.BitsPerPixel) div COPY_WIDTH ) - 1;
     for idx:=0 to BufferSize do begin
-        Front[idx]:= Back[idx];
+        //Front[idx]:= Back[idx];
         // -- TODO: Get SSE working here for 128bit copies --
-        // __SSE_128_memcpy(uint32(Front), uint32(Back));
-        // Front:= PUint64(uint32(Front) + 16);
-        // Back:= PUint64(uint32(Back) + 16);     
+        __SSE_128_memcpy(Back + (idx * 16), Front + (idx * 16));  
     end;
     //tracer.push_trace('doublebuffer.Flush.exit');
 end;
diff --git a/src/include/util.pas b/src/include/util.pas
index ac648a11..77069499 100644
--- a/src/include/util.pas
+++ b/src/include/util.pas
@@ -131,9 +131,9 @@ end;
 procedure __SSE_128_memcpy(source : uint32; dest : uint32); assembler;
 asm
     MOV EAX, Source
-    MOV ECX, Dest
-    MOVUPS XMM0, [EAX]
-    MOVUPS [ECX], XMM0
+    MOVAPS XMM1, [EAX]
+    MOV EAX, Dest
+    MOVAPS [EAX], XMM1
 end;
 
 function switchendian16(b : uint16) : uint16;
diff --git a/src/kernel.pas b/src/kernel.pas
index eb0c7ea8..d36f5e4d 100644
--- a/src/kernel.pas
+++ b/src/kernel.pas
@@ -126,6 +126,9 @@ var
    HM              : PHashMap;
 
    colour          : TRGB32;
+
+   array1          : Array[0..255] of char;
+   array2          : Array[0..255] of char;
    
 begin
      { Init the base system unit }
@@ -213,6 +216,32 @@ begin
      video.enable('BASIC_DOUBLE_BUFFER');
      colour:= color.white;
 
+    //  serial.sendHex(uint32(@array1[8]));
+    //  serial.sendHex(uint32(@array2[8]));
+    //  for i:=8 to 23 do begin
+    //     array2[i]:= '?';
+    //  end;
+    //  array2[24]:= #0;
+
+    //  array1[8]:= 'H';
+    //  array1[9]:= 'e';
+    //  array1[10]:= 'l';
+    //  array1[11]:= 'l';
+    //  array1[12]:= 'o';
+    //  array1[13]:= 'w';
+    //  array1[14]:= 'o';
+    //  array1[15]:= 'r';
+    //  array1[16]:= 'l';
+    //  array1[17]:= 'd';
+    //  array1[18]:= '1';
+    //  array1[19]:= '2';
+    //  array1[20]:= '3';
+    //  array1[21]:= '4';
+    //  array1[22]:= '5';
+    //  array1[23]:= '!';
+    //  __SSE_128_memcpy(uint32(@array1[8]), uint32(@array2[8]));
+    //  serial.sendString(pchar(@array2[8]));
+
      for i:=0 to video.frontBufferWidth-1 do begin
         for z:=0 to video.frontBufferHeight-1 do begin
             video.DrawPixel(i, z, colour);