Spexeah/Asuro
Spexeah/Asuro
3
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Double Buffer with SSE 128 copy working

Browse Source
This commit is contained in:
Kieron Morris 2022-01-30 00:17:50 +00:00
parent a3217de71a
commit 4e991c3e6f
3 changed files with 39 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/driver/video/doublebuffer.pas
View File

@ -55,25 +55,24 @@ end;
procedure Flush(FrontBuffer : PVideoBuffer; BackBuffer : PVideoBuffer); procedure Flush(FrontBuffer : PVideoBuffer; BackBuffer : PVideoBuffer);
var var
idx : uint32; idx : uint32;
Back,Front : PuInt64; Back,Front : uint32;
BufferSize : uint32; BufferSize : uint32;
const const
COPY_WIDTH = 64; //COPY_WIDTH = 64; //Use this for 64bit copies
COPY_WIDTH = 128; //Use this for SSE copies
begin begin
//tracer.push_trace('doublebuffer.Flush.enter'); //tracer.push_trace('doublebuffer.Flush.enter');
if not(BackBuffer^.Initialized) then exit; if not(BackBuffer^.Initialized) then exit;
if ((FrontBuffer^.Width > BackBuffer^.Width) or (FrontBuffer^.Height > BackBuffer^.Height)) then exit; if ((FrontBuffer^.Width > BackBuffer^.Width) or (FrontBuffer^.Height > BackBuffer^.Height)) then exit;
Back:= PUint64(BackBuffer^.Location); Back:= BackBuffer^.Location;
Front:= PuInt64(FrontBuffer^.Location); Front:= FrontBuffer^.Location;
BufferSize:= ( ( BackBuffer^.Width * BackBuffer^.Height * BackBuffer^.BitsPerPixel) div COPY_WIDTH ) - 1; BufferSize:= ( ( BackBuffer^.Width * BackBuffer^.Height * BackBuffer^.BitsPerPixel) div COPY_WIDTH ) - 1;
for idx:=0 to BufferSize do begin for idx:=0 to BufferSize do begin
Front[idx]:= Back[idx]; //Front[idx]:= Back[idx];
// -- TODO: Get SSE working here for 128bit copies -- // -- TODO: Get SSE working here for 128bit copies --
// __SSE_128_memcpy(uint32(Front), uint32(Back)); __SSE_128_memcpy(Back + (idx * 16), Front + (idx * 16));
// Front:= PUint64(uint32(Front) + 16);
// Back:= PUint64(uint32(Back) + 16);
end; end;
//tracer.push_trace('doublebuffer.Flush.exit'); //tracer.push_trace('doublebuffer.Flush.exit');
end; end;

6
src/include/util.pas
View File

@ -131,9 +131,9 @@ end;
procedure __SSE_128_memcpy(source : uint32; dest : uint32); assembler; procedure __SSE_128_memcpy(source : uint32; dest : uint32); assembler;
asm asm
MOV EAX, Source MOV EAX, Source
MOV ECX, Dest MOVAPS XMM1, [EAX]
MOVUPS XMM0, [EAX] MOV EAX, Dest
MOVUPS [ECX], XMM0 MOVAPS [EAX], XMM1
end; end;
function switchendian16(b : uint16) : uint16; function switchendian16(b : uint16) : uint16;

29
src/kernel.pas
View File

@ -127,6 +127,9 @@ var
colour : TRGB32; colour : TRGB32;
array1 : Array[0..255] of char;
array2 : Array[0..255] of char;
begin begin
{ Init the base system unit } { Init the base system unit }
System.init(); System.init();
@ -213,6 +216,32 @@ begin
video.enable('BASIC_DOUBLE_BUFFER'); video.enable('BASIC_DOUBLE_BUFFER');
colour:= color.white; 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 i:=0 to video.frontBufferWidth-1 do begin
for z:=0 to video.frontBufferHeight-1 do begin for z:=0 to video.frontBufferHeight-1 do begin
video.DrawPixel(i, z, colour); video.DrawPixel(i, z, colour);