在 汇编学习（八）--- 图像水平镜像垂直镜像的极致优化 中，我们使用了 FastMove，来提高内存复制速度。
但 FastMove 不支持 x64。而且很多年没有更新了。我们自己动手来写吧。由简到繁，容易理解。
必须支持：X86、X64。
必须支持：SSE2、SSE4、AVX1、AVX2（包括：内存对齐和未对齐）。
首先写一个最简单的 MOVE 函数。
每一次复制 1 个字节：
{ 每一次复制 1 个字节 <Byte> }
procedure Move_Byte(const src: Pointer; dst: Pointer; const len: NativeInt);
var
pSrc: PByte;
pDst: PByte;
I : NativeInt;
begin
pSrc := PByte(src); // 取 src 首地址
pDst := PByte(dst); // 取 dst 首地址
for I := 0 to len - 1 do //
begin //
pDst^ := pSrc^; // 把 pSrc 源地址中的值 ---> 赋给 pDst 目地址的值；按字节(BYTE)赋值
Inc(pSrc); // 源地址加 1
Inc(pDst); // 目地址加 1
end;
end;
每一次复制 2 个字节：
{ 每一次复制 2 个字节 <WORD> <长度必须是 2 的倍数> }
procedure Move_WORD(const src: Pointer; dst: Pointer; const len: NativeInt);
var
pSrc: PWORD;
pDst: PWORD;
I : NativeInt;
begin
pSrc := PWORD(src); // 取 src 首地址
pDst := PWORD(dst); // 取 dst 首地址
for I := 0 to len div 2 - 1 do //
begin //
pDst^ := pSrc^; // 把 pSrc 源地址中的值 ---> 赋给 pDst 目地址的值；按双字节(WORD)赋值
Inc(pSrc); // 源地址加 2
Inc(pDst); // 目地址加 2
end;
end;
每一次复制 4 个字节：
{ 每一次复制 4 个字节 <DWORD> <长度必须是 4 的倍数> }
procedure Move_DWORD(const src: Pointer; dst: Pointer; const len: NativeInt);
var
pSrc: PDWORD;
pDst: PDWORD;
I : NativeInt;
begin
pSrc := PDWORD(src); // 取 src 首地址
pDst := PDWORD(dst); // 取 dst 首地址
for I := 0 to len div 4 - 1 do //
begin //
pDst^ := pSrc^; // 把 pSrc 源地址中的值 ---> 赋给 pDst 目地址的值；按4字节(DWORD)赋值
Inc(pSrc); // 源地址加 4
Inc(pDst); // 目地址加 4
end;
end;
每一次复制 8 个字节：
{ 每一次复制 8 个字节 <UINT64> <长度必须是 8 的倍数> }
procedure Move_UINT64(const src: Pointer; dst: Pointer; const len: NativeInt);
var
pSrc: PUInt64;
pDst: PUInt64;
I : NativeInt;
begin
pSrc := PUInt64(src); // 取 src 首地址
pDst := PUInt64(dst); // 取 dst 首地址
for I := 0 to len div 8 - 1 do //
begin //
pDst^ := pSrc^; // 把 pSrc 源地址中的值 ---> 赋给 pDst 目地址的值；按8字节(UINT64)赋值
Inc(pSrc); // 源地址加 8
Inc(pDst); // 目地址加 8
end;
end;
每一次复制 16 个字节：
{ 每一次复制 16 个字节 <128 位> <长度必须是 16 的倍数> }
procedure Move_SSE2_16U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
LDDQU XMM0, [EAX]
MOVDQU [EDX], XMM0
ADD EAX, 16
ADD EDX, 16
SUB ECX, 16
JNZ @LOOP
end;每一次复制 32 个字节：
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_SSE2_32U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
LDDQU XMM0, [EAX+00]
LDDQU XMM1, [EAX+16]
MOVDQU [EDX+00], XMM0
MOVDQU [EDX+16], XMM1
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
end;每一次复制 64 个字节：
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_SSE2_64U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
LDDQU XMM0, [EAX+00]
LDDQU XMM1, [EAX+16]
LDDQU XMM2, [EAX+32]
LDDQU XMM3, [EAX+48]
MOVDQU [EDX+00], XMM0
MOVDQU [EDX+16], XMM1
MOVDQU [EDX+32], XMM2
MOVDQU [EDX+48], XMM3
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
end;每一次复制 128 个字节：
{ 每一次复制 128 个字节 <1024 位> <长度必须是 128 的倍数> }
procedure Move_SSE2_128U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
LDDQU XMM0, [EAX+000]
LDDQU XMM1, [EAX+016]
LDDQU XMM2, [EAX+032]
LDDQU XMM3, [EAX+048]
LDDQU XMM4, [EAX+064]
LDDQU XMM5, [EAX+080]
LDDQU XMM6, [EAX+096]
LDDQU XMM7, [EAX+112]
MOVDQU [EDX+000], XMM0
MOVDQU [EDX+016], XMM1
MOVDQU [EDX+032], XMM2
MOVDQU [EDX+048], XMM3
MOVDQU [EDX+064], XMM4
MOVDQU [EDX+080], XMM5
MOVDQU [EDX+096], XMM6
MOVDQU [EDX+112], XMM7
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
end;
每一次复制 256 个字节：
{ 每一次复制 256 个字节 <2048 位> <长度必须是 256 的倍数> }
procedure Move_SSE2_256U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
@LOOP:
LDDQU XMM0, [EAX+000]
LDDQU XMM1, [EAX+016]
LDDQU XMM2, [EAX+032]
LDDQU XMM3, [EAX+048]
LDDQU XMM4, [EAX+064]
LDDQU XMM5, [EAX+080]
LDDQU XMM6, [EAX+096]
LDDQU XMM7, [EAX+112]
LDDQU XMM8, [EAX+128]
LDDQU XMM9, [EAX+144]
LDDQU XMM10, [EAX+160]
LDDQU XMM11, [EAX+176]
LDDQU XMM12, [EAX+192]
LDDQU XMM13, [EAX+208]
LDDQU XMM14, [EAX+224]
LDDQU XMM15, [EAX+240]
MOVDQU [EDX+000], XMM0
MOVDQU [EDX+016], XMM1
MOVDQU [EDX+032], XMM2
MOVDQU [EDX+048], XMM3
MOVDQU [EDX+064], XMM4
MOVDQU [EDX+080], XMM5
MOVDQU [EDX+096], XMM6
MOVDQU [EDX+112], XMM7
MOVDQU [EDX+128], XMM8
MOVDQU [EDX+144], XMM9
MOVDQU [EDX+160], XMM10
MOVDQU [EDX+176], XMM11
MOVDQU [EDX+192], XMM12
MOVDQU [EDX+208], XMM13
MOVDQU [EDX+224], XMM14
MOVDQU [EDX+240], XMM15
{$ELSE}
@LOOP:
LDDQU XMM0, [EAX+000]
LDDQU XMM1, [EAX+016]
LDDQU XMM2, [EAX+032]
LDDQU XMM3, [EAX+048]
LDDQU XMM4, [EAX+064]
LDDQU XMM5, [EAX+080]
LDDQU XMM6, [EAX+096]
LDDQU XMM7, [EAX+112]
MOVDQU [EDX+000], XMM0
MOVDQU [EDX+016], XMM1
MOVDQU [EDX+032], XMM2
MOVDQU [EDX+048], XMM3
MOVDQU [EDX+064], XMM4
MOVDQU [EDX+080], XMM5
MOVDQU [EDX+096], XMM6
MOVDQU [EDX+112], XMM7
LDDQU XMM0, [EAX+128]
LDDQU XMM1, [EAX+144]
LDDQU XMM2, [EAX+160]
LDDQU XMM3, [EAX+176]
LDDQU XMM4, [EAX+192]
LDDQU XMM5, [EAX+208]
LDDQU XMM6, [EAX+224]
LDDQU XMM7, [EAX+240]
MOVDQU [EDX+128], XMM0
MOVDQU [EDX+144], XMM1
MOVDQU [EDX+160], XMM2
MOVDQU [EDX+176], XMM3
MOVDQU [EDX+192], XMM4
MOVDQU [EDX+208], XMM5
MOVDQU [EDX+224], XMM6
MOVDQU [EDX+240], XMM7
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
end;
每一次复制 32 个字节：
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_AVX1_32U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VLDDQU YMM0, [EAX]
VMOVDQU [EDX], YMM0
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FF, $F0, $00
DB $67, $C5, $FE, $7F, $02
{$ELSE}
DB $C5, $FF, $F0, $00
DB $C5, $FE, $7F, $02
{$ENDIF}
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
end;每一次复制 64 个字节：
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_AVX1_64U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VLDDQU YMM0, [EAX+00]
VLDDQU YMM1, [EAX+32]
VMOVDQU [EDX+00], YMM0
VMOVDQU [EDX+32], YMM1
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FF, $F0, $00
DB $67, $C5, $FF, $F0, $48, $20
DB $67, $C5, $FE, $7F, $02
DB $67, $C5, $FE, $7F, $4A, $20
{$ELSE}
DB $C5, $FF, $F0, $00
DB $C5, $FF, $F0, $48, $20
DB $C5, $FE, $7F, $02
DB $C5, $FE, $7F, $4A, $20
{$ENDIF}
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
end;每一次复制 128 个字节：
{ 每一次复制 128 个字节 <1024 位> <长度必须是 128 的倍数> }
procedure Move_AVX1_128U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VLDDQU YMM0, [EAX+00]
VLDDQU YMM1, [EAX+32]
VLDDQU YMM2, [EAX+64]
VLDDQU YMM3, [EAX+96]
VMOVDQU [EDX+00], YMM0
VMOVDQU [EDX+32], YMM1
VMOVDQU [EDX+64], YMM2
VMOVDQU [EDX+96], YMM3
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FF, $F0, $00
DB $67, $C5, $FF, $F0, $48, $20
DB $67, $C5, $FF, $F0, $50, $40
DB $67, $C5, $FF, $F0, $58, $60
DB $67, $C5, $FE, $7F, $02
DB $67, $C5, $FE, $7F, $4A, $20
DB $67, $C5, $FE, $7F, $52, $40
DB $67, $C5, $FE, $7F, $5A, $60
{$ELSE}
DB $C5, $FF, $F0, $00
DB $C5, $FF, $F0, $48, $20
DB $C5, $FF, $F0, $50, $40
DB $C5, $FF, $F0, $58, $60
DB $C5, $FE, $7F, $02
DB $C5, $FE, $7F, $4A, $20
DB $C5, $FE, $7F, $52, $40
DB $C5, $FE, $7F, $5A, $60
{$ENDIF}
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
end;每一次复制 256 个字节：
{ 每一次复制 256 个字节 <2048 位> <长度必须是 256 的倍数> }
procedure Move_AVX1_256U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VLDDQU YMM0, [EAX+000]
VLDDQU YMM1, [EAX+032]
VLDDQU YMM2, [EAX+064]
VLDDQU YMM3, [EAX+096]
VLDDQU YMM4, [EAX+128]
VLDDQU YMM5, [EAX+160]
VLDDQU YMM6, [EAX+192]
VLDDQU YMM7, [EAX+224]
VMOVDQU [EDX+000], YMM0
VMOVDQU [EDX+032], YMM1
VMOVDQU [EDX+064], YMM2
VMOVDQU [EDX+096], YMM3
VMOVDQU [EDX+128], YMM4
VMOVDQU [EDX+160], YMM5
VMOVDQU [EDX+192], YMM6
VMOVDQU [EDX+224], YMM7
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FF, $F0, $00
DB $67, $C5, $FF, $F0, $48, $20
DB $67, $C5, $FF, $F0, $50, $40
DB $67, $C5, $FF, $F0, $58, $60
DB $67, $C5, $FF, $F0, $A0, $80, $00, $00, $00
DB $67, $C5, $FF, $F0, $A8, $A0, $00, $00, $00
DB $67, $C5, $FF, $F0, $B0, $C0, $00, $00, $00
DB $67, $C5, $FF, $F0, $B8, $E0, $00, $00, $00
DB $67, $C5, $FE, $7F, $02
DB $67, $C5, $FE, $7F, $4A, $20
DB $67, $C5, $FE, $7F, $52, $40
DB $67, $C5, $FE, $7F, $5A, $60
DB $67, $C5, $FE, $7F, $A2, $80, $00, $00, $00
DB $67, $C5, $FE, $7F, $AA, $A0, $00, $00, $00
DB $67, $C5, $FE, $7F, $B2, $C0, $00, $00, $00
DB $67, $C5, $FE, $7F, $BA, $E0, $00, $00, $00
{$ELSE}
DB $C5, $FF, $F0, $00
DB $C5, $FF, $F0, $48, $20
DB $C5, $FF, $F0, $50, $40
DB $C5, $FF, $F0, $58, $60
DB $C5, $FF, $F0, $A0, $80, $00, $00, $00
DB $C5, $FF, $F0, $A8, $A0, $00, $00, $00
DB $C5, $FF, $F0, $B0, $C0, $00, $00, $00
DB $C5, $FF, $F0, $B8, $E0, $00, $00, $00
DB $C5, $FE, $7F, $02
DB $C5, $FE, $7F, $4A, $20
DB $C5, $FE, $7F, $52, $40
DB $C5, $FE, $7F, $5A, $60
DB $C5, $FE, $7F, $A2, $80, $00, $00, $00
DB $C5, $FE, $7F, $AA, $A0, $00, $00, $00
DB $C5, $FE, $7F, $B2, $C0, $00, $00, $00
DB $C5, $FE, $7F, $BA, $E0, $00, $00, $00
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
end;每一次复制 512 个字节：
{ 每一次复制 512 个字节 <4096 位> <长度必须是 512 的倍数> }
procedure Move_AVX1_512U(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
{$IFDEF WIN64}
VLDDQU YMM0, [EAX+000]
VLDDQU YMM1, [EAX+032]
VLDDQU YMM2, [EAX+064]
VLDDQU YMM3, [EAX+096]
VLDDQU YMM4, [EAX+128]
VLDDQU YMM5, [EAX+160]
VLDDQU YMM6, [EAX+192]
VLDDQU YMM7, [EAX+224]
VLDDQU YMM8, [EAX+256]
VLDDQU YMM9, [EAX+288]
VLDDQU YMM10, [EAX+320]
VLDDQU YMM11, [EAX+352]
VLDDQU YMM12, [EAX+384]
VLDDQU YMM13, [EAX+416]
VLDDQU YMM14, [EAX+448]
VLDDQU YMM15, [EAX+480]
VMOVDQU [EDX+000], YMM0
VMOVDQU [EDX+032], YMM1
VMOVDQU [EDX+064], YMM2
VMOVDQU [EDX+096], YMM3
VMOVDQU [EDX+128], YMM4
VMOVDQU [EDX+160], YMM5
VMOVDQU [EDX+192], YMM6
VMOVDQU [EDX+224], YMM7
VMOVDQU [EDX+256], YMM8
VMOVDQU [EDX+288], YMM9
VMOVDQU [EDX+320], YMM10
VMOVDQU [EDX+352], YMM11
VMOVDQU [EDX+384], YMM12
VMOVDQU [EDX+416], YMM13
VMOVDQU [EDX+448], YMM14
VMOVDQU [EDX+480], YMM15
{$ELSE}
VLDDQU YMM0, [EAX+000]
VLDDQU YMM1, [EAX+032]
VLDDQU YMM2, [EAX+064]
VLDDQU YMM3, [EAX+096]
VLDDQU YMM4, [EAX+128]
VLDDQU YMM5, [EAX+160]
VLDDQU YMM6, [EAX+192]
VLDDQU YMM7, [EAX+224]
VMOVDQU [EDX+000], YMM0
VMOVDQU [EDX+032], YMM1
VMOVDQU [EDX+064], YMM2
VMOVDQU [EDX+096], YMM3
VMOVDQU [EDX+128], YMM4
VMOVDQU [EDX+160], YMM5
VMOVDQU [EDX+192], YMM6
VMOVDQU [EDX+224], YMM7
VLDDQU YMM0, [EAX+256]
VLDDQU YMM1, [EAX+288]
VLDDQU YMM2, [EAX+320]
VLDDQU YMM3, [EAX+352]
VLDDQU YMM4, [EAX+384]
VLDDQU YMM5, [EAX+416]
VLDDQU YMM6, [EAX+448]
VLDDQU YMM7, [EAX+480]
VMOVDQU [EDX+256], YMM0
VMOVDQU [EDX+288], YMM1
VMOVDQU [EDX+320], YMM2
VMOVDQU [EDX+352], YMM3
VMOVDQU [EDX+384], YMM4
VMOVDQU [EDX+416], YMM5
VMOVDQU [EDX+448], YMM6
VMOVDQU [EDX+480], YMM7
{$ENDIF}
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FF, $F0, $00
DB $67, $C5, $FF, $F0, $48, $20
DB $67, $C5, $FF, $F0, $50, $40
DB $67, $C5, $FF, $F0, $58, $60
DB $67, $C5, $FF, $F0, $A0, $80, $00, $00, $00
DB $67, $C5, $FF, $F0, $A8, $A0, $00, $00, $00
DB $67, $C5, $FF, $F0, $B0, $C0, $00, $00, $00
DB $67, $C5, $FF, $F0, $B8, $E0, $00, $00, $00
DB $67, $C5, $7F, $F0, $80, $00, $01, $00, $00
DB $67, $C5, $7F, $F0, $88, $20, $01, $00, $00
DB $67, $C5, $7F, $F0, $90, $40, $01, $00, $00
DB $67, $C5, $7F, $F0, $98, $60, $01, $00, $00
DB $67, $C5, $7F, $F0, $A0, $80, $01, $00, $00
DB $67, $C5, $7F, $F0, $A8, $A0, $01, $00, $00
DB $67, $C5, $7F, $F0, $B0, $C0, $01, $00, $00
DB $67, $C5, $7F, $F0, $B8, $E0, $01, $00, $00
DB $67, $C5, $FE, $7F, $02
DB $67, $C5, $FE, $7F, $4A, $20
DB $67, $C5, $FE, $7F, $52, $40
DB $67, $C5, $FE, $7F, $5A, $60
DB $67, $C5, $FE, $7F, $A2, $80, $00, $00, $00
DB $67, $C5, $FE, $7F, $AA, $A0, $00, $00, $00
DB $67, $C5, $FE, $7F, $B2, $C0, $00, $00, $00
DB $67, $C5, $FE, $7F, $BA, $E0, $00, $00, $00
DB $67, $C5, $7E, $7F, $82, $00, $01, $00, $00
DB $67, $C5, $7E, $7F, $8A, $20, $01, $00, $00
DB $67, $C5, $7E, $7F, $92, $40, $01, $00, $00
DB $67, $C5, $7E, $7F, $9A, $60, $01, $00, $00
DB $67, $C5, $7E, $7F, $A2, $80, $01, $00, $00
DB $67, $C5, $7E, $7F, $AA, $A0, $01, $00, $00
DB $67, $C5, $7E, $7F, $B2, $C0, $01, $00, $00
DB $67, $C5, $7E, $7F, $BA, $E0, $01, $00, $00
{$ELSE}
DB $C5, $FF, $F0, $00
DB $C5, $FF, $F0, $48, $20
DB $C5, $FF, $F0, $50, $40
DB $C5, $FF, $F0, $58, $60
DB $C5, $FF, $F0, $A0, $80, $00, $00, $00
DB $C5, $FF, $F0, $A8, $A0, $00, $00, $00
DB $C5, $FF, $F0, $B0, $C0, $00, $00, $00
DB $C5, $FF, $F0, $B8, $E0, $00, $00, $00
DB $C5, $FE, $7F, $02
DB $C5, $FE, $7F, $4A, $20
DB $C5, $FE, $7F, $52, $40
DB $C5, $FE, $7F, $5A, $60
DB $C5, $FE, $7F, $A2, $80, $00, $00, $00
DB $C5, $FE, $7F, $AA, $A0, $00, $00, $00
DB $C5, $FE, $7F, $B2, $C0, $00, $00, $00
DB $C5, $FE, $7F, $BA, $E0, $00, $00, $00
DB $C5, $FF, $F0, $80, $00, $01, $00, $00
DB $C5, $FF, $F0, $88, $20, $01, $00, $00
DB $C5, $FF, $F0, $90, $40, $01, $00, $00
DB $C5, $FF, $F0, $98, $60, $01, $00, $00
DB $C5, $FF, $F0, $A0, $80, $01, $00, $00
DB $C5, $FF, $F0, $A8, $A0, $01, $00, $00
DB $C5, $FF, $F0, $B0, $C0, $01, $00, $00
DB $C5, $FF, $F0, $B8, $E0, $01, $00, $00
DB $C5, $FE, $7F, $82, $00, $01, $00, $00
DB $C5, $FE, $7F, $8A, $20, $01, $00, $00
DB $C5, $FE, $7F, $92, $40, $01, $00, $00
DB $C5, $FE, $7F, $9A, $60, $01, $00, $00
DB $C5, $FE, $7F, $A2, $80, $01, $00, $00
DB $C5, $FE, $7F, $AA, $A0, $01, $00, $00
DB $C5, $FE, $7F, $B2, $C0, $01, $00, $00
DB $C5, $FE, $7F, $BA, $E0, $01, $00, $00
{$ENDIF}
ADD EAX, 512
ADD EDX, 512
SUB ECX, 512
JNZ @LOOP
end;以上内存复制函数，都是内存未按16字节对齐的情况。
SSE2/SSE4/AVX1/AVX2，还支持字节对齐时的内存复制。效率较高。而且，SSE4 还支持不通过 cache，直接的内存复制。
什么是 16 字节对齐？
内存地址能被 16 整除。即内存地址的最后一位必须是 0；
操作的俩个数的内存地址，都要是 16 字节对齐，才能使用 SSE 对齐指令；
什么是 32 字节对齐？
内存地址能被 32 整除。
内存地址的最后两位必须类似于： $..20, $..40, $..60, $..80, $..A0, $..C0, $..E0 。
即内存地址的最后一位必须是 0，倒数第二位必须是偶数。
操作的俩个数的内存地址，都要是 32 字节对齐，才能使用 AVX 对齐指令；
x64 默认内存地址按 16 字节对齐；
内存复制速度：
1、16 字节对齐 > 未 16 字节对齐
2、AVX2 > AVX1 > SSE4 > SSE2
下面是内存对齐时的内存复制函数（SSE2、SSE4：内存必须按 16 字节对齐；AVX1、AVX2：内存必须按 32 字节对齐）：
{ 每一次复制 16 个字节 <128 位> <长度必须是 16 的倍数> }
procedure Move_SSE2_16A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVDQA XMM0, [EAX]
MOVDQA [EDX], XMM0
ADD EAX, 16
ADD EDX, 16
SUB ECX, 16
JNZ @LOOP
end;
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_SSE2_32A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVDQA XMM0, [EAX+00]
MOVDQA XMM1, [EAX+16]
MOVDQA [EDX+00], XMM0
MOVDQA [EDX+16], XMM1
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
end;
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_SSE2_64A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVDQA XMM0, [EAX+00]
MOVDQA XMM1, [EAX+16]
MOVDQA XMM2, [EAX+32]
MOVDQA XMM3, [EAX+48]
MOVDQA [EDX+00], XMM0
MOVDQA [EDX+16], XMM1
MOVDQA [EDX+32], XMM2
MOVDQA [EDX+48], XMM3
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
end;
{ 每一次复制 128 个字节 <1024 位> <长度必须是 128 的倍数> }
procedure Move_SSE2_128A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVDQA XMM0, [EAX+000]
MOVDQA XMM1, [EAX+016]
MOVDQA XMM2, [EAX+032]
MOVDQA XMM3, [EAX+048]
MOVDQA XMM4, [EAX+064]
MOVDQA XMM5, [EAX+080]
MOVDQA XMM6, [EAX+096]
MOVDQA XMM7, [EAX+112]
MOVDQA [EDX+000], XMM0
MOVDQA [EDX+016], XMM1
MOVDQA [EDX+032], XMM2
MOVDQA [EDX+048], XMM3
MOVDQA [EDX+064], XMM4
MOVDQA [EDX+080], XMM5
MOVDQA [EDX+096], XMM6
MOVDQA [EDX+112], XMM7
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
end;
{ 每一次复制 256 个字节 <2048 位> <长度必须是 256 的倍数> }
procedure Move_SSE2_256A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
@LOOP:
MOVDQA XMM0, [EAX+000]
MOVDQA XMM1, [EAX+016]
MOVDQA XMM2, [EAX+032]
MOVDQA XMM3, [EAX+048]
MOVDQA XMM4, [EAX+064]
MOVDQA XMM5, [EAX+080]
MOVDQA XMM6, [EAX+096]
MOVDQA XMM7, [EAX+112]
MOVDQA XMM8, [EAX+128]
MOVDQA XMM9, [EAX+144]
MOVDQA XMM10, [EAX+160]
MOVDQA XMM11, [EAX+176]
MOVDQA XMM12, [EAX+192]
MOVDQA XMM13, [EAX+208]
MOVDQA XMM14, [EAX+224]
MOVDQA XMM15, [EAX+240]
MOVDQA [EDX+000], XMM0
MOVDQA [EDX+016], XMM1
MOVDQA [EDX+032], XMM2
MOVDQA [EDX+048], XMM3
MOVDQA [EDX+064], XMM4
MOVDQA [EDX+080], XMM5
MOVDQA [EDX+096], XMM6
MOVDQA [EDX+112], XMM7
MOVDQA [EDX+128], XMM8
MOVDQA [EDX+144], XMM9
MOVDQA [EDX+160], XMM10
MOVDQA [EDX+176], XMM11
MOVDQA [EDX+192], XMM12
MOVDQA [EDX+208], XMM13
MOVDQA [EDX+224], XMM14
MOVDQA [EDX+240], XMM15
{$ELSE}
@LOOP:
MOVDQA XMM0, [EAX+000]
MOVDQA XMM1, [EAX+016]
MOVDQA XMM2, [EAX+032]
MOVDQA XMM3, [EAX+048]
MOVDQA XMM4, [EAX+064]
MOVDQA XMM5, [EAX+080]
MOVDQA XMM6, [EAX+096]
MOVDQA XMM7, [EAX+112]
MOVDQA [EDX+000], XMM0
MOVDQA [EDX+016], XMM1
MOVDQA [EDX+032], XMM2
MOVDQA [EDX+048], XMM3
MOVDQA [EDX+064], XMM4
MOVDQA [EDX+080], XMM5
MOVDQA [EDX+096], XMM6
MOVDQA [EDX+112], XMM7
MOVDQA XMM0, [EAX+128]
MOVDQA XMM1, [EAX+144]
MOVDQA XMM2, [EAX+160]
MOVDQA XMM3, [EAX+176]
MOVDQA XMM4, [EAX+192]
MOVDQA XMM5, [EAX+208]
MOVDQA XMM6, [EAX+224]
MOVDQA XMM7, [EAX+240]
MOVDQA [EDX+128], XMM0
MOVDQA [EDX+144], XMM1
MOVDQA [EDX+160], XMM2
MOVDQA [EDX+176], XMM3
MOVDQA [EDX+192], XMM4
MOVDQA [EDX+208], XMM5
MOVDQA [EDX+224], XMM6
MOVDQA [EDX+240], XMM7
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
end;
{ 每一次复制 16 个字节 <128 位> <长度必须是 16 的倍数> }
procedure Move_SSE4_16A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVNTDQA XMM0, [EAX]
MOVNTDQ [EDX], XMM0
ADD EAX, 16
ADD EDX, 16
SUB ECX, 16
JNZ @LOOP
SFENCE
end;
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_SSE4_32A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVNTDQA XMM0, [EAX + 00]
MOVNTDQA XMM1, [EAX + 16]
MOVNTDQ [EDX + 00], XMM0
MOVNTDQ [EDX + 16], XMM1
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
SFENCE
end;
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_SSE4_64A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVNTDQA XMM0, [EAX + 00]
MOVNTDQA XMM1, [EAX + 16]
MOVNTDQA XMM2, [EAX + 32]
MOVNTDQA XMM3, [EAX + 48]
MOVNTDQ [EDX + 00], XMM0
MOVNTDQ [EDX + 16], XMM1
MOVNTDQ [EDX + 32], XMM2
MOVNTDQ [EDX + 48], XMM3
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
SFENCE
end;
{ 每一次复制 128 个字节 <1024 位> <长度必须是 128 的倍数> }
procedure Move_SSE4_128A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
MOVNTDQA XMM0, [EAX + 00]
MOVNTDQA XMM1, [EAX + 16]
MOVNTDQA XMM2, [EAX + 32]
MOVNTDQA XMM3, [EAX + 48]
MOVNTDQA XMM4, [EAX + 64]
MOVNTDQA XMM5, [EAX + 80]
MOVNTDQA XMM6, [EAX + 96]
MOVNTDQA XMM7, [EAX +112]
MOVNTDQ [EDX + 00], XMM0
MOVNTDQ [EDX + 16], XMM1
MOVNTDQ [EDX + 32], XMM2
MOVNTDQ [EDX + 48], XMM3
MOVNTDQ [EDX + 64], XMM4
MOVNTDQ [EDX + 80], XMM5
MOVNTDQ [EDX + 96], XMM6
MOVNTDQ [EDX +112], XMM7
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
SFENCE
end;
{ 每一次复制 256 个字节 <1024 位> <长度必须是 256 的倍数> }
procedure Move_SSE4_256A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
@LOOP:
MOVNTDQA XMM0, [EAX + 00]
MOVNTDQA XMM1, [EAX + 16]
MOVNTDQA XMM2, [EAX + 32]
MOVNTDQA XMM3, [EAX + 48]
MOVNTDQA XMM4, [EAX + 64]
MOVNTDQA XMM5, [EAX + 80]
MOVNTDQA XMM6, [EAX + 96]
MOVNTDQA XMM7, [EAX +112]
MOVNTDQA XMM8, [EAX +128]
MOVNTDQA XMM9, [EAX +144]
MOVNTDQA XMM10, [EAX +160]
MOVNTDQA XMM11, [EAX +176]
MOVNTDQA XMM12, [EAX +192]
MOVNTDQA XMM13, [EAX +208]
MOVNTDQA XMM14, [EAX +224]
MOVNTDQA XMM15, [EAX +240]
MOVNTDQ [EDX + 00], XMM0
MOVNTDQ [EDX + 16], XMM1
MOVNTDQ [EDX + 32], XMM2
MOVNTDQ [EDX + 48], XMM3
MOVNTDQ [EDX + 64], XMM4
MOVNTDQ [EDX + 80], XMM5
MOVNTDQ [EDX + 96], XMM6
MOVNTDQ [EDX +112], XMM7
MOVNTDQ [EDX +128], XMM8
MOVNTDQ [EDX +144], XMM9
MOVNTDQ [EDX +160], XMM10
MOVNTDQ [EDX +176], XMM11
MOVNTDQ [EDX +192], XMM12
MOVNTDQ [EDX +208], XMM13
MOVNTDQ [EDX +224], XMM14
MOVNTDQ [EDX +240], XMM15
{$ELSE}
@LOOP:
MOVNTDQA XMM0, [EAX + 00]
MOVNTDQA XMM1, [EAX + 16]
MOVNTDQA XMM2, [EAX + 32]
MOVNTDQA XMM3, [EAX + 48]
MOVNTDQA XMM4, [EAX + 64]
MOVNTDQA XMM5, [EAX + 80]
MOVNTDQA XMM6, [EAX + 96]
MOVNTDQA XMM7, [EAX +112]
MOVNTDQ [EDX + 00], XMM0
MOVNTDQ [EDX + 16], XMM1
MOVNTDQ [EDX + 32], XMM2
MOVNTDQ [EDX + 48], XMM3
MOVNTDQ [EDX + 64], XMM4
MOVNTDQ [EDX + 80], XMM5
MOVNTDQ [EDX + 96], XMM6
MOVNTDQ [EDX +112], XMM7
MOVNTDQA XMM0, [EAX +128]
MOVNTDQA XMM1, [EAX +144]
MOVNTDQA XMM2, [EAX +160]
MOVNTDQA XMM3, [EAX +176]
MOVNTDQA XMM4, [EAX +192]
MOVNTDQA XMM5, [EAX +208]
MOVNTDQA XMM6, [EAX +224]
MOVNTDQA XMM7, [EAX +240]
MOVNTDQ [EDX +128], XMM0
MOVNTDQ [EDX +144], XMM1
MOVNTDQ [EDX +160], XMM2
MOVNTDQ [EDX +176], XMM3
MOVNTDQ [EDX +192], XMM4
MOVNTDQ [EDX +208], XMM5
MOVNTDQ [EDX +224], XMM6
MOVNTDQ [EDX +240], XMM7
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
SFENCE
end;
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_AVX1_32A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVDQA YMM0, [EAX]
VMOVDQA [EDX], YMM0
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FD, $6F, $00
DB $67, $C5, $FD, $7F, $02
{$ELSE}
DB $C5, $FD, $6F, $00
DB $C5, $FD, $7F, $02
{$ENDIF}
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
end;
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_AVX1_64A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVDQA YMM0, [EAX + 00]
VMOVDQA YMM1, [EAX + 32]
VMOVDQA [EDX + 00], YMM0
VMOVDQA [EDX + 32], YMM1
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FD, $6F, $00
DB $67, $C5, $FD, $6F, $48, $20
DB $67, $C5, $FD, $7F, $02
DB $67, $C5, $FD, $7F, $4A, $20
{$ELSE}
DB $C5, $FD, $6F, $00
DB $C5, $FD, $6F, $48, $20
DB $C5, $FD, $7F, $02
DB $C5, $FD, $7F, $4A, $20
{$ENDIF}
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
end;
{ 每一次复制 128 个字节 <1024 位> <长度必须是 128 的倍数> }
procedure Move_AVX1_128A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVDQA YMM0, [EAX + 00]
VMOVDQA YMM1, [EAX + 32]
VMOVDQA YMM2, [EAX + 64]
VMOVDQA YMM3, [EAX + 96]
VMOVDQA [EDX + 00], YMM0
VMOVDQA [EDX + 32], YMM1
VMOVDQA [EDX + 64], YMM2
VMOVDQA [EDX + 96], YMM3
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FD, $6F, $00
DB $67, $C5, $FD, $6F, $48, $20
DB $67, $C5, $FD, $6F, $50, $40
DB $67, $C5, $FD, $6F, $58, $60
DB $67, $C5, $FD, $7F, $02
DB $67, $C5, $FD, $7F, $4A, $20
DB $67, $C5, $FD, $7F, $52, $40
DB $67, $C5, $FD, $7F, $5A, $60
{$ELSE}
DB $C5, $FD, $6F, $00
DB $C5, $FD, $6F, $48, $20
DB $C5, $FD, $6F, $50, $40
DB $C5, $FD, $6F, $58, $60
DB $C5, $FD, $7F, $02
DB $C5, $FD, $7F, $4A, $20
DB $C5, $FD, $7F, $52, $40
DB $C5, $FD, $7F, $5A, $60
{$ENDIF}
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
end;
{ 每一次复制 256 个字节 <2048 位> <长度必须是 256 的倍数> }
procedure Move_AVX1_256A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVDQA YMM0, [EAX+000]
VMOVDQA YMM1, [EAX+032]
VMOVDQA YMM2, [EAX+064]
VMOVDQA YMM3, [EAX+096]
VMOVDQA YMM4, [EAX+128]
VMOVDQA YMM5, [EAX+160]
VMOVDQA YMM6, [EAX+192]
VMOVDQA YMM7, [EAX+224]
VMOVDQA [EDX+000], YMM0
VMOVDQA [EDX+032], YMM1
VMOVDQA [EDX+064], YMM2
VMOVDQA [EDX+096], YMM3
VMOVDQA [EDX+128], YMM4
VMOVDQA [EDX+160], YMM5
VMOVDQA [EDX+192], YMM6
VMOVDQA [EDX+224], YMM7
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FD, $6F, $00
DB $67, $C5, $FD, $6F, $48, $20
DB $67, $C5, $FD, $6F, $50, $40
DB $67, $C5, $FD, $6F, $58, $60
DB $67, $C5, $FD, $6F, $A0, $80, $00, $00, $00
DB $67, $C5, $FD, $6F, $A8, $A0, $00, $00, $00
DB $67, $C5, $FD, $6F, $B0, $C0, $00, $00, $00
DB $67, $C5, $FD, $6F, $B8, $E0, $00, $00, $00
DB $67, $C5, $FD, $7F, $02
DB $67, $C5, $FD, $7F, $4A, $20
DB $67, $C5, $FD, $7F, $52, $40
DB $67, $C5, $FD, $7F, $5A, $60
DB $67, $C5, $FD, $7F, $A2, $80, $00, $00, $00
DB $67, $C5, $FD, $7F, $AA, $A0, $00, $00, $00
DB $67, $C5, $FD, $7F, $B2, $C0, $00, $00, $00
DB $67, $C5, $FD, $7F, $BA, $E0, $00, $00, $00
{$ELSE}
DB $C5, $FD, $6F, $00
DB $C5, $FD, $6F, $48, $20
DB $C5, $FD, $6F, $50, $40
DB $C5, $FD, $6F, $58, $60
DB $C5, $FD, $6F, $A0, $80, $00, $00, $00
DB $C5, $FD, $6F, $A8, $A0, $00, $00, $00
DB $C5, $FD, $6F, $B0, $C0, $00, $00, $00
DB $C5, $FD, $6F, $B8, $E0, $00, $00, $00
DB $C5, $FD, $7F, $02
DB $C5, $FD, $7F, $4A, $20
DB $C5, $FD, $7F, $52, $40
DB $C5, $FD, $7F, $5A, $60
DB $C5, $FD, $7F, $A2, $80, $00, $00, $00
DB $C5, $FD, $7F, $AA, $A0, $00, $00, $00
DB $C5, $FD, $7F, $B2, $C0, $00, $00, $00
DB $C5, $FD, $7F, $BA, $E0, $00, $00, $00
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
end;
{ 每一次复制 512 个字节 <4096 位> <长度必须是 512 的倍数> }
procedure Move_AVX1_512A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
{$IFDEF WIN64}
VMOVDQA YMM0, [EAX+000]
VMOVDQA YMM1, [EAX+032]
VMOVDQA YMM2, [EAX+064]
VMOVDQA YMM3, [EAX+096]
VMOVDQA YMM4, [EAX+128]
VMOVDQA YMM5, [EAX+160]
VMOVDQA YMM6, [EAX+192]
VMOVDQA YMM7, [EAX+224]
VMOVDQA YMM8, [EAX+256]
VMOVDQA YMM9, [EAX+288]
VMOVDQA YMM10, [EAX+320]
VMOVDQA YMM11, [EAX+352]
VMOVDQA YMM12, [EAX+384]
VMOVDQA YMM13, [EAX+416]
VMOVDQA YMM14, [EAX+448]
VMOVDQA YMM15, [EAX+480]
VMOVDQA [EDX+000], YMM0
VMOVDQA [EDX+032], YMM1
VMOVDQA [EDX+064], YMM2
VMOVDQA [EDX+096], YMM3
VMOVDQA [EDX+128], YMM4
VMOVDQA [EDX+160], YMM5
VMOVDQA [EDX+192], YMM6
VMOVDQA [EDX+224], YMM7
VMOVDQA [EDX+256], YMM8
VMOVDQA [EDX+288], YMM9
VMOVDQA [EDX+320], YMM10
VMOVDQA [EDX+352], YMM11
VMOVDQA [EDX+384], YMM12
VMOVDQA [EDX+416], YMM13
VMOVDQA [EDX+448], YMM14
VMOVDQA [EDX+480], YMM15
{$ELSE}
VMOVDQA YMM0, [EAX+000]
VMOVDQA YMM1, [EAX+032]
VMOVDQA YMM2, [EAX+064]
VMOVDQA YMM3, [EAX+096]
VMOVDQA YMM4, [EAX+128]
VMOVDQA YMM5, [EAX+160]
VMOVDQA YMM6, [EAX+192]
VMOVDQA YMM7, [EAX+224]
VMOVDQA [EDX+000], YMM0
VMOVDQA [EDX+032], YMM1
VMOVDQA [EDX+064], YMM2
VMOVDQA [EDX+096], YMM3
VMOVDQA [EDX+128], YMM4
VMOVDQA [EDX+160], YMM5
VMOVDQA [EDX+192], YMM6
VMOVDQA [EDX+224], YMM7
VMOVDQA YMM0, [EAX+256]
VMOVDQA YMM1, [EAX+288]
VMOVDQA YMM2, [EAX+320]
VMOVDQA YMM3, [EAX+352]
VMOVDQA YMM4, [EAX+384]
VMOVDQA YMM5, [EAX+416]
VMOVDQA YMM6, [EAX+448]
VMOVDQA YMM7, [EAX+480]
VMOVDQA [EDX+256], YMM0
VMOVDQA [EDX+288], YMM1
VMOVDQA [EDX+320], YMM2
VMOVDQA [EDX+352], YMM3
VMOVDQA [EDX+384], YMM4
VMOVDQA [EDX+416], YMM5
VMOVDQA [EDX+448], YMM6
VMOVDQA [EDX+480], YMM7
{$ENDIF}
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C5, $FD, $6F, $00
DB $67, $C5, $FD, $6F, $48, $20
DB $67, $C5, $FD, $6F, $50, $40
DB $67, $C5, $FD, $6F, $58, $60
DB $67, $C5, $FD, $6F, $A0, $80, $00, $00, $00
DB $67, $C5, $FD, $6F, $A8, $A0, $00, $00, $00
DB $67, $C5, $FD, $6F, $B0, $C0, $00, $00, $00
DB $67, $C5, $FD, $6F, $B8, $E0, $00, $00, $00
DB $67, $C5, $7D, $6F, $80, $00, $01, $00, $00
DB $67, $C5, $7D, $6F, $88, $20, $01, $00, $00
DB $67, $C5, $7D, $6F, $90, $40, $01, $00, $00
DB $67, $C5, $7D, $6F, $98, $60, $01, $00, $00
DB $67, $C5, $7D, $6F, $A0, $80, $01, $00, $00
DB $67, $C5, $7D, $6F, $A8, $A0, $01, $00, $00
DB $67, $C5, $7D, $6F, $B0, $C0, $01, $00, $00
DB $67, $C5, $7D, $6F, $B8, $E0, $01, $00, $00
DB $67, $C5, $FD, $7F, $02
DB $67, $C5, $FD, $7F, $4A, $20
DB $67, $C5, $FD, $7F, $52, $40
DB $67, $C5, $FD, $7F, $5A, $60
DB $67, $C5, $FD, $7F, $A2, $80, $00, $00, $00
DB $67, $C5, $FD, $7F, $AA, $A0, $00, $00, $00
DB $67, $C5, $FD, $7F, $B2, $C0, $00, $00, $00
DB $67, $C5, $FD, $7F, $BA, $E0, $00, $00, $00
DB $67, $C5, $7D, $7F, $82, $00, $01, $00, $00
DB $67, $C5, $7D, $7F, $8A, $20, $01, $00, $00
DB $67, $C5, $7D, $7F, $92, $40, $01, $00, $00
DB $67, $C5, $7D, $7F, $9A, $60, $01, $00, $00
DB $67, $C5, $7D, $7F, $A2, $80, $01, $00, $00
DB $67, $C5, $7D, $7F, $AA, $A0, $01, $00, $00
DB $67, $C5, $7D, $7F, $B2, $C0, $01, $00, $00
DB $67, $C5, $7D, $7F, $BA, $E0, $01, $00, $00
{$ELSE}
DB $C5, $FD, $6F, $00
DB $C5, $FD, $6F, $48, $20
DB $C5, $FD, $6F, $50, $40
DB $C5, $FD, $6F, $58, $60
DB $C5, $FD, $6F, $A0, $80, $00, $00, $00
DB $C5, $FD, $6F, $A8, $A0, $00, $00, $00
DB $C5, $FD, $6F, $B0, $C0, $00, $00, $00
DB $C5, $FD, $6F, $B8, $E0, $00, $00, $00
DB $C5, $7D, $6F, $80, $00, $01, $00, $00
DB $C5, $7D, $6F, $88, $20, $01, $00, $00
DB $C5, $7D, $6F, $90, $40, $01, $00, $00
DB $C5, $7D, $6F, $98, $60, $01, $00, $00
DB $C5, $7D, $6F, $A0, $80, $01, $00, $00
DB $C5, $7D, $6F, $A8, $A0, $01, $00, $00
DB $C5, $7D, $6F, $B0, $C0, $01, $00, $00
DB $C5, $7D, $6F, $B8, $E0, $01, $00, $00
DB $C5, $FD, $7F, $02
DB $C5, $FD, $7F, $4A, $20
DB $C5, $FD, $7F, $52, $40
DB $C5, $FD, $7F, $5A, $60
DB $C5, $FD, $7F, $A2, $80, $00, $00, $00
DB $C5, $FD, $7F, $AA, $A0, $00, $00, $00
DB $C5, $FD, $7F, $B2, $C0, $00, $00, $00
DB $C5, $FD, $7F, $BA, $E0, $00, $00, $00
DB $C5, $7D, $7F, $82, $00, $01, $00, $00
DB $C5, $7D, $7F, $8A, $20, $01, $00, $00
DB $C5, $7D, $7F, $92, $40, $01, $00, $00
DB $C5, $7D, $7F, $9A, $60, $01, $00, $00
DB $C5, $7D, $7F, $A2, $80, $01, $00, $00
DB $C5, $7D, $7F, $AA, $A0, $01, $00, $00
DB $C5, $7D, $7F, $B2, $C0, $01, $00, $00
DB $C5, $7D, $7F, $BA, $E0, $01, $00, $00
{$ENDIF}
ADD EAX, 512
ADD EDX, 512
SUB ECX, 512
JNZ @LOOP
end;
{ 每一次复制 32 个字节 <256 位> <长度必须是 32 的倍数> }
procedure Move_AVX2_32A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVNTDQA YMM0, [EAX]
VMOVNTDQ [EDX], YMM0
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C4, $E2, $7D, $2A, $00
DB $67, $C5, $FD, $E7, $02
{$ELSE}
DB $C4, $E2, $7D, $2A, $00
DB $C5, $FD, $E7, $02
{$ENDIF}
ADD EAX, 32
ADD EDX, 32
SUB ECX, 32
JNZ @LOOP
SFENCE
end;
{ 每一次复制 64 个字节 <512 位> <长度必须是 64 的倍数> }
procedure Move_AVX2_64A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVNTDQA YMM0, [EAX+00]
VMOVNTDQA YMM1, [EAX+32]
VMOVNTDQ [EDX+00], YMM0
VMOVNTDQ [EDX+32], YMM1
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C4, $E2, $7D, $2A, $00
DB $67, $C4, $E2, $7D, $2A, $48, $20
DB $67, $C5, $FD, $E7, $02
DB $67, $C5, $FD, $E7, $4A, $20
{$ELSE}
DB $C4, $E2, $7D, $2A, $00
DB $C4, $E2, $7D, $2A, $48, $20
DB $C5, $FD, $E7, $02
DB $C5, $FD, $E7, $4A, $20
{$ENDIF}
ADD EAX, 64
ADD EDX, 64
SUB ECX, 64
JNZ @LOOP
SFENCE
end;
{ 每一次复制 128 个字节 <512 位> <长度必须是 128 的倍数> }
procedure Move_AVX2_128A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVNTDQA YMM0, [EAX+00]
VMOVNTDQA YMM1, [EAX+32]
VMOVNTDQA YMM2, [EAX+64]
VMOVNTDQA YMM3, [EAX+96]
VMOVNTDQ [EDX+00], YMM0
VMOVNTDQ [EDX+32], YMM1
VMOVNTDQ [EDX+64], YMM2
VMOVNTDQ [EDX+96], YMM3
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C4, $E2, $7D, $2A, $00
DB $67, $C4, $E2, $7D, $2A, $48, $20
DB $67, $C4, $E2, $7D, $2A, $50, $40
DB $67, $C4, $E2, $7D, $2A, $58, $60
DB $67, $C5, $FD, $E7, $02
DB $67, $C5, $FD, $E7, $4A, $20
DB $67, $C5, $FD, $E7, $52, $40
DB $67, $C5, $FD, $E7, $5A, $60
{$ELSE}
DB $C4, $E2, $7D, $2A, $00
DB $C4, $E2, $7D, $2A, $48, $20
DB $C4, $E2, $7D, $2A, $50, $40
DB $C4, $E2, $7D, $2A, $58, $60
DB $C5, $FD, $E7, $02
DB $C5, $FD, $E7, $4A, $20
DB $C5, $FD, $E7, $52, $40
DB $C5, $FD, $E7, $5A, $60
{$ENDIF}
ADD EAX, 128
ADD EDX, 128
SUB ECX, 128
JNZ @LOOP
SFENCE
end;
{ 每一次复制 256 个字节 <1024 位> <长度必须是 256 的倍数> }
procedure Move_AVX2_256A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
VMOVNTDQA YMM0, [EAX+00]
VMOVNTDQA YMM1, [EAX+32]
VMOVNTDQA YMM2, [EAX+64]
VMOVNTDQA YMM3, [EAX+96]
VMOVNTDQA YMM4, [EAX+128]
VMOVNTDQA YMM5, [EAX+160]
VMOVNTDQA YMM6, [EAX+192]
VMOVNTDQA YMM7, [EAX+224]
VMOVNTDQ [EDX+00], YMM0
VMOVNTDQ [EDX+32], YMM1
VMOVNTDQ [EDX+64], YMM2
VMOVNTDQ [EDX+96], YMM3
VMOVNTDQ [EDX+128], YMM4
VMOVNTDQ [EDX+160], YMM5
VMOVNTDQ [EDX+192], YMM6
VMOVNTDQ [EDX+224], YMM7
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C4, $E2, $7D, $2A, $00
DB $67, $C4, $E2, $7D, $2A, $48, $20
DB $67, $C4, $E2, $7D, $2A, $50, $40
DB $67, $C4, $E2, $7D, $2A, $58, $60
DB $67, $C4, $E2, $7D, $2A, $A0, $80, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $A8, $A0, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $B0, $C0, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $B8, $E0, $00, $00, $00
DB $67, $C5, $FD, $E7, $02
DB $67, $C5, $FD, $E7, $4A, $20
DB $67, $C5, $FD, $E7, $52, $40
DB $67, $C5, $FD, $E7, $5A, $60
DB $67, $C5, $FD, $E7, $A2, $80, $00, $00, $00
DB $67, $C5, $FD, $E7, $AA, $A0, $00, $00, $00
DB $67, $C5, $FD, $E7, $B2, $C0, $00, $00, $00
DB $67, $C5, $FD, $E7, $BA, $E0, $00, $00, $00
{$ELSE}
DB $C4, $E2, $7D, $2A, $00
DB $C4, $E2, $7D, $2A, $48, $20
DB $C4, $E2, $7D, $2A, $50, $40
DB $C4, $E2, $7D, $2A, $58, $60
DB $C4, $E2, $7D, $2A, $A0, $80, $00, $00, $00
DB $C4, $E2, $7D, $2A, $A8, $A0, $00, $00, $00
DB $C4, $E2, $7D, $2A, $B0, $C0, $00, $00, $00
DB $C4, $E2, $7D, $2A, $B8, $E0, $00, $00, $00
DB $C5, $FD, $E7, $02
DB $C5, $FD, $E7, $4A, $20
DB $C5, $FD, $E7, $52, $40
DB $C5, $FD, $E7, $5A, $60
DB $C5, $FD, $E7, $A2, $80, $00, $00, $00
DB $C5, $FD, $E7, $AA, $A0, $00, $00, $00
DB $C5, $FD, $E7, $B2, $C0, $00, $00, $00
DB $C5, $FD, $E7, $BA, $E0, $00, $00, $00
{$ENDIF}
ADD EAX, 256
ADD EDX, 256
SUB ECX, 256
JNZ @LOOP
SFENCE
end;
{ 每一次复制 512 个字节 <1024 位> <长度必须是 512 的倍数> }
procedure Move_AVX2_512A(const src: Pointer; dst: Pointer; const len: NativeInt); assembler; register;
asm
{$IFDEF WIN64}
XCHG RAX, RCX
{$ENDIF}
@LOOP:
{$IFDEF FPC}
{$IFDEF WIN64}
VMOVNTDQA YMM0, [EAX+000]
VMOVNTDQA YMM1, [EAX+032]
VMOVNTDQA YMM2, [EAX+064]
VMOVNTDQA YMM3, [EAX+096]
VMOVNTDQA YMM4, [EAX+128]
VMOVNTDQA YMM5, [EAX+160]
VMOVNTDQA YMM6, [EAX+192]
VMOVNTDQA YMM7, [EAX+224]
VMOVNTDQA YMM8, [EAX+256]
VMOVNTDQA YMM9, [EAX+288]
VMOVNTDQA YMM10, [EAX+320]
VMOVNTDQA YMM11, [EAX+352]
VMOVNTDQA YMM12, [EAX+384]
VMOVNTDQA YMM13, [EAX+416]
VMOVNTDQA YMM14, [EAX+448]
VMOVNTDQA YMM15, [EAX+480]
VMOVNTDQ [EDX+000], YMM0
VMOVNTDQ [EDX+032], YMM1
VMOVNTDQ [EDX+064], YMM2
VMOVNTDQ [EDX+096], YMM3
VMOVNTDQ [EDX+128], YMM4
VMOVNTDQ [EDX+160], YMM5
VMOVNTDQ [EDX+192], YMM6
VMOVNTDQ [EDX+224], YMM7
VMOVNTDQ [EDX+256], YMM8
VMOVNTDQ [EDX+288], YMM9
VMOVNTDQ [EDX+320], YMM10
VMOVNTDQ [EDX+352], YMM11
VMOVNTDQ [EDX+384], YMM12
VMOVNTDQ [EDX+416], YMM13
VMOVNTDQ [EDX+448], YMM14
VMOVNTDQ [EDX+480], YMM15
{$ELSE}
VMOVNTDQA YMM0, [EAX+000]
VMOVNTDQA YMM1, [EAX+032]
VMOVNTDQA YMM2, [EAX+064]
VMOVNTDQA YMM3, [EAX+096]
VMOVNTDQA YMM4, [EAX+128]
VMOVNTDQA YMM5, [EAX+160]
VMOVNTDQA YMM6, [EAX+192]
VMOVNTDQA YMM7, [EAX+224]
VMOVNTDQ [EDX+000], YMM0
VMOVNTDQ [EDX+032], YMM1
VMOVNTDQ [EDX+064], YMM2
VMOVNTDQ [EDX+096], YMM3
VMOVNTDQ [EDX+128], YMM4
VMOVNTDQ [EDX+160], YMM5
VMOVNTDQ [EDX+192], YMM6
VMOVNTDQ [EDX+224], YMM7
VMOVNTDQA YMM0, [EAX+256]
VMOVNTDQA YMM1, [EAX+288]
VMOVNTDQA YMM2, [EAX+320]
VMOVNTDQA YMM3, [EAX+352]
VMOVNTDQA YMM4, [EAX+384]
VMOVNTDQA YMM5, [EAX+416]
VMOVNTDQA YMM6, [EAX+448]
VMOVNTDQA YMM7, [EAX+480]
VMOVNTDQ [EDX+256], YMM0
VMOVNTDQ [EDX+288], YMM1
VMOVNTDQ [EDX+320], YMM2
VMOVNTDQ [EDX+352], YMM3
VMOVNTDQ [EDX+384], YMM4
VMOVNTDQ [EDX+416], YMM5
VMOVNTDQ [EDX+448], YMM6
VMOVNTDQ [EDX+480], YMM7
{$ENDIF}
{$ENDIF}
{$IFDEF WIN64}
DB $67, $C4, $E2, $7D, $2A, $00
DB $67, $C4, $E2, $7D, $2A, $48, $20
DB $67, $C4, $E2, $7D, $2A, $50, $40
DB $67, $C4, $E2, $7D, $2A, $58, $60
DB $67, $C4, $E2, $7D, $2A, $A0, $80, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $A8, $A0, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $B0, $C0, $00, $00, $00
DB $67, $C4, $E2, $7D, $2A, $B8, $E0, $00, $00, $00
DB $67, $C4, $62, $7D, $2A, $80, $00, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $88, $20, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $90, $40, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $98, $60, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $A0, $80, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $A8, $A0, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $B0, $C0, $01, $00, $00
DB $67, $C4, $62, $7D, $2A, $B8, $E0, $01, $00, $00
DB $67, $C5, $FD, $E7, $02
DB $67, $C5, $FD, $E7, $4A, $20
DB $67, $C5, $FD, $E7, $52, $40
DB $67, $C5, $FD, $E7, $5A, $60
DB $67, $C5, $FD, $E7, $A2, $80, $00, $00, $00
DB $67, $C5, $FD, $E7, $AA, $A0, $00, $00, $00
DB $67, $C5, $FD, $E7, $B2, $C0, $00, $00, $00
DB $67, $C5, $FD, $E7, $BA, $E0, $00, $00, $00
DB $67, $C5, $7D, $E7, $82, $00, $01, $00, $00
DB $67, $C5, $7D, $E7, $8A, $20, $01, $00, $00
DB $67, $C5, $7D, $E7, $92, $40, $01, $00, $00
DB $67, $C5, $7D, $E7, $9A, $60, $01, $00, $00
DB $67, $C5, $7D, $E7, $A2, $80, $01, $00, $00
DB $67, $C5, $7D, $E7, $AA, $A0, $01, $00, $00
DB $67, $C5, $7D, $E7, $B2, $C0, $01, $00, $00
DB $67, $C5, $7D, $E7, $BA, $E0, $01, $00, $00
{$ELSE}
DB $C4, $E2, $7D, $2A, $00
DB $C4, $E2, $7D, $2A, $48, $20
DB $C4, $E2, $7D, $2A, $50, $40
DB $C4, $E2, $7D, $2A, $58, $60
DB $C4, $E2, $7D, $2A, $A0, $80, $00, $00, $00
DB $C4, $E2, $7D, $2A, $A8, $A0, $00, $00, $00
DB $C4, $E2, $7D, $2A, $B0, $C0, $00, $00, $00
DB $C4, $E2, $7D, $2A, $B8, $E0, $00, $00, $00
DB $C5, $FD, $E7, $02
DB $C5, $FD, $E7, $4A, $20
DB $C5, $FD, $E7, $52, $40
DB $C5, $FD, $E7, $5A, $60
DB $C5, $FD, $E7, $A2, $80, $00, $00, $00
DB $C5, $FD, $E7, $AA, $A0, $00, $00, $00
DB $C5, $FD, $E7, $B2, $C0, $00, $00, $00
DB $C5, $FD, $E7, $BA, $E0, $00, $00, $00
DB $C4, $E2, $7D, $2A, $80, $00, $01, $00, $00
DB $C4, $E2, $7D, $2A, $88, $20, $01, $00, $00
DB $C4, $E2, $7D, $2A, $90, $40, $01, $00, $00
DB $C4, $E2, $7D, $2A, $98, $60, $01, $00, $00
DB $C4, $E2, $7D, $2A, $A0, $80, $01, $00, $00
DB $C4, $E2, $7D, $2A, $A8, $A0, $01, $00, $00
DB $C4, $E2, $7D, $2A, $B0, $C0, $01, $00, $00
DB $C4, $E2, $7D, $2A, $B8, $E, $001, $00, $00
DB $C5, $FD, $E7, $82, $00, $01, $00, $00
DB $C5, $FD, $E7, $8A, $20, $01, $00, $00
DB $C5, $FD, $E7, $92, $40, $01, $00, $00
DB $C5, $FD, $E7, $9A, $60, $01, $00, $00
DB $C5, $FD, $E7, $A2, $80, $01, $00, $00
DB $C5, $FD, $E7, $AA, $A0, $01, $00, $00
DB $C5, $FD, $E7, $B2, $C0, $01, $00, $00
DB $C5, $FD, $E7, $BA, $E0, $01, $00, $00
{$ENDIF}
ADD EAX, 512
ADD EDX, 512
SUB ECX, 512
JNZ @LOOP
SFENCE
end;
水平有限，错误难免。如有错误，还望高手不吝指正。
详细代码：https://github.com/dbyoung720/ImageGray/blob/master/src/db.Memory.Move.pas
qq交流群：101611228
732
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
