[MMX指令版]黑白棋行动力的计算

类别:编程语言 点击:0 评论:0 推荐:

前言:这个东西老早就应该贴了,因为我已经在Nowcan的Blog帖子里说过了,一直没贴是因为没有心情,或者觉得自己还没有整理得足够好,现在先发未经过整理的上来冲冲数先吧。

这段代码是WZebra的作者Gunnar Andersson写的,作用是计算《黑白棋》双方的行动力,其中使用了bitboard技术和MMX指令,作者说他已经考虑了指令配对,请不要随意修改代码的先后顺序,除非你真的确定你是对的。原来的代码是gcc下的汇编格式,我把他转换过来而已,这个是在VC++ 6.0编译的版本,希望对你有点用处。

另:由于是很久很久以前弄的,大概有快2,3年了吧(当时我的汇编也不怎么样),所以你会看到头和尾的int count处理不是很科学(也许这样比较直观一些),其实你可以让他更有效率,我也懒得改了,修改一下代码会更快一点儿,留给你自己去做了,其实相当简单的。

static unsigned __int64 dir_mask0;
static unsigned __int64 dir_mask1;
static unsigned __int64 dir_mask2;
static unsigned __int64 dir_mask3;
static unsigned __int64 dir_mask4;
static unsigned __int64 dir_mask5;
static unsigned __int64 dir_mask6;
static unsigned __int64 dir_mask7;
static unsigned __int64 c0f;
static unsigned __int64 c33;
static unsigned __int64 c55;

void init_mmx( void )
{
    dir_mask0 = 0x007e7e7e7e7e7e00;
    dir_mask1 = 0x00ffffffffffff00;
    dir_mask2 = 0x007e7e7e7e7e7e00;
    dir_mask3 = 0x7e7e7e7e7e7e7e7e;
    dir_mask4 = 0x7e7e7e7e7e7e7e7e;
    dir_mask5 = 0x007e7e7e7e7e7e00;
    dir_mask6 = 0x00ffffffffffff00;
    dir_mask7 = 0x007e7e7e7e7e7e00;
    c0f = 0x0f0f0f0f0f0f0f0f;
    c33 = 0x3333333333333333;
    c55 = 0x5555555555555555;
}

typedef struct {
    unsigned long high;
    unsigned long low;
} BitBoard;

int bitboard_mobility( const BitBoard my_bits, const BitBoard opp_bits )
{
    unsigned int count;

    __asm {

        /* Ready for init data */
        mov     ebx, my_bits.high       ;
        mov     ecx, my_bits.low        ;
        mov     edi, opp_bits.high      ;
        mov     esi, opp_bits.low       ;

        //
        movd    mm0, ebx                ;
        psllq   mm0, 32                 ;
        movd    mm3, ecx                ;
        por     mm0, mm3                ; mm0 is BitBoard of my_bits
        movd    mm1, edi                ;
        psllq   mm1, 32                 ;
        movd    mm4, esi                ;
        por     mm1, mm4                ; mm1 is BitBoard of opp_bits
        pxor    mm2, mm2                ; mm2 <- 0x0000000000000000

        /* shift=-9   rowDelta=-1   colDelta=-1 */
        /* shift=+9   rowDelta=+1   colDelta=+1 */

        /* Disc #1, flip direction 0. */
        /* Disc #1, flip direction 7. */
        movq    mm3, mm1                ; mm3 <- opp_bits
        movq    mm4, mm0                ; mm4 <- my_bits
        movq    mm6, mm0                ; mm6 <- backup of my_bits
        pand    mm3, dir_mask0          ; 0x007e7e7e7e7e7e00
                                        ; dir_mask0 of value:
                                        ;   00000000
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   00000000
        push    esi                     ;
        psllq   mm4, 9                  ;
        psrlq   mm6, 9                  ;
        push    edi                     ;
        pand    mm4, mm3                ;
        pand    mm6, mm3                ;
        push    ecx                     ;

        /* Disc #2, flip direction 0. */
        /* Disc #2, flip direction 7. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        psllq   mm5, 9                  ;
        psrlq   mm7, 9                  ;
        push    ebx                     ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        and     edi, 0x7e7e7e7e         ; 0x7e7e7e7e
        and     esi, 0x7e7e7e7e         ; 0x7e7e7e7e
                                        ; value of:
                                        ; 011111110
                                        ; 011111110
                                        ; 011111110
                                        ; 011111110
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        shl     ebx, 1                  ;
        shl     ecx, 1                  ;

        /* Disc #3, flip direction 0. */
        /* Disc #3, flip direction 7. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        and     ebx, edi                ;
        and     ecx, esi                ;
        psllq   mm5, 9                  ;
        psrlq   mm7, 9                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shl     edx, 1                  ;
        shl     eax, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;

        /* Disc #4, flip direction 0. */
        /* Disc #4, flip direction 7. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        psllq   mm5, 9                  ;
        psrlq   mm7, 9                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shl     edx, 1                  ;
        shl     eax, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;

        /* Disc #5, flip direction 0. */
        /* Disc #5, flip direction 7. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        psllq   mm5, 9                  ;
        psrlq   mm7, 9                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shl     edx, 1                  ;
        shl     eax, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;

        /* Disc #6, flip direction 0. */
        /* Disc #6, flip direction 7. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        psrlq   mm7, 9                  ;
        psllq   mm5, 9                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shl     edx, 1                  ;
        shl     eax, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm4, 9                  ;
        psrlq   mm6, 9                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        por     mm2, mm4                ;
        por     mm2, mm6                ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;

        /* shift=-8   rowDelta=-1   colDelta=0 */
        /* shift=+8   rowDelta=1   colDelta=0 */

        /* Disc #1, flip direction 1. */
        /* Disc #1, flip direction 6. */
        movq    mm3, mm1                ;
        movq    mm4, mm0                ;
        movq    mm6, mm0                ;
        pand    mm3, dir_mask1          ; 0x00ffffffffffff00;
                                        ; dir_mask1 of value:
                                        ;   00000000
                                        ;   11111111
                                        ;   11111111
                                        ;   11111111
                                        ;   11111111
                                        ;   11111111
                                        ;   11111111
                                        ;   00000000
        psllq   mm4, 8                  ;
        psrlq   mm6, 8                  ;
        shl     edx, 1                  ;
        shl     eax, 1                  ;
        pand    mm4, mm3                ;
        pand    mm6, mm3                ;
        and     eax, edi                ;
        and     edx, esi                ;
        /* Disc #2, flip direction 1. */
        /* Disc #2, flip direction 6. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        psllq   mm5, 8                  ;
        psrlq   mm7, 8                  ;
        shl     ebx, 1                  ;
        shl     ecx, 1                  ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;

        /* serialize here: add horizontal shl flips. */

        movd    mm5, ebx                ;
        psllq   mm5, 32                 ;
        movd    mm7, ecx                ;
        por     mm5, mm7                ;
        por     mm2, mm5                ;

        /* Disc #3, flip direction 1. */
        /* Disc #3, flip direction 6. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        psllq   mm5, 8                  ;
        psrlq   mm7, 8                  ;
        pop     ebx                     ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        pop     ecx                     ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        push    ecx                     ;

        /* Disc #4, flip direction 1. */
        /* Disc #4, flip direction 6. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        push    ebx                     ;
        psllq   mm5, 8                  ;
        psrlq   mm7, 8                  ;
        shr     ebx, 1                  ;
        shr     ecx, 1                  ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        and     ebx, edi                ;
        and     ecx, esi                ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;

        /* Disc #5, flip direction 1. */
        /* Disc #5, flip direction 6. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        psllq   mm5, 8                  ;
        psrlq   mm7, 8                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;

        /* Disc #6, flip direction 1. */
        /* Disc #6, flip direction 6. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        psllq   mm5, 8                  ;
        psrlq   mm7, 8                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm4, 8                  ;
        psrlq   mm6, 8                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        por     mm2, mm4                ;
        por     mm2, mm6                ;

        /* shift=-7   rowDelta=-1   colDelta=1 */
        /* shift=+7   rowDelta=1   colDelta=-1 */

        /* Disc #1, flip direction 2. */
        /* Disc #1, flip direction 5. */
        movq    mm3, mm1                ;
        movq    mm4, mm0                ;
        movq    mm6, mm0                ;
        pand    mm3, dir_mask2          ; 0x007e7e7e7e7e7e00;
                                        ; dir_mask2 of value:
                                        ;   00000000
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   01111110
                                        ;   00000000
        psllq   mm4, 7                  ;
        psrlq   mm6, 7                  ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        pand    mm4, mm3                ;
        pand    mm6, mm3                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;

        /* Disc #2, flip direction 2. */
        /* Disc #2, flip direction 5. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm5, 7                  ;
        psrlq   mm7, 7                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;

        /* Disc #3, flip direction 2. */
        /* Disc #3, flip direction 5. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm5, 7                  ;
        psrlq   mm7, 7                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;

        /* Disc #4, flip direction 2. */
        /* Disc #4, flip direction 5. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm5, 7                  ;
        psrlq   mm7, 7                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        mov     eax, ebx                ;
        mov     edx, ecx                ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        shr     eax, 1                  ;
        shr     edx, 1                  ;

        /* Disc #5, flip direction 2. */
        /* Disc #5, flip direction 5. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        and     eax, edi                ;
        and     edx, esi                ;
        psllq   mm5, 7                  ;
        psrlq   mm7, 7                  ;
        or      ebx, eax                ;
        or      ecx, edx                ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        shr     ebx, 1                  ;
        shr     ecx, 1                  ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;

        /* serialize here: add horizontal shr flips. */

        movd    mm5, ebx                ;
        psllq   mm5, 32                 ;
        movd    mm7, ecx                ;
        por     mm5, mm7                ;
        por     mm2, mm5                ;
        pop     ebx                     ;

        /* Disc #6, flip direction 2. */
        /* Disc #6, flip direction 5. */
        movq    mm5, mm4                ;
        movq    mm7, mm6                ;
        psllq   mm5, 7                  ;
        psrlq   mm7, 7                  ;
        pop     ecx                     ;
        pand    mm5, mm3                ;
        pand    mm7, mm3                ;
        pop     edi                     ;
        por     mm4, mm5                ;
        por     mm6, mm7                ;
        pop     esi                     ;
        psllq   mm4, 7                  ;
        psrlq   mm6, 7                  ;
        por     mm2, mm4                ;
        por     mm2, mm6                ;

        /* mm2 is the pseudo-feasible moves at this point. */
        /* Let mm7 be the feasible moves, i.e., mm2 restricted to empty squares. */

        movq    mm7, mm0                ;
        por     mm7, mm1                ;
        pandn   mm7, mm2                ;

        /* Count the moves, i.e., the number of bits set in mm7. */

        movq    mm1, mm7                ;
        psrld   mm7, 1                  ;
        pand    mm7, c55                ; c55 = 0x5555555555555555
        psubd   mm1, mm7                ;
        movq    mm7, mm1                ;
        psrld   mm1, 2                  ;
        pand    mm7, c33                ; c33 = 0x3333333333333333;
        pand    mm1, c33                ; c33 = 0x3333333333333333;
        paddd   mm7, mm1                ;
        movq    mm1, mm7                ;
        psrld   mm7, 4                  ;
        paddd   mm7, mm1                ;
        pand    mm7, c0f                ; c0f = 0x0f0f0f0f0f0f0f0f;

        movq    mm1, mm7                ;
        psrld   mm7, 8                  ;
        paddd   mm7, mm1                ;
        movq    mm1, mm7                ;
        psrld   mm7, 16                 ;
        paddd   mm7, mm1                ;
        movq    mm1, mm7                ;
        psrlq   mm7, 32                 ;
        paddd   mm7, mm1                ;
        movd    eax, mm7                ;
        and     eax, 63                 ;
        mov     count, eax              ;

        //
        emms                            ;
    }
    return count;
}

本文地址:http://com.8s8s.com/it/it23596.htm