// Pipeline Median  and ssmedian

// =============================================
//    パイプラインメディアンフィルタ  速度比2.4
// =============================================

// 第１段
inline unsigned int plfirststage(int *a)
{
    unsigned int s;
    if((a[0]>a[1])&&(a[0]>a[2]))
      { 
          s=a[0]<<16; 
          if(a[1]>a[2])
            s |= (a[1]<<8)|(a[2]);
          else
            s |= (a[2]<<8)|(a[1]);
      }
    else if(a[1]>a[2])
      {
          s=a[1]<<16; 
          if(a[0]>a[2])
            s |= (a[0]<<8)|(a[2]);
          else
            s |= (a[2]<<8)|(a[0]);
      }
    else
      {
          s=a[2]<<16;
          if(a[0]>a[1])
            s |= (a[0]<<8)|(a[1]);
          else
            s |= (a[1]<<8)|(a[0]);
      }
    return s<<8;
}

// 本体
int plmedian(int *a)  
{
    unsigned int s1,s2,s3;
    s1=plfirststage(a+0);
    s2=plfirststage(a+3);
    s3=plfirststage(a+6);

    if((s1>s2)&&(s1>s3))
      s1=s1<<8;
    else if(s2>s3)
      s2=s2<<8;
    else
      s3=s3<<8;

    if((s2>s3)&&(s2>s1))
      s2=s2<<8;
    else if(s3>s1)
      s3=s3<<8;
    else
      s1=s1<<8;

    if((s3>s1)&&(s3>s2))
      s3=s3<<8;
    else if(s1>s2)
      s1=s1<<8;
    else
      s2=s2<<8;

    if((s1>s2)&&(s1>s3))
      s1=s1<<8;
    else if(s2>s3)
      s2=s2<<8;
    else
      s3=s3<<8;

    if((s1>s2)&&(s1>s3))
      return s1>>24;
    else if(s2>s3)
      return s2>>24;
    else
      return s3>>24;
}

#if 0
// =============================================
//   総当たりソート         速度比 １
// =============================================

int ssmedian(int *a)
{
    int i,j,t;
    for(i=0;i<5;i++)
      for(j=i+1;j<9;j++)
        {
            if(a[i]<a[j])
              { t=a[i]; a[i]=a[j]; a[j]=t; }
        }
    return a[4];
}
#endif


#ifdef USEMMX
// =============================================
//   MMX ２ピクセル同時メディアン  速度比 13
// =============================================

void inline MedianA(imgbyte *base,int w)
{
  asm(// 準備
      "\n     mov %0,%%esi"
      "\n     mov %0,%%edi"
      "\n     shl $2,%%ebx"         // w*4
      "\n     add %%ebx,%%edi"
      "\n     add %%ebx,%%edi"      // edi=esi+w*8


      "\n     pcmpeqb %%mm7,%%mm7"
      "\n     psrlq $32,%%mm7"
      "\n     "

      // １巡目
      // データロード
      "\n     movq (%%esi)  ,%%mm0"
      "\n     movq (%%esi,%%ebx) ,%%mm1"
      "\n     movq (%%edi),%%mm2"
      "\n     movq 8(%%esi),%%mm3"
      "\n     movq 8(%%esi,%%ebx),%%mm4"
      "\n     movq 8(%%edi),%%mm5"
      "\n     pand %%mm7,%%mm3"
      "\n     pand %%mm7,%%mm5"
      "\n     psllq $32,%%mm4"
      "\n     por %%mm5,%%mm4"
      "\n     "

      // １：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ２：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ３：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ４：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"

      "\n     movd %%mm2,(%%esi)"



      // ２巡目
      "\n     pcmpeqb %%mm7,%%mm7"
      "\n     psllq $32,%%mm7"
      // データロード
      "\n     movq 8(%%esi)  ,%%mm0"
      "\n     movq 8(%%esi,%%ebx) ,%%mm1"
      "\n     movq 8(%%edi),%%mm2"
      "\n     movq 0(%%esi),%%mm3"
      "\n     movq 0(%%esi,%%ebx),%%mm4"
      "\n     movq 0(%%edi),%%mm5"
      "\n     pand %%mm7,%%mm3"
      "\n     pand %%mm7,%%mm5"
      "\n     psrlq $32,%%mm4"
      "\n     por %%mm5,%%mm4"
      "\n     "

      // １：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ２：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ３：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"
      /////////////////
      // ４：
      "\n     movq %%mm0,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm0"

      "\n     movq %%mm1,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm1"

      "\n     movq %%mm2,%%mm5"
      "\n     psrlq $32,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psllq $32,%%mm5"
      "\n     psubb %%mm5,%%mm2"

      "\n     movq %%mm3,%%mm5"
      "\n     psllq $32,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psrlq $32,%%mm5"
      "\n     psubb %%mm5,%%mm3"

      "\n     movq %%mm4,%%mm5"
      "\n     psubusb %%mm3,%%mm5"
      "\n     paddb %%mm5,%%mm3"
      "\n     psubb %%mm5,%%mm4"
      "\n     movq %%mm3,%%mm5"
      "\n     psubusb %%mm2,%%mm5"
      "\n     paddb %%mm5,%%mm2"
      "\n     psubb %%mm5,%%mm3"
      "\n     movq %%mm2,%%mm5"
      "\n     psubusb %%mm1,%%mm5"
      "\n     paddb %%mm5,%%mm1"
      "\n     psubb %%mm5,%%mm2"
      "\n     movq %%mm1,%%mm5"
      "\n     psubusb %%mm0,%%mm5"
      "\n     paddb %%mm5,%%mm0"
      "\n     psubb %%mm5,%%mm1"

      "\n     movd %%mm2,4(%%esi)"

      : : "a" (base),"b" (w): "esi","edi","bx");
}


#endif