// ---------- loop begin
  lp_mmx_fdct_row1:
    movd mm5, dword ptr [INP+12]; // mm5 = 7 6

    punpcklwd mm5, dword ptr [INP+8] // mm5 =  5 7 4 6

    movq mm2, mm5;     // mm2 = 5 7 4 6
    psrlq mm5, 32;     // mm5 = _ _ 5 7

    movq mm0, qword ptr [INP]; // mm0 = 3 2 1 0
    punpcklwd mm5, mm2;// mm5 = 4 5 6 7

    movq mm1, mm0;     // mm1 = 3 2 1 0
    paddsw mm0, mm5;   // mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)

    psubsw mm1, mm5;   // mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
    movq mm2, mm0;     // mm2 = [ xt3 xt2 xt1 xt0 ]

    //movq [ xt3xt2xt1xt0 ], mm0; // debugging
    //movq [ xt7xt6xt5xt4 ], mm1; // debugging

    punpcklwd mm0, mm1;// mm0 = [ xt5 xt1 xt4 xt0 ]

    punpckhwd mm2, mm1;// mm2 = [ xt7 xt3 xt6 xt2 ]
    movq mm1, mm2;     // mm1

    ;// shuffle bytes around

//  movq mm0, qword ptr [INP] ; 0 ; x3 x2 x1 x0

//  movq mm1, qword ptr [INP+8] ; 1 ; x7 x6 x5 x4
    movq mm2, mm0 ; 2 ; x3 x2 x1 x0

    movq mm3, qword ptr [TABLE] ; 3 ; w06 w04 w02 w00
    punpcklwd mm0, mm1 ; x5 x1 x4 x0

    movq mm5, mm0 ; 5 ; x5 x1 x4 x0
    punpckldq mm0, mm0 ; x4 x0 x4 x0  [ xt2 xt0 xt2 xt0 ]

    movq mm4, qword ptr [TABLE+8] ; 4 ; w07 w05 w03 w01
    punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2

    pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
    movq mm6, mm2 ; 6 ; x7 x3 x6 x2

    movq mm1, qword ptr [TABLE+32] ; 1 ; w22 w20 w18 w16
    punpckldq mm2, mm2 ; x6 x2 x6 x2  [ xt3 xt1 xt3 xt1 ]

    pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
    punpckhdq mm5, mm5 ; x5 x1 x5 x1  [ xt6 xt4 xt6 xt4 ]

    pmaddwd mm0, qword ptr [TABLE+16] ; x4*w14+x0*w12 x4*w10+x0*w08
    punpckhdq mm6, mm6 ; x7 x3 x7 x3  [ xt7 xt5 xt7 xt5 ]

    movq mm7, qword ptr [TABLE+40] ; 7 ; w23 w21 w19 w17
    pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
//mm3 = a1, a0 (y2,y0)
//mm1 = b1, b0 (y3,y1)
//mm0 = a3,a2  (y6,y4)
//mm5 = b3,b2  (y7,y5)

    paddd mm3, qword ptr [round_frw_row] ; +rounder (y2,y0)
    pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17

    pmaddwd mm2, qword ptr [TABLE+24] ; x6*w15+x2*w13 x6*w11+x2*w09
    paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) // now ( y2, y0)

    pmaddwd mm5, qword ptr [TABLE+48] ; x5*w30+x1*w28 x5*w26+x1*w24
    ;//

    pmaddwd mm6, qword ptr [TABLE+56] ; x7*w31+x3*w29 x7*w27+x3*w25
    paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) // now ( y3, y1)

    paddd mm0, qword ptr [round_frw_row] ; +rounder (y6,y4)
    psrad mm3, SHIFT_FRW_ROW_CLIP1 ;// (y2, y0) 

    paddd mm1, qword ptr [round_frw_row] ; +rounder (y3,y1)
    paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) // now (y6, y4)

    paddd mm5, qword ptr [round_frw_row] ; +rounder (y7,y5)
    psrad mm1, SHIFT_FRW_ROW_CLIP1 ;// y1=a1+b1 y0=a0+b0

    paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) // now ( y7, y5)
    psrad mm0, SHIFT_FRW_ROW_CLIP1 ;//y3=a3+b3 y2=a2+b2

    add OUT, 16;  // increment row-output address by 1 row
    psrad mm5, SHIFT_FRW_ROW_CLIP1;// y4=a3-b3 y5=a2-b2

    add INP, 16;  // increment row-address by 1 row
    packssdw mm3, mm0 ;// 0 ; y6 y4 y2 y0, saturate {-32768,+32767}

    packssdw mm1, mm5 ;// 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
    movq mm6, mm3;    // mm0 = y6 y4 y2 y0

    punpcklwd mm3, mm1; // y3 y2 y1 y0
    sub edi, 0x01;   // i = i - 1
    
    punpckhwd mm6, mm1; // y7 y6 y5 y4
    add TABLE,64;  // increment to next table

    psraw mm3, SHIFT_FRW_ROW_CLIP2;  // descale [y3 y2 y1 y0] to {-2048,+2047}

    psraw mm6, SHIFT_FRW_ROW_CLIP2;  // descale [y7 y6 y5 y4] to {-2048,+2047}

    movq qword ptr [OUT-16], mm3 ; 1 ; save y3 y2 y1 y0

    movq qword ptr [OUT-8], mm6 ; 7 ; save y7 y6 y5 y4

    cmp edi, 0x00;
    jg lp_mmx_fdct_row1;  // begin fdct processing on next row
    emms;
    }
    
/*
    ////////////////////////////////////////////////////////////////////////
    //
    // DCT_8_FRW_COL(), equivalent c_code
    //
    // This C-code can be substituted for the same __asm block
    //
    // I found several *DISCREPANCIES* between the AP-922 C-listing 
    // and actual corrected code (shown below).
    //
    ////////////////////////////////////////////////////////////////////////

    sptr = (short *) blk;
    optr = (short *) blk; // output will overwrite source data!

    for ( j = 0; j < 8; j=j+1 ) // dct_frw_col1 loop
    {
      // read source-data column #j into xt[0..7]
      xt[7] = sptr[7*8];
      xt[6] = sptr[6*8];
      xt[5] = sptr[5*8];
      xt[4] = sptr[4*8];

      xt[3] = sptr[3*8];
      xt[2] = sptr[2*8];
      xt[1] = sptr[1*8];
      xt[0] = sptr[0*8];
 
#define  LEFT_SHIFT( x ) ((x) << (SHIFT_FRW_COL) )   // left shift
#define LEFT_SHIFT1( x ) ((x) << (SHIFT_FRW_COL+1) ) // left shift+1

      t0 = LEFT_SHIFT ( xt[0] + xt[7] );
      t1 = LEFT_SHIFT ( xt[1] + xt[6] );
      t2 = LEFT_SHIFT ( xt[2] + xt[5] );
      t3 = LEFT_SHIFT ( xt[3] + xt[4] );
      t4 = LEFT_SHIFT ( xt[3] - xt[4] );
      t5 = LEFT_SHIFT1( xt[2] - xt[5] ); // *** DISCREPANCY
      t6 = LEFT_SHIFT1( xt[1] - xt[6] ); // *** DISCREPANCY
      t7 = LEFT_SHIFT ( xt[0] - xt[7] );
 
      tp03 = t0 + t3;
      tm03 = t0 - t3;
      tp12 = t1 + t2;
      tm12 = t1 - t2;

// pmulhw/pmulhrw emulation macros 
#define X86_PMULHW( X ) ((short) ( ((int)X)>>16 ))   //Intel MMX
//#define X86_PMULHRW( X ) ((short) ( ( (((int)X)>>15)+1) >>1) ) //3DNow-MMX

      optr[0*8] = tp03 + tp12;
      optr[4*8] = tp03 - tp12;
      optr[2*8] = tm03 + X86_PMULHW( tm12 * _tg_2_16 );
      optr[2*8] = optr[2*8] | _one_corr; // one_correction
      optr[6*8] = X86_PMULHW( tm03 * _tg_2_16 ) - tm12;
      optr[6*8] = optr[6*8] | _one_corr; // one_correction
 
      tp65 = X86_PMULHW( (t6 +t5 )*_ocos_4_16 ); // *** DISCREPANCY
      tp65 = tp65 | _one_corr; // one_correction
      tm65 = X86_PMULHW( (t6 -t5 )*_ocos_4_16 ); // *** DISCREPANCY
  
      tp765 = t7 + tp65;
      tm765 = t7 - tp65;
      tp465 = t4 + tm65;
      tm465 = t4 - tm65;
 
      optr[1*8]  = tp765 + X86_PMULHW( tp465 * _tg_1_16 );
      optr[1*8]  = optr[1*8] | _one_corr; // one_correction
      optr[7*8] = X86_PMULHW( tp765 * _tg_1_16 ) - tp465;

//    optr[5*8] = X86_PMULHW( tm765 * _tg_3_16 ) + tm465; // *** DISCREPANCY
      // from pg8 of AP-922,  ICONST = [ const*(2^16) + 0.5 ]
      //                      const * x = PMULHW( ICONST,x ) + x
      // The constant "tg_3_16" > 0.5, thus _tg_3_16 is encoded as tg_3_16-1.0
      // optr[5*8] = X86_PMULHW( tm765 * ( tg_3_16 - 1.0 ) ) + tm465
      //           = [tm765*tg_3_16 - tm765] + tm465
      //
      // optr[5*8] + tm765 = [ tm765*tg_3_16 ] + tm465 + tm765
      //                   = [ tm765*tg_3_16 ] + tm465 <-- what we want

      optr[5*8] = X86_PMULHW( tm765 * _tg_3_16 ) + tm465 + tm765;

//    optr[3*8] = tm765 - X86_PMULHW( tm465 * _tg_3_16 ); // *** DISCREPANCY
      // The following operations must be performed in the shown order!
      // same trick (as shown for optr[5*8]) applies to optr[3*8]

      optr[3*8] = X86_PMULHW( tm465 * _tg_3_16 ) + tm465;
      optr[3*8] = tm765 - optr[3*8];

    ++sptr;   // increment source pointer +1 column
    ++optr;   // increment output pointer +1 column
  } // end for ( j = 0 ..., end of C_equivalent code for forward_dct_col_1
 
    ////////////////////////////////////////////////////////////////////////
    //
    // DCT8_FRW_ROW1(), equivalent c_code
    //
    // This C-code can be substituted for the same __asm block
    // For a derivation of this code, please read fdctmm32.doc
    ////////////////////////////////////////////////////////////////////////

    sptr = (short *) blk;
    optr = (short *) blk; // output will overwrite source data!
    tf = &tab_frw_01234567[ 0 ]; // fdct_row load table_forward_w

    for ( j = 0; j < 8; j=j+1 ) // dct_frw_row1 loop
    {
        // forward_dct_row input arithmetic + shuffle
        xt[3] = sptr[3] + sptr[4];
        xt[2] = sptr[2] + sptr[5];
        xt[1] = sptr[1] + sptr[6];
        xt[0] = sptr[0] + sptr[7];

        xt[7] = sptr[3] - sptr[4];
        xt[6] = sptr[2] - sptr[5];
        xt[5] = sptr[1] - sptr[6];
        xt[4] = sptr[0] - sptr[7];
  

        a3 = ( xt[0]*tf[10]+ xt[2]*tf[11]) + ( xt[1]*tf[14]+ xt[3]*tf[15]);
        a2 = ( xt[0]*tf[8] + xt[2]*tf[9] ) + ( xt[1]*tf[12]+ xt[3]*tf[13]);
        a1 = ( xt[0]*tf[2] + xt[2]*tf[3] ) + ( xt[1]*tf[6] + xt[3]*tf[7] );
        a0 = ( xt[0]*tf[0] + xt[2]*tf[1] ) + ( xt[1]*tf[4] + xt[3]*tf[5] );
        tf += 16;  // increment table pointer

        b3 = ( xt[4]*tf[10]+ xt[6]*tf[11]) + ( xt[5]*tf[14]+ xt[7]*tf[15]);
        b2 = ( xt[4]*tf[8] + xt[6]*tf[9] ) + ( xt[5]*tf[12]+ xt[7]*tf[13]);
        b1 = ( xt[4]*tf[2] + xt[6]*tf[3] ) + ( xt[5]*tf[6] + xt[7]*tf[7] );
        b0 = ( xt[4]*tf[0] + xt[6]*tf[1] ) + ( xt[5]*tf[4] + xt[7]*tf[5] );
        tf += 16;  // increment table pointer

        // apply rounding constants to scaled elements
        // note, in the MMX implementation, the shift&round is done *last.*
        // Here, the C-code applies the shifts 1st, then the clipping.
#define SHIFT_AND_ROUND_FRW_ROW( x )  ( ((x)+RND_FRW_ROW) >> SHIFT_FRW_ROW )

        a3 = SHIFT_AND_ROUND_FRW_ROW( a3 );
        a2 = SHIFT_AND_ROUND_FRW_ROW( a2 );
        a1 = SHIFT_AND_ROUND_FRW_ROW( a1 );
        a0 = SHIFT_AND_ROUND_FRW_ROW( a0 );

        b3 = SHIFT_AND_ROUND_FRW_ROW( b3 );
        b2 = SHIFT_AND_ROUND_FRW_ROW( b2 );
        b1 = SHIFT_AND_ROUND_FRW_ROW( b1 );
        b0 = SHIFT_AND_ROUND_FRW_ROW( b0 );

        // v1.01, clip output results to range {-2048, +2047}

        // In the MMX implementation, the "clipper" is integrated into
        // the shift&round operation (thanks to packssdw)
        a3 = (a3 > 2047) ?  2047 : a3; // ceiling @ +2047
        a2 = (a2 > 2047) ?  2047 : a2; // ceiling @ +2047
        a1 = (a1 > 2047) ?  2047 : a1; // ceiling @ +2047
        a0 = (a0 > 2047) ?  2047 : a0; // ceiling @ +2047
        b3 = (b3 > 2047) ?  2047 : b3; // ceiling @ +2047
        b2 = (b2 > 2047) ?  2047 : b2; // ceiling @ +2047
        b1 = (b1 > 2047) ?  2047 : b1; // ceiling @ +2047
        b0 = (b0 > 2047) ?  2047 : b0; // ceiling @ +2047

        a3 = (a3 <-2048) ? -2048 : a3; // floor   @ -2048
        a2 = (a2 <-2048) ? -2048 : a2; // floor   @ -2048
        a1 = (a1 <-2048) ? -2048 : a1; // floor   @ -2048
        a0 = (a0 <-2048) ? -2048 : a0; // floor   @ -2048
        b3 = (b3 <-2048) ? -2048 : b3; // floor   @ -2048
        b2 = (b2 <-2048) ? -2048 : b2; // floor   @ -2048
        b1 = (b1 <-2048) ? -2048 : b1; // floor   @ -2048
        b0 = (b0 <-2048) ? -2048 : b0; // floor   @ -2048


        // forward_dct_row, assign outputs
        optr[ 3 ] = b1;
        optr[ 2 ] = a1;
        optr[ 1 ] = b0;
        optr[ 0 ] = a0;

        optr[ 7 ] = b3;
        optr[ 6 ] = a3;
        optr[ 5 ] = b2;
        optr[ 4 ] = a2;

        sptr += 8;   // increment source pointer +1 row
        optr += 8;   // increment output pointer +1 row
    } // end for ( j = 0 ..., end of C_equivalent code for forward_dct_row_1
  */  
} // fdct_mm32( short *blk )