falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
3464 Commits
32 Branches
334MB
Tree: d4c5d2adf3
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'd4c5d2adf3'
${ noResults }
FFmpeg/libavcodec/i386/vp3dsp_sse2.c

835 lines
36KB
Raw Blame History 

  1. /*

  2.  * Copyright (C) 2004 the ffmpeg project

  3.  *

  4.  * This library is free software; you can redistribute it and/or

  5.  * modify it under the terms of the GNU Lesser General Public

  6.  * License as published by the Free Software Foundation; either

  7.  * version 2 of the License, or (at your option) any later version.

  8.  *

  9.  * This library is distributed in the hope that it will be useful,

  10.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  11.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  12.  * Lesser General Public License for more details.

  13.  *

  14.  * You should have received a copy of the GNU Lesser General Public

  15.  * License along with this library; if not, write to the Free Software

  16.  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

  17.  */

  18. 

  19. /**

  20.  * @file vp3dsp_sse2.c

  21.  * SSE2-optimized functions cribbed from the original VP3 source code.

  22.  */

  23. 

  24. #include "../dsputil.h"

  25. #include "mmx.h"

  26. 

  27. static unsigned short __align16 SSE2_dequant_const[] =

  28. {

  29.     0,65535,65535,0,0,0,0,0,    // 0x0000 0000 0000 0000 0000 FFFF FFFF 0000

  30.     0,0,0,0,65535,65535,0,0,    // 0x0000 0000 FFFF FFFF 0000 0000 0000 0000

  31.     65535,65535,65535,0,0,0,0,0,// 0x0000 0000 0000 0000 0000 FFFF FFFF FFFF

  32.     0,0,0,65535,0,0,0,0,        // 0x0000 0000 0000 0000 FFFF 0000 0000 0000

  33.     0,0,0,65535,65535,0,0,0,    // 0x0000 0000 0000 FFFF FFFF 0000 0000 0000

  34.     65535,0,0,0,0,65535,0,0,    // 0x0000 0000 FFFF 0000 0000 0000 0000 FFFF

  35.     0,0,65535,65535, 0,0,0,0    // 0x0000 0000 0000 0000 FFFF FFFF 0000 0000

  36. };

  37. 

  38. static unsigned int __align16 eight_data[] =

  39. { 

  40.     0x00080008, 

  41.     0x00080008,

  42.     0x00080008, 

  43.     0x00080008 

  44. }; 

  45. 

  46. static unsigned short __align16 SSE2_idct_data[7 * 8] =

  47. {

  48.     64277,64277,64277,64277,64277,64277,64277,64277, 

  49.     60547,60547,60547,60547,60547,60547,60547,60547, 

  50.     54491,54491,54491,54491,54491,54491,54491,54491, 

  51.     46341,46341,46341,46341,46341,46341,46341,46341, 

  52.     36410,36410,36410,36410,36410,36410,36410,36410, 

  53.     25080,25080,25080,25080,25080,25080,25080,25080, 

  54.     12785,12785,12785,12785,12785,12785,12785,12785

  55. };

  56. 

  57. 

  58. #define SSE2_Column_IDCT() {        \

  59.     \

  60.     movdqu_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \

  61.     movdqu_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \

  62.     \

  63.     movdqu_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \

  64.     movdqu_m2r(*I(5), xmm7);     /* xmm7 = i5 */ \

  65.     \

  66.     pmulhw_r2r(xmm6, xmm4);      /* xmm4 = c3 * i3 - i3 */ \

  67.     movdqu_m2r(*C(5), xmm1);     /* xmm1 = c5 */ \

  68.     \

  69.     pmulhw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 - i5 */ \

  70.     movdqu_r2r(xmm1, xmm5);      /* xmm5 = c5 */ \

  71.     \

  72.     pmulhw_r2r(xmm2, xmm1);      /* xmm1 = c5 * i3 - i3 */ \

  73.     movdqu_m2r(*I(1), xmm3);     /* xmm3 = i1 */ \

  74.     \

  75.     pmulhw_r2r(xmm7, xmm5);      /* xmm5 = c5 * i5 - i5 */ \

  76.     movdqu_m2r(*C(1), xmm0);     /* xmm0 = c1 */ \

  77.     \

  78.     /* all registers are in use */ \

  79.     \

  80.     paddw_r2r(xmm2, xmm4);       /* xmm4 = c3 * i3 */ \

  81.     paddw_r2r(xmm7, xmm6);       /* xmm6 = c3 * i5 */ \

  82.     \

  83.     paddw_r2r(xmm1, xmm2);       /* xmm2 = c5 * i3 */ \

  84.     movdqu_m2r(*I(7), xmm1);     /* xmm1 = i7 */ \

  85.     \

  86.     paddw_r2r(xmm5, xmm7);       /* xmm7 = c5 * i5 */ \

  87.     movdqu_r2r(xmm0, xmm5);      /* xmm5 = c1 */ \

  88.     \

  89.     pmulhw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 - i1 */ \

  90.     paddsw_r2r(xmm7, xmm4);      /* xmm4 = c3 * i3 + c5 * i5 = C */ \

  91.     \

  92.     pmulhw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 - i7 */ \

  93.     movdqu_m2r(*C(7), xmm7);     /* xmm7 = c7 */ \

  94.     \

  95.     psubsw_r2r(xmm2, xmm6);      /* xmm6 = c3 * i5 - c5 * i3 = D */ \

  96.     paddw_r2r(xmm3, xmm0);       /* xmm0 = c1 * i1 */ \

  97.     \

  98.     pmulhw_r2r(xmm7, xmm3);      /* xmm3 = c7 * i1 */ \

  99.     movdqu_m2r(*I(2), xmm2);     /* xmm2 = i2 */ \

  100.     \

  101.     pmulhw_r2r(xmm1, xmm7);      /* xmm7 = c7 * i7 */ \

  102.     paddw_r2r(xmm1, xmm5);       /* xmm5 = c1 * i7 */ \

  103.     \

  104.     movdqu_r2r(xmm2, xmm1);      /* xmm1 = i2 */ \

  105.     pmulhw_m2r(*C(2), xmm2);     /* xmm2 = i2 * c2 -i2 */ \

  106.     \

  107.     psubsw_r2r(xmm5, xmm3);      /* xmm3 = c7 * i1 - c1 * i7 = B */ \

  108.     movdqu_m2r(*I(6), xmm5);     /* xmm5 = i6 */ \

  109.     \

  110.     paddsw_r2r(xmm7, xmm0);      /* xmm0 = c1 * i1 + c7 * i7 = A */ \

  111.     movdqu_r2r(xmm5, xmm7);      /* xmm7 = i6 */ \

  112.     \

  113.     psubsw_r2r(xmm4, xmm0);      /* xmm0 = A - C */ \

  114.     pmulhw_m2r(*C(2), xmm5);     /* xmm5 = c2 * i6 - i6 */ \

  115.     \

  116.     paddw_r2r(xmm1, xmm2);       /* xmm2 = i2 * c2 */ \

  117.     pmulhw_m2r(*C(6), xmm1);     /* xmm1 = c6 * i2 */ \

  118.     \

  119.     paddsw_r2r(xmm4, xmm4);      /* xmm4 = C + C */ \

  120.     paddsw_r2r(xmm0, xmm4);      /* xmm4 = A + C = C. */ \

  121.     \

  122.     psubsw_r2r(xmm6, xmm3);      /* xmm3 = B - D */ \

  123.     paddw_r2r(xmm7, xmm5);       /* xmm5 = c2 * i6 */ \

  124.     \

  125.     paddsw_r2r(xmm6, xmm6);      /* xmm6 = D + D */ \

  126.     pmulhw_m2r(*C(6), xmm7);     /* xmm7 = c6 * i6 */ \

  127.     \

  128.     paddsw_r2r(xmm3, xmm6);      /* xmm6 = B + D = D. */ \

  129.     movdqu_r2m(xmm4, *I(1));     /* Save C. at I(1) */ \

  130.     \

  131.     psubsw_r2r(xmm5, xmm1);      /* xmm1 = c6 * i2 - c2 * i6 = H */ \

  132.     movdqu_m2r(*C(4), xmm4);     /* xmm4 = c4 */ \

  133.     \

  134.     movdqu_r2r(xmm3, xmm5);      /* xmm5 = B - D */ \

  135.     pmulhw_r2r(xmm4, xmm3);      /* xmm3 = ( c4 -1 ) * ( B - D ) */ \

  136.     \

  137.     paddsw_r2r(xmm2, xmm7);      /* xmm7 = c2 * i2 + c6 * i6 = G */ \

  138.     movdqu_r2m(xmm6, *I(2));     /* Save D. at I(2) */ \

  139.     \

  140.     movdqu_r2r(xmm0, xmm2);      /* xmm2 = A - C */ \

  141.     movdqu_m2r(*I(0), xmm6);     /* xmm6 = i0 */ \

  142.     \

  143.     pmulhw_r2r(xmm4, xmm0);      /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \

  144.     paddw_r2r(xmm3, xmm5);       /* xmm5 = c4 * ( B - D ) = B. */ \

  145.     \

  146.     movdqu_m2r(*I(4), xmm3);     /* xmm3 = i4 */ \

  147.     psubsw_r2r(xmm1, xmm5);      /* xmm5 = B. - H = B.. */ \

  148.     \

  149.     paddw_r2r(xmm0, xmm2);       /* xmm2 = c4 * ( A - C) = A. */ \

  150.     psubsw_r2r(xmm3, xmm6);      /* xmm6 = i0 - i4 */ \

  151.     \

  152.     movdqu_r2r(xmm6, xmm0);      /* xmm0 = i0 - i4 */ \

  153.     pmulhw_r2r(xmm4, xmm6);      /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \

  154.     \

  155.     paddsw_r2r(xmm3, xmm3);      /* xmm3 = i4 + i4 */ \

  156.     paddsw_r2r(xmm1, xmm1);      /* xmm1 = H + H */ \

  157.     \

  158.     paddsw_r2r(xmm0, xmm3);      /* xmm3 = i0 + i4 */ \

  159.     paddsw_r2r(xmm5, xmm1);      /* xmm1 = B. + H = H. */ \

  160.     \

  161.     pmulhw_r2r(xmm3, xmm4);      /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \

  162.     paddw_r2r(xmm0, xmm6);       /* xmm6 = c4 * ( i0 - i4 ) */ \

  163.     \

  164.     psubsw_r2r(xmm2, xmm6);      /* xmm6 = F - A. = F. */ \

  165.     paddsw_r2r(xmm2, xmm2);      /* xmm2 = A. + A. */ \

  166.     \

  167.     movdqu_m2r(*I(1), xmm0);     /* Load        C. from I(1) */ \

  168.     paddsw_r2r(xmm6, xmm2);      /* xmm2 = F + A. = A.. */ \

  169.     \

  170.     paddw_r2r(xmm3, xmm4);       /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \

  171.     psubsw_r2r(xmm1, xmm2);      /* xmm2 = A.. - H. = R2 */ \

  172.     \

  173.     paddsw_m2r(*Eight, xmm2);    /* Adjust R2 and R1 before shifting */ \

  174.     paddsw_r2r(xmm1, xmm1);      /* xmm1 = H. + H. */ \

  175.     \

  176.     paddsw_r2r(xmm2, xmm1);      /* xmm1 = A.. + H. = R1 */ \

  177.     psraw_i2r(4, xmm2);          /* xmm2 = op2 */ \

  178.     \

  179.     psubsw_r2r(xmm7, xmm4);      /* xmm4 = E - G = E. */ \

  180.     psraw_i2r(4, xmm1);          /* xmm1 = op1 */ \

  181.     \

  182.     movdqu_m2r(*I(2), xmm3);     /* Load D. from I(2) */ \

  183.     paddsw_r2r(xmm7, xmm7);      /* xmm7 = G + G */ \

  184.     \

  185.     movdqu_r2m(xmm2, *O(2));     /* Write out op2 */ \

  186.     paddsw_r2r(xmm4, xmm7);      /* xmm7 = E + G = G. */ \

  187.     \

  188.     movdqu_r2m(xmm1, *O(1));     /* Write out op1 */ \

  189.     psubsw_r2r(xmm3, xmm4);      /* xmm4 = E. - D. = R4 */ \

  190.     \

  191.     paddsw_m2r(*Eight, xmm4);    /* Adjust R4 and R3 before shifting */ \

  192.     paddsw_r2r(xmm3, xmm3);      /* xmm3 = D. + D. */ \

  193.     \

  194.     paddsw_r2r(xmm4, xmm3);      /* xmm3 = E. + D. = R3 */ \

  195.     psraw_i2r(4, xmm4);          /* xmm4 = op4 */ \

  196.     \

  197.     psubsw_r2r(xmm5, xmm6);      /* xmm6 = F. - B..= R6 */ \

  198.     psraw_i2r(4, xmm3);          /* xmm3 = op3 */ \

  199.     \

  200.     paddsw_m2r(*Eight, xmm6);    /* Adjust R6 and R5 before shifting */ \

  201.     paddsw_r2r(xmm5, xmm5);      /* xmm5 = B.. + B.. */ \

  202.     \

  203.     paddsw_r2r(xmm6, xmm5);      /* xmm5 = F. + B.. = R5 */ \

  204.     psraw_i2r(4, xmm6);          /* xmm6 = op6 */ \

  205.     \

  206.     movdqu_r2m(xmm4, *O(4));     /* Write out op4 */ \

  207.     psraw_i2r(4, xmm5);          /* xmm5 = op5 */ \

  208.     \

  209.     movdqu_r2m(xmm3, *O(3));     /* Write out op3 */ \

  210.     psubsw_r2r(xmm0, xmm7);      /* xmm7 = G. - C. = R7 */ \

  211.     \

  212.     paddsw_m2r(*Eight, xmm7);    /* Adjust R7 and R0 before shifting */ \

  213.     paddsw_r2r(xmm0, xmm0);      /* xmm0 = C. + C. */ \

  214.     \

  215.     paddsw_r2r(xmm7, xmm0);      /* xmm0 = G. + C. */ \

  216.     psraw_i2r(4, xmm7);          /* xmm7 = op7 */ \

  217.     \

  218.     movdqu_r2m(xmm6, *O(6));     /* Write out op6 */ \

  219.     psraw_i2r(4, xmm0);          /* xmm0 = op0 */ \

  220.     \

  221.     movdqu_r2m(xmm5, *O(5));     /* Write out op5 */ \

  222.     movdqu_r2m(xmm7, *O(7));     /* Write out op7 */ \

  223.     \

  224.     movdqu_r2m(xmm0, *O(0));     /* Write out op0 */ \

  225.     \

  226. } /* End of SSE2_Column_IDCT macro */

  227. 

  228. 

  229. #define SSE2_Row_IDCT() {        \

  230.     \

  231.     movdqu_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \

  232.     movdqu_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \

  233.     \

  234.     movdqu_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \

  235.     movdqu_m2r(*I(5), xmm7);     /* xmm7 = i5 */ \

  236.     \

  237.     pmulhw_r2r(xmm6, xmm4);      /* xmm4 = c3 * i3 - i3 */ \

  238.     movdqu_m2r(*C(5), xmm1);     /* xmm1 = c5 */ \

  239.     \

  240.     pmulhw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 - i5 */ \

  241.     movdqu_r2r(xmm1, xmm5);      /* xmm5 = c5 */ \

  242.     \

  243.     pmulhw_r2r(xmm2, xmm1);      /* xmm1 = c5 * i3 - i3 */ \

  244.     movdqu_m2r(*I(1), xmm3);     /* xmm3 = i1 */ \

  245.     \

  246.     pmulhw_r2r(xmm7, xmm5);      /* xmm5 = c5 * i5 - i5 */ \

  247.     movdqu_m2r(*C(1), xmm0);     /* xmm0 = c1 */ \

  248.     \

  249.     /* all registers are in use */ \

  250.     \

  251.     paddw_r2r(xmm2, xmm4);      /* xmm4 = c3 * i3 */ \

  252.     paddw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 */ \

  253.     \

  254.     paddw_r2r(xmm1, xmm2);      /* xmm2 = c5 * i3 */ \

  255.     movdqu_m2r(*I(7), xmm1);    /* xmm1 = i7 */ \

  256.     \

  257.     paddw_r2r(xmm5, xmm7);      /* xmm7 = c5 * i5 */ \

  258.     movdqu_r2r(xmm0, xmm5);     /* xmm5 = c1 */ \

  259.     \

  260.     pmulhw_r2r(xmm3, xmm0);     /* xmm0 = c1 * i1 - i1 */ \

  261.     paddsw_r2r(xmm7, xmm4);     /* xmm4 = c3 * i3 + c5 * i5 = C */ \

  262.     \

  263.     pmulhw_r2r(xmm1, xmm5);     /* xmm5 = c1 * i7 - i7 */ \

  264.     movdqu_m2r(*C(7), xmm7);    /* xmm7 = c7 */ \

  265.     \

  266.     psubsw_r2r(xmm2, xmm6);     /* xmm6 = c3 * i5 - c5 * i3 = D */ \

  267.     paddw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 */ \

  268.     \

  269.     pmulhw_r2r(xmm7, xmm3);     /* xmm3 = c7 * i1 */ \

  270.     movdqu_m2r(*I(2), xmm2);    /* xmm2 = i2 */ \

  271.     \

  272.     pmulhw_r2r(xmm1, xmm7);     /* xmm7 = c7 * i7 */ \

  273.     paddw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 */ \

  274.     \

  275.     movdqu_r2r(xmm2, xmm1);     /* xmm1 = i2 */ \

  276.     pmulhw_m2r(*C(2), xmm2);    /* xmm2 = i2 * c2 -i2 */ \

  277.     \

  278.     psubsw_r2r(xmm5, xmm3);     /* xmm3 = c7 * i1 - c1 * i7 = B */ \

  279.     movdqu_m2r(*I(6), xmm5);    /* xmm5 = i6 */ \

  280.     \

  281.     paddsw_r2r(xmm7, xmm0);     /* xmm0 = c1 * i1 + c7 * i7        = A */ \

  282.     movdqu_r2r(xmm5, xmm7);     /* xmm7 = i6 */ \

  283.     \

  284.     psubsw_r2r(xmm4, xmm0);     /* xmm0 = A - C */ \

  285.     pmulhw_m2r(*C(2), xmm5);    /* xmm5 = c2 * i6 - i6 */ \

  286.     \

  287.     paddw_r2r(xmm1, xmm2);      /* xmm2 = i2 * c2 */ \

  288.     pmulhw_m2r(*C(6), xmm1);    /* xmm1 = c6 * i2 */ \

  289.     \

  290.     paddsw_r2r(xmm4, xmm4);     /* xmm4 = C + C */ \

  291.     paddsw_r2r(xmm0, xmm4);     /* xmm4 = A + C = C. */ \

  292.     \

  293.     psubsw_r2r(xmm6, xmm3);     /* xmm3 = B - D */ \

  294.     paddw_r2r(xmm7, xmm5);      /* xmm5 = c2 * i6 */ \

  295.     \

  296.     paddsw_r2r(xmm6, xmm6);     /* xmm6 = D + D */ \

  297.     pmulhw_m2r(*C(6), xmm7);    /* xmm7 = c6 * i6 */ \

  298.     \

  299.     paddsw_r2r(xmm3, xmm6);     /* xmm6 = B + D = D. */ \

  300.     movdqu_r2m(xmm4, *I(1));    /* Save C. at I(1)        */ \

  301.     \

  302.     psubsw_r2r(xmm5, xmm1);     /* xmm1 = c6 * i2 - c2 * i6 = H */ \

  303.     movdqu_m2r(*C(4), xmm4);    /* xmm4 = c4 */ \

  304.     \

  305.     movdqu_r2r(xmm3, xmm5);     /* xmm5 = B - D */ \

  306.     pmulhw_r2r(xmm4, xmm3);     /* xmm3 = ( c4 -1 ) * ( B - D ) */ \

  307.     \

  308.     paddsw_r2r(xmm2, xmm7);     /* xmm7 = c2 * i2 + c6 * i6 = G */ \

  309.     movdqu_r2m(xmm6, *I(2));    /* Save D. at I(2) */ \

  310.     \

  311.     movdqu_r2r(xmm0, xmm2);     /* xmm2 = A - C */ \

  312.     movdqu_m2r(*I(0), xmm6);    /* xmm6 = i0 */ \

  313.     \

  314.     pmulhw_r2r(xmm4, xmm0);     /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \

  315.     paddw_r2r(xmm3, xmm5);      /* xmm5 = c4 * ( B - D ) = B. */ \

  316.     \

  317.     movdqu_m2r(*I(4), xmm3);    /* xmm3 = i4 */ \

  318.     psubsw_r2r(xmm1, xmm5);     /* xmm5 = B. - H = B.. */ \

  319.     \

  320.     paddw_r2r(xmm0, xmm2);      /* xmm2 = c4 * ( A - C) = A. */ \

  321.     psubsw_r2r(xmm3, xmm6);     /* xmm6 = i0 - i4 */ \

  322.     \

  323.     movdqu_r2r(xmm6, xmm0);     /* xmm0 = i0 - i4 */ \

  324.     pmulhw_r2r(xmm4, xmm6);     /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \

  325.     \

  326.     paddsw_r2r(xmm3, xmm3);     /* xmm3 = i4 + i4 */ \

  327.     paddsw_r2r(xmm1, xmm1);     /* xmm1 = H + H */ \

  328.     \

  329.     paddsw_r2r(xmm0, xmm3);     /* xmm3 = i0 + i4 */ \

  330.     paddsw_r2r(xmm5, xmm1);     /* xmm1 = B. + H = H. */ \

  331.     \

  332.     pmulhw_r2r(xmm3, xmm4);     /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \

  333.     paddw_r2r(xmm0, xmm6);      /* xmm6 = c4 * ( i0 - i4 ) */ \

  334.     \

  335.     psubsw_r2r(xmm2, xmm6);     /* xmm6 = F - A. = F. */ \

  336.     paddsw_r2r(xmm2, xmm2);     /* xmm2 = A. + A. */ \

  337.     \

  338.     movdqu_m2r(*I(1), xmm0);    /* Load C. from I(1) */ \

  339.     paddsw_r2r(xmm6, xmm2);     /* xmm2 = F + A. = A.. */ \

  340.     \

  341.     paddw_r2r(xmm3, xmm4);      /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \

  342.     psubsw_r2r(xmm1, xmm2);     /* xmm2 = A.. - H. = R2 */ \

  343.     \

  344.     paddsw_r2r(xmm1, xmm1);     /* xmm1 = H. + H. */ \

  345.     paddsw_r2r(xmm2, xmm1);     /* xmm1 = A.. + H. = R1 */ \

  346.     \

  347.     psubsw_r2r(xmm7, xmm4);     /* xmm4 = E - G = E. */ \

  348.     \

  349.     movdqu_m2r(*I(2), xmm3);    /* Load D. from I(2) */ \

  350.     paddsw_r2r(xmm7, xmm7);     /* xmm7 = G + G */ \

  351.     \

  352.     movdqu_r2m(xmm2, *I(2));    /* Write out op2 */ \

  353.     paddsw_r2r(xmm4, xmm7);     /* xmm7 = E + G = G. */ \

  354.     \

  355.     movdqu_r2m(xmm1, *I(1));    /* Write out op1 */ \

  356.     psubsw_r2r(xmm3, xmm4);     /* xmm4 = E. - D. = R4 */ \

  357.     \

  358.     paddsw_r2r(xmm3, xmm3);     /* xmm3 = D. + D. */ \

  359.     \

  360.     paddsw_r2r(xmm4, xmm3);     /* xmm3 = E. + D. = R3 */ \

  361.     \

  362.     psubsw_r2r(xmm5, xmm6);     /* xmm6 = F. - B..= R6 */ \

  363.     \

  364.     paddsw_r2r(xmm5, xmm5);     /* xmm5 = B.. + B.. */ \

  365.     \

  366.     paddsw_r2r(xmm6, xmm5);     /* xmm5 = F. + B.. = R5 */ \

  367.     \

  368.     movdqu_r2m(xmm4, *I(4));    /* Write out op4 */ \

  369.     \

  370.     movdqu_r2m(xmm3, *I(3));    /* Write out op3 */ \

  371.     psubsw_r2r(xmm0, xmm7);     /* xmm7 = G. - C. = R7 */ \

  372.     \

  373.     paddsw_r2r(xmm0, xmm0);     /* xmm0 = C. + C. */ \

  374.     \

  375.     paddsw_r2r(xmm7, xmm0);     /* xmm0 = G. + C. */ \

  376.     \

  377.     movdqu_r2m(xmm6, *I(6));    /* Write out op6 */ \

  378.     \

  379.     movdqu_r2m(xmm5, *I(5));    /* Write out op5 */ \

  380.     movdqu_r2m(xmm7, *I(7));    /* Write out op7 */ \

  381.     \

  382.     movdqu_r2m(xmm0, *I(0));    /* Write out op0 */ \

  383.     \

  384. } /* End of SSE2_Row_IDCT macro */

  385. 

  386. 

  387. #define SSE2_Transpose() {    \

  388.     \

  389.     movdqu_m2r(*I(4), xmm4);    /* xmm4=e7e6e5e4e3e2e1e0 */ \

  390.     movdqu_m2r(*I(5), xmm0);    /* xmm4=f7f6f5f4f3f2f1f0 */ \

  391.     \

  392.     movdqu_r2r(xmm4, xmm5);     /* make a copy */ \

  393.     punpcklwd_r2r(xmm0, xmm4);  /* xmm4=f3e3f2e2f1e1f0e0 */ \

  394.     \

  395.     punpckhwd_r2r(xmm0, xmm5);  /* xmm5=f7e7f6e6f5e5f4e4 */ \

  396.     movdqu_m2r(*I(6), xmm6);    /* xmm6=g7g6g5g4g3g2g1g0 */ \

  397.     \

  398.     movdqu_m2r(*I(7), xmm0);    /* xmm0=h7h6h5h4h3h2h1h0 */ \

  399.     movdqu_r2r(xmm6, xmm7);     /* make a copy */ \

  400.     \

  401.     punpcklwd_r2r(xmm0, xmm6);  /* xmm6=h3g3h3g2h1g1h0g0 */ \

  402.     punpckhwd_r2r(xmm0, xmm7);  /* xmm7=h7g7h6g6h5g5h4g4 */ \

  403.     \

  404.     movdqu_r2r(xmm4, xmm3);     /* make a copy */ \

  405.     punpckldq_r2r(xmm6, xmm4);  /* xmm4=h1g1f1e1h0g0f0e0 */ \

  406.     \

  407.     punpckhdq_r2r(xmm6, xmm3);  /* xmm3=h3g3g3e3h2g2f2e2 */ \

  408.     movdqu_r2m(xmm3, *I(6));    /* save h3g3g3e3h2g2f2e2 */ \

  409.     /* Free xmm6 */ \

  410.     movdqu_r2r(xmm5, xmm6);     /* make a copy */ \

  411.     punpckldq_r2r(xmm7, xmm5);  /* xmm5=h5g5f5e5h4g4f4e4 */ \

  412.     \

  413.     punpckhdq_r2r(xmm7, xmm6);  /* xmm6=h7g7f7e7h6g6f6e6 */ \

  414.     movdqu_m2r(*I(0), xmm0);    /* xmm0=a7a6a5a4a3a2a1a0 */ \

  415.     /* Free xmm7 */ \

  416.     movdqu_m2r(*I(1), xmm1);    /* xmm1=b7b6b5b4b3b2b1b0 */ \

  417.     movdqu_r2r(xmm0, xmm7);     /* make a copy */ \

  418.     \

  419.     punpcklwd_r2r(xmm1, xmm0);  /* xmm0=b3a3b2a2b1a1b0a0 */ \

  420.     punpckhwd_r2r(xmm1, xmm7);  /* xmm7=b7a7b6a6b5a5b4a4 */ \

  421.     /* Free xmm1 */ \

  422.     movdqu_m2r(*I(2), xmm2);    /* xmm2=c7c6c5c4c3c2c1c0 */ \

  423.     movdqu_m2r(*I(3), xmm3);    /* xmm3=d7d6d5d4d3d2d1d0 */ \

  424.     \

  425.     movdqu_r2r(xmm2, xmm1);     /* make a copy */ \

  426.     punpcklwd_r2r(xmm3, xmm2);  /* xmm2=d3c3d2c2d1c1d0c0 */ \

  427.     \

  428.     punpckhwd_r2r(xmm3, xmm1);  /* xmm1=d7c7d6c6d5c5d4c4 */ \

  429.     movdqu_r2r(xmm0, xmm3);     /* make a copy        */ \

  430.     \

  431.     punpckldq_r2r(xmm2, xmm0);  /* xmm0=d1c1b1a1d0c0b0a0 */ \

  432.     punpckhdq_r2r(xmm2, xmm3);  /* xmm3=d3c3b3a3d2c2b2a2 */ \

  433.     /* Free xmm2 */ \

  434.     movdqu_r2r(xmm7, xmm2);     /* make a copy */ \

  435.     punpckldq_r2r(xmm1, xmm2);  /* xmm2=d5c5b5a5d4c4b4a4 */ \

  436.     \

  437.     punpckhdq_r2r(xmm1, xmm7);  /* xmm7=d7c7b7a7d6c6b6a6 */ \

  438.     movdqu_r2r(xmm0, xmm1);     /* make a copy */ \

  439.     \

  440.     punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \

  441.     punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \

  442.     \

  443.     movdqu_r2m(xmm0, *I(0));    /* save I(0) */ \

  444.     movdqu_r2m(xmm1, *I(1));    /* save I(1) */ \

  445.     \

  446.     movdqu_m2r(*I(6), xmm0);    /* load h3g3g3e3h2g2f2e2 */ \

  447.     movdqu_r2r(xmm3, xmm1);     /* make a copy */ \

  448.     \

  449.     punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \

  450.     punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \

  451.     \

  452.     movdqu_r2r(xmm2, xmm4);     /* make a copy */ \

  453.     punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \

  454.     \

  455.     punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \

  456.     movdqu_r2m(xmm1, *I(2));    /* save I(2) */ \

  457.     \

  458.     movdqu_r2m(xmm3, *I(3));    /* save I(3) */ \

  459.     movdqu_r2m(xmm4, *I(4));    /* save I(4) */ \

  460.     \

  461.     movdqu_r2m(xmm2, *I(5));    /* save I(5) */ \

  462.     movdqu_r2r(xmm7, xmm5);     /* make a copy */ \

  463.     \

  464.     punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \

  465.     punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \

  466.     \

  467.     movdqu_r2m(xmm5, *I(6));    /* save I(6) */ \

  468.     movdqu_r2m(xmm7, *I(7));    /* save I(7) */ \

  469.     \

  470. } /* End of Transpose Macro */

  471. 

  472. 

  473. #define SSE2_Dequantize() {        \

  474.     movdqu_m2r(*(eax), xmm0);     \

  475.     \

  476.     pmullw_m2r(*(ebx), xmm0);          /* xmm0 = 07 06 05 04 03 02 01 00 */ \

  477.     movdqu_m2r(*(eax + 16), xmm1);     \

  478.     \

  479.     pmullw_m2r(*(ebx + 16), xmm1);     /* xmm1 = 17 16 15 14 13 12 11 10 */ \

  480.     pshuflw_r2r(xmm0, xmm3, 0x078);    /* xmm3 = 07 06 05 04 01 03 02 00 */ \

  481.     \

  482.     movdqu_r2r(xmm1, xmm2);            /* xmm2 = 17 16 15 14 13 12 11 10 */ \

  483.     movdqu_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \

  484.     \

  485.     movdqu_m2r(*(eax + 32), xmm4);     \

  486.     movdqu_m2r(*(eax + 64), xmm5);     \

  487.     \

  488.     pmullw_m2r(*(ebx + 32), xmm4);     /* xmm4 = 27 26 25 24 23 22 21 20 */ \

  489.     pmullw_m2r(*(ebx + 64), xmm5);     /* xmm5 = 47 46 45 44 43 42 41 40 */ \

  490.     \

  491.     movdqu_m2r(*(ecx + 16), xmm6);     /* xmm6 = -- -- FF FF -- -- -- -- */ \

  492.     pand_r2r(xmm2, xmm7);              /* xmm7 = -- -- -- -- -- 12 11 -- */ \

  493.     \

  494.     pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- 25 24 -- -- -- -- */ \

  495.     pxor_r2r(xmm7, xmm2);              /* xmm2 = 17 16 15 14 13 -- -- 10 */ \

  496.     \

  497.     pxor_r2r(xmm6, xmm4);              /* xmm4 = 27 26 -- -- 23 22 21 20 */ \

  498.     pslldq_i2r(4, xmm7);               /* xmm7 = -- -- -- 12 11 -- -- -- */ \

  499.     \

  500.     pslldq_i2r(2, xmm6);               /* xmm6 = -- 25 24 -- -- -- -- -- */ \

  501.     por_r2r(xmm6, xmm7);               /* xmm7 = -- 25 24 12 11 -- -- -- */ \

  502.     \

  503.     movdqu_m2r(*(ecx + 32), xmm0);     /* xmm0 = -- -- -- -- -- FF FF FF */ \

  504.     movdqu_m2r(*(ecx + 48), xmm6);     /* xmm6 = -- -- -- -- FF -- -- -- */ \

  505.     \

  506.     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- -- -- 03 02 00 */ \

  507.     pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- -- -- 43 -- -- -- */ \

  508.     \

  509.     pxor_r2r(xmm0, xmm3);              /* xmm3 = 07 06 05 04 01 -- -- -- */ \

  510.     pxor_r2r(xmm6, xmm5);              /* xmm5 = 47 46 45 44 -- 42 41 40 */ \

  511.     \

  512.     por_r2r(xmm7, xmm0);               /* xmm0 = -- 25 24 12 11 03 02 00 */ \

  513.     pslldq_i2r(8, xmm6);               /* xmm6 = 43 -- -- -- -- -- -- -- */ \

  514.     \

  515.     por_r2r(xmm6, xmm0);               /* xmm0 = 43 25 24 12 11 03 02 00 */ \

  516.     /* 02345 in use */ \

  517.     \

  518.     movdqu_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \

  519.     pshuflw_r2r(xmm5, xmm5, 0x0B4);    /* xmm5 = 47 46 45 44 42 -- 41 40 */ \

  520.     \

  521.     movdqu_r2r(xmm1, xmm7);            /* xmm7 = -- -- -- FF FF -- -- -- */ \

  522.     movdqu_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \

  523.     \

  524.     movdqu_r2m(xmm0, *(eax));          /* write  43 25 24 12 11 03 02 00 */ \

  525.     pshufhw_r2r(xmm4, xmm4, 0x0C2);    /* xmm4 = 27 -- -- 26 23 22 21 20 */ \

  526.     \

  527.     pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- 26 23 -- -- -- */ \

  528.     pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 44 42 -- -- -- */ \

  529.     \

  530.     pxor_r2r(xmm7, xmm4);              /* xmm4 = 27 -- -- -- -- 22 21 20 */ \

  531.     pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 45 -- -- -- 41 40 */ \

  532.     \

  533.     pshuflw_r2r(xmm2, xmm2, 0x0C6);    /* xmm2 = 17 16 15 14 13 10 -- -- */ \

  534.     movdqu_r2r(xmm6, xmm0);            /* xmm0 = -- -- -- FF FF -- -- -- */ \

  535.     \

  536.     pslldq_i2r(2, xmm7);               /* xmm7 = -- -- 26 23 -- -- -- -- */ \

  537.     pslldq_i2r(6, xmm1);               /* xmm1 = 44 42 -- -- -- -- -- -- */ \

  538.     \

  539.     psrldq_i2r(2, xmm0);               /* xmm0 = -- -- -- -- FF FF -- -- */ \

  540.     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- 04 01 -- -- -- */ \

  541.     \

  542.     pand_r2r(xmm2, xmm0);              /* xmm0 = -- -- -- -- 13 10 -- -- */ \

  543.     pxor_r2r(xmm6, xmm3);              /* xmm3 = 07 06 05 -- -- -- -- -- */ \

  544.     \

  545.     pxor_r2r(xmm0, xmm2);              /* xmm2 = 17 16 15 14 -- -- -- -- */ \

  546.     psrldq_i2r(6, xmm6);               /* xmm0 = -- -- -- -- -- -- 04 01 */ \

  547.     \

  548.     por_r2r(xmm7, xmm1);               /* xmm1 = 44 42 26 23 -- -- -- -- */ \

  549.     por_r2r(xmm6, xmm0);               /* xmm1 = -- -- -- -- 13 10 04 01 */ \

  550.     /* 12345 in use */    \

  551.     por_r2r(xmm0, xmm1);               /* xmm1 = 44 42 26 23 13 10 04 01 */ \

  552.     pshuflw_r2r(xmm4, xmm4, 0x093);    /* xmm4 = 27 -- -- -- 22 21 20 -- */ \

  553.     \

  554.     pshufhw_r2r(xmm4, xmm4, 0x093);    /* xmm4 = -- -- -- 27 22 21 20 -- */ \

  555.     movdqu_r2m(xmm1, *(eax + 16));     /* write  44 42 26 23 13 10 04 01 */ \

  556.     \

  557.     pshufhw_r2r(xmm3, xmm3, 0x0D2);    /* xmm3 = 07 05 -- 06 -- -- -- -- */ \

  558.     movdqu_m2r(*(ecx + 64), xmm0);     /* xmm0 = -- -- -- FF FF -- -- -- */ \

  559.     \

  560.     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 06 -- -- -- -- */ \

  561.     psrldq_i2r(12, xmm3);              /* xmm3 = -- -- -- -- -- -- 07 05 */ \

  562.     \

  563.     psrldq_i2r(8, xmm0);               /* xmm0 = -- -- -- -- -- -- -- 06 */ \

  564.     \

  565.     movdqu_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \

  566.     movdqu_m2r(*(ecx + 96), xmm7);     /* xmm7 = -- -- -- -- FF FF -- -- */ \

  567.     \

  568.     pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- -- 27 22 -- -- -- */ \

  569.     pxor_r2r(xmm6, xmm4);              /* xmm4 = -- -- -- -- -- 21 20 -- */ \

  570.     \

  571.     por_r2r(xmm6, xmm3);               /* xmm3 = -- -- -- 27 22 -- 07 05 */ \

  572.     pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- -- -- 21 -- -- */ \

  573.     \

  574.     por_r2r(xmm7, xmm0);               /* xmm0 = -- -- -- -- -- 21 -- 06 */ \

  575.     pxor_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- -- 20 -- */ \

  576.     \

  577.     movdqu_m2r(*(ecx + 16 ), xmm6);    /* xmm6 = -- -- FF FF -- -- -- -- */ \

  578.     movdqu_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \

  579.     \

  580.     pand_r2r(xmm2, xmm6);              /* xmm6 = -- -- 15 14 -- -- -- -- */ \

  581.     pand_r2r(xmm6, xmm1);              /* xmm1 = -- -- -- 14 -- -- -- -- */ \

  582.     \

  583.     pxor_r2r(xmm6, xmm2);              /* xmm2 = 17 16 -- -- -- -- -- -- */ \

  584.     pxor_r2r(xmm1, xmm6);              /* xmm6 = -- -- 15 -- -- -- -- -- */ \

  585.     \

  586.     psrldq_i2r(4, xmm1);               /* xmm1 = -- -- -- -- -- 14 -- -- */ \

  587.     \

  588.     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- 15 -- */ \

  589.     por_r2r(xmm1, xmm3);               /* xmm3 = -- -- -- 27 22 14 07 05 */ \

  590.     \

  591.     por_r2r(xmm6, xmm0);               /* xmm0 = -- -- -- -- -- 21 15 06 */ \

  592.     pshufhw_r2r(xmm5, xmm5, 0x0E1);    /* xmm5 = 47 46 -- 45 -- -- 41 40 */ \

  593.     \

  594.     movdqu_m2r(*(ecx + 64), xmm1);     /* xmm1 = -- -- -- FF FF -- -- -- */ \

  595.     pshuflw_r2r(xmm5, xmm5, 0x072);    /* xmm5 = 47 46 -- 45 41 -- 40 -- */ \

  596.     \

  597.     movdqu_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \

  598.     pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 45 41 -- -- -- */ \

  599.     \

  600.     pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 -- -- -- -- 40 -- */ \

  601.     pslldq_i2r(4, xmm1);               /* xmm1 = -- 45 41 -- -- -- -- -- */ \

  602.     \

  603.     pshufd_r2r(xmm5, xmm5, 0x09C);     /* xmm5 = -- -- -- -- 47 46 40 -- */ \

  604.     por_r2r(xmm1, xmm3);               /* xmm3 = -- 45 41 27 22 14 07 05 */ \

  605.     \

  606.     movdqu_m2r(*(eax + 96), xmm1);     /* xmm1 = 67 66 65 64 63 62 61 60 */ \

  607.     pmullw_m2r(*(ebx + 96), xmm1);     \

  608.     \

  609.     movdqu_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \

  610.     \

  611.     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \

  612.     pand_r2r(xmm5, xmm7);              /* xmm7 = -- -- -- -- -- 46 40 -- */ \

  613.     \

  614.     pand_r2r(xmm1, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 60 */ \

  615.     pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- -- -- 47 -- -- -- */ \

  616.     \

  617.     pxor_r2r(xmm6, xmm1);              /* xmm1 = 67 66 65 64 63 62 61 -- */ \

  618.     pslldq_i2r(2, xmm5);               /* xmm5 = -- -- -- 47 -- -- -- -- */ \

  619.     \

  620.     pslldq_i2r(14, xmm6);              /* xmm6 = 60 -- -- -- -- -- -- -- */ \

  621.     por_r2r(xmm5, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 -- */ \

  622.     \

  623.     por_r2r(xmm6, xmm3);               /* xmm3 = 60 45 41 27 22 14 07 05 */ \

  624.     pslldq_i2r(6, xmm7);               /* xmm7 = -- -- 46 40 -- -- -- -- */ \

  625.     \

  626.     movdqu_r2m(xmm3, *(eax+32));       /* write  60 45 41 27 22 14 07 05 */ \

  627.     por_r2r(xmm7, xmm0);               /* xmm0 = -- -- 46 40 -- 21 15 06 */ \

  628.     /* 0, 1, 2, 4 in use */    \

  629.     movdqu_m2r(*(eax + 48), xmm3);     /* xmm3 = 37 36 35 34 33 32 31 30 */ \

  630.     movdqu_m2r(*(eax + 80), xmm5);     /* xmm5 = 57 56 55 54 53 52 51 50 */ \

  631.     \

  632.     pmullw_m2r(*(ebx + 48), xmm3);     \

  633.     pmullw_m2r(*(ebx + 80), xmm5);     \

  634.     \

  635.     movdqu_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \

  636.     movdqu_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \

  637.     \

  638.     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \

  639.     pslldq_i2r(8, xmm7);               /* xmm7 = FF -- -- -- -- -- -- -- */ \

  640.     \

  641.     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 30 */ \

  642.     pand_r2r(xmm5, xmm7);              /* xmm7 = 57 -- -- -- -- -- -- -- */ \

  643.     \

  644.     pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 31 -- */ \

  645.     pxor_r2r(xmm7, xmm5);              /* xmm5 = __ 56 55 54 53 52 51 50 */ \

  646.     \

  647.     pslldq_i2r(6, xmm6);               /* xmm6 = -- -- -- -- 30 -- -- -- */ \

  648.     psrldq_i2r(2, xmm7);               /* xmm7 = -- 57 -- -- -- -- -- -- */ \

  649.     \

  650.     por_r2r(xmm7, xmm6);               /* xmm6 = -- 57 -- -- 30 -- -- -- */ \

  651.     movdqu_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \

  652.     \

  653.     por_r2r(xmm6, xmm0);               /* xmm0 = -- 57 46 40 30 21 15 06 */ \

  654.     psrldq_i2r(2, xmm7);               /* xmm7 = -- -- -- -- -- -- FF FF */ \

  655.     \

  656.     movdqu_r2r(xmm2, xmm6);            /* xmm6 = 17 16 -- -- -- -- -- -- */ \

  657.     pand_r2r(xmm1, xmm7);              /* xmm7 = -- -- -- -- -- -- 61 -- */ \

  658.     \

  659.     pslldq_i2r(2, xmm6);               /* xmm6 = 16 -- -- -- -- -- -- -- */ \

  660.     psrldq_i2r(14, xmm2);              /* xmm2 = -- -- -- -- -- -- -- 17 */ \

  661.     \

  662.     pxor_r2r(xmm7, xmm1);              /* xmm1 = 67 66 65 64 63 62 -- -- */ \

  663.     pslldq_i2r(12, xmm7);              /* xmm7 = 61 -- -- -- -- -- -- -- */ \

  664.     \

  665.     psrldq_i2r(14, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 16 */ \

  666.     por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 16 */ \

  667.     \

  668.     por_r2r(xmm7, xmm0);               /* xmm0 = 61 57 46 40 30 21 15 06 */ \

  669.     movdqu_m2r(*(ecx), xmm6);          /* xmm6 = -- -- -- -- -- FF FF -- */ \

  670.     \

  671.     psrldq_i2r(2, xmm6);               /* xmm6 = -- -- -- -- -- -- FF FF */ \

  672.     movdqu_r2m(xmm0, *(eax+48));       /* write  61 57 46 40 30 21 15 06 */ \

  673.     /* 1, 2, 3, 4, 5 in use */\

  674.     movdqu_m2r(*(ecx), xmm0);          /* xmm0 = -- -- -- -- -- FF FF -- */ \

  675.     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- 31 -- */ \

  676.     \

  677.     movdqu_r2r(xmm3, xmm7);            /* xmm7 = 37 36 35 34 33 32 31 -- */ \

  678.     pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 -- -- */ \

  679.     \

  680.     pslldq_i2r(2, xmm3);               /* xmm3 = 36 35 34 33 32 -- -- -- */ \

  681.     pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- -- 62 -- -- */ \

  682.     \

  683.     psrldq_i2r(14, xmm7);              /* xmm7 = -- -- -- -- -- -- -- 37 */ \

  684.     pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 63 -- -- -- */ \

  685.     \

  686.     por_r2r(xmm7, xmm6);               /* xmm6 = -- -- -- -- -- -- 31 37 */ \

  687.     movdqu_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \

  688.     \

  689.     pshuflw_r2r(xmm6, xmm6, 0x01E);    /* xmm6 = -- -- -- -- 37 31 -- -- */ \

  690.     pslldq_i2r(6, xmm7);               /* xmm7 = FF FF -- -- -- -- -- -- */ \

  691.     \

  692.     por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 37 31 20 16 */ \

  693.     pand_r2r(xmm5, xmm7);              /* xmm7 = -- 56 -- -- -- -- -- -- */ \

  694.     \

  695.     pslldq_i2r(8, xmm0);               /* xmm0 = -- 62 -- -- -- -- -- -- */ \

  696.     pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- 55 54 53 52 51 50 */ \

  697.     \

  698.     psrldq_i2r(2, xmm7);               /* xmm7 = -- -- 56 -- -- -- -- -- */ \

  699.     \

  700.     pshufhw_r2r(xmm3, xmm3, 0x087);    /* xmm3 = 35 33 34 36 32 -- -- -- */ \

  701.     por_r2r(xmm7, xmm0);               /* xmm0 = -- 62 56 -- -- -- -- -- */ \

  702.     \

  703.     movdqu_m2r(*(eax + 112), xmm7);    /* xmm7 = 77 76 75 74 73 72 71 70 */ \

  704.     pmullw_m2r(*(ebx + 112), xmm7);     \

  705.     \

  706.     movdqu_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \

  707.     por_r2r(xmm0, xmm4);               /* xmm4 = -- 62 56 47 37 31 20 16 */ \

  708.     \

  709.     pshuflw_r2r(xmm7, xmm7, 0x0E1);    /* xmm7 = 77 76 75 74 73 72 70 71 */ \

  710.     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \

  711.     \

  712.     movdqu_m2r(*(ecx + 64), xmm0);     /* xmm0 = -- -- -- FF FF -- -- -- */ \

  713.     pand_r2r(xmm7, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 71 */ \

  714.     \

  715.     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 36 32 -- -- -- */ \

  716.     pxor_r2r(xmm6, xmm7);              /* xmm7 = 77 76 75 74 73 72 70 -- */ \

  717.     \

  718.     pxor_r2r(xmm0, xmm3);              /* xmm3 = 35 33 34 -- -- -- -- -- */ \

  719.     pslldq_i2r(14, xmm6);              /* xmm6 = 71 -- -- -- -- -- -- -- */ \

  720.     \

  721.     psrldq_i2r(4, xmm0);               /* xmm0 = -- -- -- -- -- 36 32 -- */ \

  722.     por_r2r(xmm6, xmm4);               /* xmm4 = 71 62 56 47 37 31 20 16 */ \

  723.     \

  724.     por_r2r(xmm0, xmm2);               /* xmm2 = -- -- -- -- -- 36 32 17 */ \

  725.     movdqu_r2m(xmm4, *(eax + 64));     /* write  71 62 56 47 37 31 20 16 */ \

  726.     /* 1, 2, 3, 5, 7 in use */ \

  727.     movdqu_m2r(*(ecx + 80), xmm6);     /* xmm6 = -- -- FF -- -- -- -- FF */ \

  728.     pshufhw_r2r(xmm7, xmm7, 0x0D2);    /* xmm7 = 77 75 74 76 73 72 70 __ */ \

  729.     \

  730.     movdqu_m2r(*(ecx), xmm4);          /* xmm4 = -- -- -- -- -- FF FF -- */ \

  731.     movdqu_m2r(*(ecx+48), xmm0);       /* xmm0 = -- -- -- -- FF -- -- -- */ \

  732.     \

  733.     pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- 55 -- -- -- -- 50 */ \

  734.     pand_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- 72 70 -- */ \

  735.     \

  736.     pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- 63 -- -- -- */ \

  737.     pxor_r2r(xmm6, xmm5);              /* xmm5 = -- -- -- 54 53 52 51 -- */ \

  738.     \

  739.     pxor_r2r(xmm4, xmm7);              /* xmm7 = 77 75 74 76 73 -- -- -- */ \

  740.     pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 -- -- -- -- */ \

  741.     \

  742.     pshuflw_r2r(xmm6, xmm6, 0x02B);    /* xmm6 = -- -- 55 -- 50 -- -- -- */ \

  743.     pslldq_i2r(10, xmm4);              /* xmm4 = 72 20 -- -- -- -- -- -- */ \

  744.     \

  745.     pshufhw_r2r(xmm6, xmm6, 0x0B1);    /* xmm6 = -- -- -- 55 50 -- -- -- */ \

  746.     pslldq_i2r(4, xmm0);               /* xmm0 = -- -- 63 -- -- -- -- -- */ \

  747.     \

  748.     por_r2r(xmm4, xmm6);               /* xmm6 = 72 70 -- 55 50 -- -- -- */ \

  749.     por_r2r(xmm0, xmm2);               /* xmm2 = -- -- 63 -- -- 36 32 17 */ \

  750.     \

  751.     por_r2r(xmm6, xmm2);               /* xmm2 = 72 70 64 55 50 36 32 17 */ \

  752.     pshufhw_r2r(xmm1, xmm1, 0x0C9);    /* xmm1 = 67 64 66 65 -- -- -- -- */ \

  753.     \

  754.     movdqu_r2r(xmm3, xmm6);            /* xmm6 = 35 33 34 -- -- -- -- -- */ \

  755.     movdqu_r2m(xmm2, *(eax+80));       /* write  72 70 64 55 50 36 32 17 */ \

  756.     \

  757.     psrldq_i2r(12, xmm6);              /* xmm6 = -- -- -- -- -- -- 35 33 */ \

  758.     pslldq_i2r(4, xmm3);               /* xmm3 = 34 -- -- -- -- -- -- -- */ \

  759.     \

  760.     pshuflw_r2r(xmm5, xmm5, 0x04E);    /* xmm5 = -- -- -- 54 51 -- 53 52 */ \

  761.     movdqu_r2r(xmm7, xmm4);            /* xmm4 = 77 75 74 76 73 -- -- -- */ \

  762.     \

  763.     movdqu_r2r(xmm5, xmm2);            /* xmm2 = -- -- -- 54 51 -- 53 52 */ \

  764.     psrldq_i2r(10, xmm7);              /* xmm7 = -- -- -- -- -- 77 75 74 */ \

  765.     \

  766.     pslldq_i2r(6, xmm4);               /* xmm4 = 76 73 -- -- -- -- -- -- */ \

  767.     pslldq_i2r(12, xmm2);              /* xmm2 = 53 52 -- -- -- -- -- -- */ \

  768.     \

  769.     movdqu_r2r(xmm1, xmm0);            /* xmm0 = 67 64 66 65 -- -- -- -- */ \

  770.     psrldq_i2r(12, xmm1);              /* xmm1 = -- -- -- -- -- -- 67 64 */ \

  771.     \

  772.     psrldq_i2r(6, xmm5);               /* xmm5 = -- -- -- -- -- -- 54 51 */ \

  773.     psrldq_i2r(14, xmm3);              /* xmm3 = -- -- -- -- -- -- -- 34 */ \

  774.     \

  775.     pslldq_i2r(10, xmm7);              /* xmm7 = 77 75 74 -- -- -- -- -- */ \

  776.     por_r2r(xmm6, xmm4);               /* xmm4 = 76 73 -- -- -- -- 35 33 */ \

  777.     \

  778.     psrldq_i2r(10, xmm2);              /* xmm2 = -- -- -- -- -- 53 52 -- */ \

  779.     pslldq_i2r(4, xmm0);               /* xmm0 = 66 65 -- -- -- -- -- -- */ \

  780.     \

  781.     pslldq_i2r(8, xmm1);               /* xmm1 = -- -- 67 64 -- -- -- -- */ \

  782.     por_r2r(xmm7, xmm3);               /* xmm3 = 77 75 74 -- -- -- -- 34 */ \

  783.     \

  784.     psrldq_i2r(6, xmm0);               /* xmm0 = -- -- -- 66 65 -- -- -- */ \

  785.     pslldq_i2r(4, xmm5);               /* xmm5 = -- -- -- -- 54 51 -- -- */ \

  786.     \

  787.     por_r2r(xmm1, xmm4);               /* xmm4 = 76 73 67 64 -- -- 35 33 */ \

  788.     por_r2r(xmm2, xmm3);               /* xmm3 = 77 75 74 -- -- 53 52 34 */ \

  789.     \

  790.     por_r2r(xmm5, xmm4);               /* xmm4 = 76 73 67 64 54 51 35 33 */ \

  791.     por_r2r(xmm0, xmm3);               /* xmm3 = 77 75 74 66 65 53 52 34 */ \

  792.     \

  793.     movdqu_r2m(xmm4, *(eax+96));       /* write  76 73 67 64 54 51 35 33 */ \

  794.     movdqu_r2m(xmm3, *(eax+112));      /* write  77 75 74 66 65 53 52 34 */ \

  795.     \

  796. } /* end of SSE2_Dequantize Macro */

  797. 

  798. 

  799. void vp3_dsp_init_sse2(void)

  800. {

  801.     /* nop */

  802. }

  803. 

  804. 

  805. void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,

  806.     int coeff_count, int16_t *output_data)

  807. {

  808.     unsigned char *input_bytes = (unsigned char *)input_data;

  809.     unsigned char *dequant_matrix_bytes = (unsigned char *)dequant_matrix;

  810.     unsigned char *dequant_const_bytes = (unsigned char *)SSE2_dequant_const;

  811.     unsigned char *output_data_bytes = (unsigned char *)output_data;

  812.     unsigned char *idct_data_bytes = (unsigned char *)SSE2_idct_data;

  813.     unsigned char *Eight = (unsigned char *)eight_data;

  814. 

  815. #define eax input_bytes

  816. #define ebx dequant_matrix_bytes

  817. #define ecx dequant_const_bytes

  818. #define edx idct_data_bytes

  819. 

  820. #define I(i) (eax + 16 * i)

  821. #define O(i) (ebx + 16 * i)

  822. #define C(i) (edx + 16 * (i-1))

  823. 

  824.     SSE2_Dequantize();

  825. 

  826. #undef ebx

  827. #define ebx output_data_bytes

  828. 

  829.     SSE2_Row_IDCT();

  830. 

  831.     SSE2_Transpose();

  832.         

  833.     SSE2_Column_IDCT();

  834. }