开通博客赚积分
发布资源赚积分
分类

源码开发语言/平台
当前位置：首页 > 源码/资料 > 多媒体 > 流媒体/Mpeg4/MP4 > 查看源码
dct64_i386.cpp
资源名称：playa-0.3.3src.zip　[点击查看]
上传用户：lusi_8715
上传日期：2007-01-08
资源大小：199k
文件大小：11k
源码类别：

流媒体/Mpeg4/MP4

开发平台：

Visual C++

dct64_i386.cpp：源码内容
  1. /**************************************************************************************
  2.  *                                                                                    *
  3.  * This application contains code from OpenDivX and is released as a "Larger Work"    *
  4.  * under that license. Consistant with that license, this application is released     *
  5.  * under the GNU General Public License.                                              *
  6.  *                                                                                    *
  7.  * The OpenDivX license can be found at: http://www.projectmayo.com/opendivx/docs.php *
  8.  * The GPL can be found at: http://www.gnu.org/copyleft/gpl.html                      *
  9.  *                                                                                    *
  10.  * Authors: Damien Chavarria <roy204 at projectmayo.com>                              *
  11.  *                                                                                    *
  12.  **************************************************************************************/
  14. /*
  15.  * Discrete Cosine Tansform (DCT) for subband synthesis
  16.  * optimized for machines with no auto-increment. 
  17.  * The performance is highly compiler dependend. Maybe
  18.  * the dct64.c version for 'normal' processor may be faster
  19.  * even for Intel processors.
  20.  */
  22. #include "mpg123.h"
  24. static void dct64_1(real *out0,real *out1,real *b1,real *b2,real *samples)
  25. {
  26.  {
  27.   register real *costab = pnts[0];
  29.   b1[0x00] = samples[0x00] + samples[0x1F];
  30.   b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
  32.   b1[0x01] = samples[0x01] + samples[0x1E];
  33.   b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
  35.   b1[0x02] = samples[0x02] + samples[0x1D];
  36.   b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
  38.   b1[0x03] = samples[0x03] + samples[0x1C];
  39.   b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
  41.   b1[0x04] = samples[0x04] + samples[0x1B];
  42.   b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
  44.   b1[0x05] = samples[0x05] + samples[0x1A];
  45.   b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
  47.   b1[0x06] = samples[0x06] + samples[0x19];
  48.   b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
  50.   b1[0x07] = samples[0x07] + samples[0x18];
  51.   b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
  53.   b1[0x08] = samples[0x08] + samples[0x17];
  54.   b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
  56.   b1[0x09] = samples[0x09] + samples[0x16];
  57.   b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
  59.   b1[0x0A] = samples[0x0A] + samples[0x15];
  60.   b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
  62.   b1[0x0B] = samples[0x0B] + samples[0x14];
  63.   b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
  65.   b1[0x0C] = samples[0x0C] + samples[0x13];
  66.   b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
  68.   b1[0x0D] = samples[0x0D] + samples[0x12];
  69.   b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
  71.   b1[0x0E] = samples[0x0E] + samples[0x11];
  72.   b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
  74.   b1[0x0F] = samples[0x0F] + samples[0x10];
  75.   b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
  76.  }
  79.  {
  80.   register real *costab = pnts[1];
  82.   b2[0x00] = b1[0x00] + b1[0x0F]; 
  83.   b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
  84.   b2[0x01] = b1[0x01] + b1[0x0E]; 
  85.   b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
  86.   b2[0x02] = b1[0x02] + b1[0x0D]; 
  87.   b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
  88.   b2[0x03] = b1[0x03] + b1[0x0C]; 
  89.   b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
  90.   b2[0x04] = b1[0x04] + b1[0x0B]; 
  91.   b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
  92.   b2[0x05] = b1[0x05] + b1[0x0A]; 
  93.   b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
  94.   b2[0x06] = b1[0x06] + b1[0x09]; 
  95.   b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
  96.   b2[0x07] = b1[0x07] + b1[0x08]; 
  97.   b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
  99.   b2[0x10] = b1[0x10] + b1[0x1F];
  100.   b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
  101.   b2[0x11] = b1[0x11] + b1[0x1E];
  102.   b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
  103.   b2[0x12] = b1[0x12] + b1[0x1D];
  104.   b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
  105.   b2[0x13] = b1[0x13] + b1[0x1C];
  106.   b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
  107.   b2[0x14] = b1[0x14] + b1[0x1B];
  108.   b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
  109.   b2[0x15] = b1[0x15] + b1[0x1A];
  110.   b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
  111.   b2[0x16] = b1[0x16] + b1[0x19];
  112.   b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
  113.   b2[0x17] = b1[0x17] + b1[0x18];
  114.   b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
  115.  }
  117.  {
  118.   register real *costab = pnts[2];
  120.   b1[0x00] = b2[0x00] + b2[0x07];
  121.   b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
  122.   b1[0x01] = b2[0x01] + b2[0x06];
  123.   b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
  124.   b1[0x02] = b2[0x02] + b2[0x05];
  125.   b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
  126.   b1[0x03] = b2[0x03] + b2[0x04];
  127.   b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
  129.   b1[0x08] = b2[0x08] + b2[0x0F];
  130.   b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
  131.   b1[0x09] = b2[0x09] + b2[0x0E];
  132.   b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
  133.   b1[0x0A] = b2[0x0A] + b2[0x0D];
  134.   b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
  135.   b1[0x0B] = b2[0x0B] + b2[0x0C];
  136.   b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
  138.   b1[0x10] = b2[0x10] + b2[0x17];
  139.   b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
  140.   b1[0x11] = b2[0x11] + b2[0x16];
  141.   b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
  142.   b1[0x12] = b2[0x12] + b2[0x15];
  143.   b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
  144.   b1[0x13] = b2[0x13] + b2[0x14];
  145.   b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
  147.   b1[0x18] = b2[0x18] + b2[0x1F];
  148.   b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
  149.   b1[0x19] = b2[0x19] + b2[0x1E];
  150.   b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
  151.   b1[0x1A] = b2[0x1A] + b2[0x1D];
  152.   b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
  153.   b1[0x1B] = b2[0x1B] + b2[0x1C];
  154.   b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
  155.  }
  157.  {
  158.   register real const cos0 = pnts[3][0];
  159.   register real const cos1 = pnts[3][1];
  161.   b2[0x00] = b1[0x00] + b1[0x03];
  162.   b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
  163.   b2[0x01] = b1[0x01] + b1[0x02];
  164.   b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
  166.   b2[0x04] = b1[0x04] + b1[0x07];
  167.   b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
  168.   b2[0x05] = b1[0x05] + b1[0x06];
  169.   b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
  171.   b2[0x08] = b1[0x08] + b1[0x0B];
  172.   b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
  173.   b2[0x09] = b1[0x09] + b1[0x0A];
  174.   b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
  175.   
  176.   b2[0x0C] = b1[0x0C] + b1[0x0F];
  177.   b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
  178.   b2[0x0D] = b1[0x0D] + b1[0x0E];
  179.   b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
  181.   b2[0x10] = b1[0x10] + b1[0x13];
  182.   b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
  183.   b2[0x11] = b1[0x11] + b1[0x12];
  184.   b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
  186.   b2[0x14] = b1[0x14] + b1[0x17];
  187.   b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
  188.   b2[0x15] = b1[0x15] + b1[0x16];
  189.   b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
  191.   b2[0x18] = b1[0x18] + b1[0x1B];
  192.   b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
  193.   b2[0x19] = b1[0x19] + b1[0x1A];
  194.   b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
  196.   b2[0x1C] = b1[0x1C] + b1[0x1F];
  197.   b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
  198.   b2[0x1D] = b1[0x1D] + b1[0x1E];
  199.   b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
  200.  }
  202.  {
  203.   register real const cos0 = pnts[4][0];
  205.   b1[0x00] = b2[0x00] + b2[0x01];
  206.   b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
  207.   b1[0x02] = b2[0x02] + b2[0x03];
  208.   b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
  209.   b1[0x02] += b1[0x03];
  211.   b1[0x04] = b2[0x04] + b2[0x05];
  212.   b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
  213.   b1[0x06] = b2[0x06] + b2[0x07];
  214.   b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
  215.   b1[0x06] += b1[0x07];
  216.   b1[0x04] += b1[0x06];
  217.   b1[0x06] += b1[0x05];
  218.   b1[0x05] += b1[0x07];
  220.   b1[0x08] = b2[0x08] + b2[0x09];
  221.   b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
  222.   b1[0x0A] = b2[0x0A] + b2[0x0B];
  223.   b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
  224.   b1[0x0A] += b1[0x0B];
  226.   b1[0x0C] = b2[0x0C] + b2[0x0D];
  227.   b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
  228.   b1[0x0E] = b2[0x0E] + b2[0x0F];
  229.   b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
  230.   b1[0x0E] += b1[0x0F];
  231.   b1[0x0C] += b1[0x0E];
  232.   b1[0x0E] += b1[0x0D];
  233.   b1[0x0D] += b1[0x0F];
  235.   b1[0x10] = b2[0x10] + b2[0x11];
  236.   b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
  237.   b1[0x12] = b2[0x12] + b2[0x13];
  238.   b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
  239.   b1[0x12] += b1[0x13];
  241.   b1[0x14] = b2[0x14] + b2[0x15];
  242.   b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
  243.   b1[0x16] = b2[0x16] + b2[0x17];
  244.   b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
  245.   b1[0x16] += b1[0x17];
  246.   b1[0x14] += b1[0x16];
  247.   b1[0x16] += b1[0x15];
  248.   b1[0x15] += b1[0x17];
  250.   b1[0x18] = b2[0x18] + b2[0x19];
  251.   b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
  252.   b1[0x1A] = b2[0x1A] + b2[0x1B];
  253.   b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
  254.   b1[0x1A] += b1[0x1B];
  256.   b1[0x1C] = b2[0x1C] + b2[0x1D];
  257.   b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
  258.   b1[0x1E] = b2[0x1E] + b2[0x1F];
  259.   b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
  260.   b1[0x1E] += b1[0x1F];
  261.   b1[0x1C] += b1[0x1E];
  262.   b1[0x1E] += b1[0x1D];
  263.   b1[0x1D] += b1[0x1F];
  264.  }
  266.  out0[0x10*16] = b1[0x00];
  267.  out0[0x10*12] = b1[0x04];
  268.  out0[0x10* 8] = b1[0x02];
  269.  out0[0x10* 4] = b1[0x06];
  270.  out0[0x10* 0] = b1[0x01];
  271.  out1[0x10* 0] = b1[0x01];
  272.  out1[0x10* 4] = b1[0x05];
  273.  out1[0x10* 8] = b1[0x03];
  274.  out1[0x10*12] = b1[0x07];
  276.  b1[0x08] += b1[0x0C];
  277.  out0[0x10*14] = b1[0x08];
  278.  b1[0x0C] += b1[0x0a];
  279.  out0[0x10*10] = b1[0x0C];
  280.  b1[0x0A] += b1[0x0E];
  281.  out0[0x10* 6] = b1[0x0A];
  282.  b1[0x0E] += b1[0x09];
  283.  out0[0x10* 2] = b1[0x0E];
  284.  b1[0x09] += b1[0x0D];
  285.  out1[0x10* 2] = b1[0x09];
  286.  b1[0x0D] += b1[0x0B];
  287.  out1[0x10* 6] = b1[0x0D];
  288.  b1[0x0B] += b1[0x0F];
  289.  out1[0x10*10] = b1[0x0B];
  290.  out1[0x10*14] = b1[0x0F];
  292.  b1[0x18] += b1[0x1C];
  293.  out0[0x10*15] = b1[0x10] + b1[0x18];
  294.  out0[0x10*13] = b1[0x18] + b1[0x14];
  295.  b1[0x1C] += b1[0x1a];
  296.  out0[0x10*11] = b1[0x14] + b1[0x1C];
  297.  out0[0x10* 9] = b1[0x1C] + b1[0x12];
  298.  b1[0x1A] += b1[0x1E];
  299.  out0[0x10* 7] = b1[0x12] + b1[0x1A];
  300.  out0[0x10* 5] = b1[0x1A] + b1[0x16];
  301.  b1[0x1E] += b1[0x19];
  302.  out0[0x10* 3] = b1[0x16] + b1[0x1E];
  303.  out0[0x10* 1] = b1[0x1E] + b1[0x11];
  304.  b1[0x19] += b1[0x1D];
  305.  out1[0x10* 1] = b1[0x11] + b1[0x19];
  306.  out1[0x10* 3] = b1[0x19] + b1[0x15];
  307.  b1[0x1D] += b1[0x1B];
  308.  out1[0x10* 5] = b1[0x15] + b1[0x1D];
  309.  out1[0x10* 7] = b1[0x1D] + b1[0x13];
  310.  b1[0x1B] += b1[0x1F];
  311.  out1[0x10* 9] = b1[0x13] + b1[0x1B];
  312.  out1[0x10*11] = b1[0x1B] + b1[0x17];
  313.  out1[0x10*13] = b1[0x17] + b1[0x1F];
  314.  out1[0x10*15] = b1[0x1F];
  315. }
  317. /*
  318.  * the call via dct64 is a trick to force GCC to use
  319.  * (new) registers for the b1,b2 pointer to the bufs[xx] field
  320.  */
  321. void dct64(real *a,real *b,real *c)
  322. {
  323.   real bufs[0x40];
  324.   dct64_1(a,b,bufs,bufs+0x20,c);
  325. }