ReactOS Development > Doxygen

00001 /*
00002     dct64_i486.c: DCT64, a plain C variant for i486
00003 
00004     copyright 1998-2006 by the mpg123 project - free software under the terms of the LGPL 2.1
00005     see COPYING and AUTHORS files in distribution or http://mpg123.org
00006     initially written by Fabrice Bellard
00007 */
00008 
00009 /* Discrete Cosine Tansform (DCT) for subband synthesis.
00010  *
00011  * This code is optimized for 80486. It should be compiled with gcc
00012  * 2.7.2 or higher.
00013  *
00014  * Note: This code does not give the necessary accuracy. Moreover, no
00015  * overflow test are done.
00016  *
00017  * (c) 1998 Fabrice Bellard.  
00018  */
00019 
00020 #include "mpg123lib_intern.h"
00021 
00022 #define COS_0_0 16403
00023 #define COS_0_1 16563
00024 #define COS_0_2 16890
00025 #define COS_0_3 17401
00026 #define COS_0_4 18124
00027 #define COS_0_5 19101
00028 #define COS_0_6 20398
00029 #define COS_0_7 22112
00030 #define COS_0_8 24396
00031 #define COS_0_9 27503
00032 #define COS_0_10 31869
00033 #define COS_0_11 38320
00034 #define COS_0_12 48633
00035 #define COS_0_13 67429
00036 #define COS_0_14 111660
00037 #define COS_0_15 333906
00038 #define COS_1_0 16463
00039 #define COS_1_1 17121
00040 #define COS_1_2 18577
00041 #define COS_1_3 21195
00042 #define COS_1_4 25826
00043 #define COS_1_5 34756
00044 #define COS_1_6 56441
00045 #define COS_1_7 167154
00046 #define COS_2_0 16704
00047 #define COS_2_1 19704
00048 #define COS_2_2 29490
00049 #define COS_2_3 83981
00050 #define COS_3_0 17733
00051 #define COS_3_1 42813
00052 #define COS_4_0 23170         
00053 
00054 #define SETOUT(out,n,expr) out[FIR_BUFFER_SIZE*(n)]=(expr)
00055 #define MULL(a,b) (((long long)(a)*(long long)(b)) >> 15)
00056 #define MUL(a,b) \
00057 (\
00058        ((!(b & 0x3F)) ? (((a)*(b >> 6)) >> 9) :\
00059        ((!(b & 0x1F)) ? (((a)*(b >> 5)) >> 10) :\
00060        ((!(b & 0x0F)) ? (((a)*(b >> 4)) >> 11) :\
00061        ((!(b & 0x07)) ? (((a)*(b >> 3)) >> 12) :\
00062        ((!(b & 0x03)) ? (((a)*(b >> 2)) >> 13) :\
00063        ((!(b & 0x01)) ? (((a)*(b >> 1)) >> 14) :\
00064                         (((a)*(b   )) >> 15))))))))
00065 
00066 
00067 void dct64_1_486(int *out0,int *out1,int *b1,int *b2)
00068 {
00069   b1[0x00] = b2[0x00] + b2[0x1F];
00070   b1[0x1F] = MUL((b2[0x00] - b2[0x1F]),COS_0_0);
00071 
00072   b1[0x01] = b2[0x01] + b2[0x1E];
00073   b1[0x1E] = MUL((b2[0x01] - b2[0x1E]),COS_0_1);
00074 
00075   b1[0x02] = b2[0x02] + b2[0x1D];
00076   b1[0x1D] = MUL((b2[0x02] - b2[0x1D]),COS_0_2);
00077 
00078   b1[0x03] = b2[0x03] + b2[0x1C];
00079   b1[0x1C] = MUL((b2[0x03] - b2[0x1C]),COS_0_3);
00080 
00081   b1[0x04] = b2[0x04] + b2[0x1B];
00082   b1[0x1B] = MUL((b2[0x04] - b2[0x1B]),COS_0_4);
00083 
00084   b1[0x05] = b2[0x05] + b2[0x1A];
00085   b1[0x1A] = MUL((b2[0x05] - b2[0x1A]),COS_0_5);
00086 
00087   b1[0x06] = b2[0x06] + b2[0x19];
00088   b1[0x19] = MUL((b2[0x06] - b2[0x19]),COS_0_6);
00089 
00090   b1[0x07] = b2[0x07] + b2[0x18];
00091   b1[0x18] = MUL((b2[0x07] - b2[0x18]),COS_0_7);
00092 
00093   b1[0x08] = b2[0x08] + b2[0x17];
00094   b1[0x17] = MUL((b2[0x08] - b2[0x17]),COS_0_8);
00095 
00096   b1[0x09] = b2[0x09] + b2[0x16];
00097   b1[0x16] = MUL((b2[0x09] - b2[0x16]),COS_0_9);
00098 
00099   b1[0x0A] = b2[0x0A] + b2[0x15];
00100   b1[0x15] = MUL((b2[0x0A] - b2[0x15]),COS_0_10);
00101 
00102   b1[0x0B] = b2[0x0B] + b2[0x14];
00103   b1[0x14] = MUL((b2[0x0B] - b2[0x14]),COS_0_11);
00104 
00105   b1[0x0C] = b2[0x0C] + b2[0x13];
00106   b1[0x13] = MUL((b2[0x0C] - b2[0x13]),COS_0_12);
00107 
00108   b1[0x0D] = b2[0x0D] + b2[0x12];
00109   b1[0x12] = MULL((b2[0x0D] - b2[0x12]),COS_0_13);
00110 
00111   b1[0x0E] = b2[0x0E] + b2[0x11];
00112   b1[0x11] = MULL((b2[0x0E] - b2[0x11]),COS_0_14);
00113 
00114   b1[0x0F] = b2[0x0F] + b2[0x10];
00115   b1[0x10] = MULL((b2[0x0F] - b2[0x10]),COS_0_15);
00116 
00117 
00118   b2[0x00] = b1[0x00] + b1[0x0F]; 
00119   b2[0x0F] = MUL((b1[0x00] - b1[0x0F]),COS_1_0);
00120   b2[0x01] = b1[0x01] + b1[0x0E]; 
00121   b2[0x0E] = MUL((b1[0x01] - b1[0x0E]),COS_1_1);
00122   b2[0x02] = b1[0x02] + b1[0x0D]; 
00123   b2[0x0D] = MUL((b1[0x02] - b1[0x0D]),COS_1_2);
00124   b2[0x03] = b1[0x03] + b1[0x0C]; 
00125   b2[0x0C] = MUL((b1[0x03] - b1[0x0C]),COS_1_3);
00126   b2[0x04] = b1[0x04] + b1[0x0B]; 
00127   b2[0x0B] = MUL((b1[0x04] - b1[0x0B]),COS_1_4);
00128   b2[0x05] = b1[0x05] + b1[0x0A]; 
00129   b2[0x0A] = MUL((b1[0x05] - b1[0x0A]),COS_1_5);
00130   b2[0x06] = b1[0x06] + b1[0x09]; 
00131   b2[0x09] = MUL((b1[0x06] - b1[0x09]),COS_1_6);
00132   b2[0x07] = b1[0x07] + b1[0x08]; 
00133   b2[0x08] = MULL((b1[0x07] - b1[0x08]),COS_1_7);
00134 
00135   b2[0x10] = b1[0x10] + b1[0x1F];
00136   b2[0x1F] = MUL((b1[0x1F] - b1[0x10]),COS_1_0);
00137   b2[0x11] = b1[0x11] + b1[0x1E];
00138   b2[0x1E] = MUL((b1[0x1E] - b1[0x11]),COS_1_1);
00139   b2[0x12] = b1[0x12] + b1[0x1D];
00140   b2[0x1D] = MUL((b1[0x1D] - b1[0x12]),COS_1_2);
00141   b2[0x13] = b1[0x13] + b1[0x1C];
00142   b2[0x1C] = MUL((b1[0x1C] - b1[0x13]),COS_1_3);
00143   b2[0x14] = b1[0x14] + b1[0x1B];
00144   b2[0x1B] = MUL((b1[0x1B] - b1[0x14]),COS_1_4);
00145   b2[0x15] = b1[0x15] + b1[0x1A];
00146   b2[0x1A] = MUL((b1[0x1A] - b1[0x15]),COS_1_5);
00147   b2[0x16] = b1[0x16] + b1[0x19];
00148   b2[0x19] = MUL((b1[0x19] - b1[0x16]),COS_1_6);
00149   b2[0x17] = b1[0x17] + b1[0x18];
00150   b2[0x18] = MULL((b1[0x18] - b1[0x17]),COS_1_7);
00151 
00152 
00153   b1[0x00] = b2[0x00] + b2[0x07];
00154   b1[0x07] = MUL((b2[0x00] - b2[0x07]),COS_2_0);
00155   b1[0x01] = b2[0x01] + b2[0x06];
00156   b1[0x06] = MUL((b2[0x01] - b2[0x06]),COS_2_1);
00157   b1[0x02] = b2[0x02] + b2[0x05];
00158   b1[0x05] = MUL((b2[0x02] - b2[0x05]),COS_2_2);
00159   b1[0x03] = b2[0x03] + b2[0x04];
00160   b1[0x04] = MULL((b2[0x03] - b2[0x04]),COS_2_3);
00161 
00162   b1[0x08] = b2[0x08] + b2[0x0F];
00163   b1[0x0F] = MUL((b2[0x0F] - b2[0x08]),COS_2_0);
00164   b1[0x09] = b2[0x09] + b2[0x0E];
00165   b1[0x0E] = MUL((b2[0x0E] - b2[0x09]),COS_2_1);
00166   b1[0x0A] = b2[0x0A] + b2[0x0D];
00167   b1[0x0D] = MUL((b2[0x0D] - b2[0x0A]),COS_2_2);
00168   b1[0x0B] = b2[0x0B] + b2[0x0C];
00169   b1[0x0C] = MULL((b2[0x0C] - b2[0x0B]),COS_2_3);
00170 
00171   b1[0x10] = b2[0x10] + b2[0x17];
00172   b1[0x17] = MUL((b2[0x10] - b2[0x17]),COS_2_0);
00173   b1[0x11] = b2[0x11] + b2[0x16];
00174   b1[0x16] = MUL((b2[0x11] - b2[0x16]),COS_2_1);
00175   b1[0x12] = b2[0x12] + b2[0x15];
00176   b1[0x15] = MUL((b2[0x12] - b2[0x15]),COS_2_2);
00177   b1[0x13] = b2[0x13] + b2[0x14];
00178   b1[0x14] = MULL((b2[0x13] - b2[0x14]),COS_2_3);
00179 
00180   b1[0x18] = b2[0x18] + b2[0x1F];
00181   b1[0x1F] = MUL((b2[0x1F] - b2[0x18]),COS_2_0);
00182   b1[0x19] = b2[0x19] + b2[0x1E];
00183   b1[0x1E] = MUL((b2[0x1E] - b2[0x19]),COS_2_1);
00184   b1[0x1A] = b2[0x1A] + b2[0x1D];
00185   b1[0x1D] = MUL((b2[0x1D] - b2[0x1A]),COS_2_2);
00186   b1[0x1B] = b2[0x1B] + b2[0x1C];
00187   b1[0x1C] = MULL((b2[0x1C] - b2[0x1B]),COS_2_3);
00188 
00189 
00190   b2[0x00] = b1[0x00] + b1[0x03];
00191   b2[0x03] = MUL((b1[0x00] - b1[0x03]),COS_3_0);
00192   b2[0x01] = b1[0x01] + b1[0x02];
00193   b2[0x02] = MUL((b1[0x01] - b1[0x02]),COS_3_1);
00194 
00195   b2[0x04] = b1[0x04] + b1[0x07];
00196   b2[0x07] = MUL((b1[0x07] - b1[0x04]),COS_3_0);
00197   b2[0x05] = b1[0x05] + b1[0x06];
00198   b2[0x06] = MUL((b1[0x06] - b1[0x05]),COS_3_1);
00199 
00200   b2[0x08] = b1[0x08] + b1[0x0B];
00201   b2[0x0B] = MUL((b1[0x08] - b1[0x0B]),COS_3_0);
00202   b2[0x09] = b1[0x09] + b1[0x0A];
00203   b2[0x0A] = MUL((b1[0x09] - b1[0x0A]),COS_3_1);
00204   
00205   b2[0x0C] = b1[0x0C] + b1[0x0F];
00206   b2[0x0F] = MUL((b1[0x0F] - b1[0x0C]),COS_3_0);
00207   b2[0x0D] = b1[0x0D] + b1[0x0E];
00208   b2[0x0E] = MUL((b1[0x0E] - b1[0x0D]),COS_3_1);
00209 
00210   b2[0x10] = b1[0x10] + b1[0x13];
00211   b2[0x13] = MUL((b1[0x10] - b1[0x13]),COS_3_0);
00212   b2[0x11] = b1[0x11] + b1[0x12];
00213   b2[0x12] = MUL((b1[0x11] - b1[0x12]),COS_3_1);
00214 
00215   b2[0x14] = b1[0x14] + b1[0x17];
00216   b2[0x17] = MUL((b1[0x17] - b1[0x14]),COS_3_0);
00217   b2[0x15] = b1[0x15] + b1[0x16];
00218   b2[0x16] = MUL((b1[0x16] - b1[0x15]),COS_3_1);
00219 
00220   b2[0x18] = b1[0x18] + b1[0x1B];
00221   b2[0x1B] = MUL((b1[0x18] - b1[0x1B]),COS_3_0);
00222   b2[0x19] = b1[0x19] + b1[0x1A];
00223   b2[0x1A] = MUL((b1[0x19] - b1[0x1A]),COS_3_1);
00224 
00225   b2[0x1C] = b1[0x1C] + b1[0x1F];
00226   b2[0x1F] = MUL((b1[0x1F] - b1[0x1C]),COS_3_0);
00227   b2[0x1D] = b1[0x1D] + b1[0x1E];
00228   b2[0x1E] = MUL((b1[0x1E] - b1[0x1D]),COS_3_1);
00229 
00230   {
00231     int i;
00232     for(i=0;i<32;i+=4) {
00233       b1[i+0x00] = b2[i+0x00] + b2[i+0x01];
00234       b1[i+0x01] = MUL((b2[i+0x00] - b2[i+0x01]),COS_4_0);
00235       b1[i+0x02] = b2[i+0x02] + b2[i+0x03];
00236       b1[i+0x03] = MUL((b2[i+0x03] - b2[i+0x02]),COS_4_0);
00237     }
00238   }
00239 
00240   b1[0x02] += b1[0x03];
00241   b1[0x06] += b1[0x07];
00242   b1[0x04] += b1[0x06];
00243   b1[0x06] += b1[0x05];
00244   b1[0x05] += b1[0x07];
00245 
00246   b1[0x0A] += b1[0x0B];
00247   b1[0x0E] += b1[0x0F];
00248   b1[0x0C] += b1[0x0E];
00249   b1[0x0E] += b1[0x0D];
00250   b1[0x0D] += b1[0x0F];
00251 
00252   b1[0x12] += b1[0x13];
00253   b1[0x16] += b1[0x17];
00254   b1[0x14] += b1[0x16];
00255   b1[0x16] += b1[0x15];
00256   b1[0x15] += b1[0x17];
00257 
00258   b1[0x1A] += b1[0x1B];
00259   b1[0x1E] += b1[0x1F];
00260   b1[0x1C] += b1[0x1E];
00261   b1[0x1E] += b1[0x1D];
00262   b1[0x1D] += b1[0x1F];
00263 
00264  SETOUT(out0,16,b1[0x00]);
00265  SETOUT(out0,12,b1[0x04]);
00266  SETOUT(out0, 8,b1[0x02]);
00267  SETOUT(out0, 4,b1[0x06]);
00268  SETOUT(out0, 0,b1[0x01]);
00269  SETOUT(out1, 0,b1[0x01]);
00270  SETOUT(out1, 4,b1[0x05]);
00271  SETOUT(out1, 8,b1[0x03]);
00272  SETOUT(out1,12,b1[0x07]);
00273 
00274  b1[0x08] += b1[0x0C];
00275  SETOUT(out0,14,b1[0x08]);
00276  b1[0x0C] += b1[0x0a];
00277  SETOUT(out0,10,b1[0x0C]);
00278  b1[0x0A] += b1[0x0E];
00279  SETOUT(out0, 6,b1[0x0A]);
00280  b1[0x0E] += b1[0x09];
00281  SETOUT(out0, 2,b1[0x0E]);
00282  b1[0x09] += b1[0x0D];
00283  SETOUT(out1, 2,b1[0x09]);
00284  b1[0x0D] += b1[0x0B];
00285  SETOUT(out1, 6,b1[0x0D]);
00286  b1[0x0B] += b1[0x0F];
00287  SETOUT(out1,10,b1[0x0B]);
00288  SETOUT(out1,14,b1[0x0F]);
00289 
00290  b1[0x18] += b1[0x1C];
00291  SETOUT(out0,15,b1[0x10] + b1[0x18]);
00292  SETOUT(out0,13,b1[0x18] + b1[0x14]);
00293  b1[0x1C] += b1[0x1a];
00294  SETOUT(out0,11,b1[0x14] + b1[0x1C]);
00295  SETOUT(out0, 9,b1[0x1C] + b1[0x12]);
00296  b1[0x1A] += b1[0x1E];
00297  SETOUT(out0, 7,b1[0x12] + b1[0x1A]);
00298  SETOUT(out0, 5,b1[0x1A] + b1[0x16]);
00299  b1[0x1E] += b1[0x19];
00300  SETOUT(out0, 3,b1[0x16] + b1[0x1E]);
00301  SETOUT(out0, 1,b1[0x1E] + b1[0x11]);
00302  b1[0x19] += b1[0x1D];
00303  SETOUT(out1, 1,b1[0x11] + b1[0x19]);
00304  SETOUT(out1, 3,b1[0x19] + b1[0x15]);
00305  b1[0x1D] += b1[0x1B];
00306  SETOUT(out1, 5,b1[0x15] + b1[0x1D]);
00307  SETOUT(out1, 7,b1[0x1D] + b1[0x13]);
00308  b1[0x1B] += b1[0x1F];
00309  SETOUT(out1, 9,b1[0x13] + b1[0x1B]);
00310  SETOUT(out1,11,b1[0x1B] + b1[0x17]);
00311  SETOUT(out1,13,b1[0x17] + b1[0x1F]);
00312  SETOUT(out1,15,b1[0x1F]);
00313 }
00314 
00315 
00316 /*
00317  * the call via dct64 is a trick to force GCC to use
00318  * (new) registers for the b1,b2 pointer to the bufs[xx] field
00319  */
00320 void dct64_i486(int *a,int *b,real *samples)
00321 {
00322   int bufs[64];
00323   int i;
00324 
00325 #ifdef REAL_IS_FIXED  
00326 #define TOINT(a) ((a) * 32768 / (int)REAL_FACTOR)
00327 
00328   for(i=0;i<32;i++) {
00329     bufs[i]=TOINT(samples[i]);
00330   }
00331 #else      
00332   int *p = bufs;
00333   register double const scale = ((65536.0 * 32) + 1) * 65536.0;
00334   
00335   for(i=0;i<32;i++) {
00336     *((double *) (p++)) = scale + *samples++; /* beware on bufs overrun: 8B store from x87 */
00337   }
00338 #endif
00339   
00340   dct64_1_486(a,b,bufs+32,bufs);
00341 }
00342
Generated on Sun May 27 2012 04:34:09 for ReactOS by
1.7.6.1