libavcodec/x86/idct_mmx.c

   1 /*
   2  * idct_mmx.c
   3  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
   4  *
   5  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
   6  * See http://libmpeg2.sourceforge.net/ for updates.
   7  *
   8  * mpeg2dec is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * mpeg2dec is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with mpeg2dec; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "libavutil/common.h"
  24 #include "libavcodec/dsputil.h"
  25
  26 #include "dsputil_mmx.h"
  27 #include "mmx.h"
  28
  29 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
  30
  31 #define ROW_SHIFT 11
  32 #define COL_SHIFT 6
  33
  34 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
  35 #define rounder(bias) {round (bias), round (bias)}
  36
  37
  38 #if 0
  39 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
  40 static inline void idct_row (int16_t * row, int offset,
  41                              int16_t * table, int32_t * rounder)
  42 {
  43     int C1, C2, C3, C4, C5, C6, C7;
  44     int a0, a1, a2, a3, b0, b1, b2, b3;
  45
  46     row += offset;
  47
  48     C1 = table[1];
  49     C2 = table[2];
  50     C3 = table[3];
  51     C4 = table[4];
  52     C5 = table[5];
  53     C6 = table[6];
  54     C7 = table[7];
  55
  56     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
  57     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
  58     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
  59     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
  60
  61     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
  62     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
  63     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
  64     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
  65
  66     row[0] = (a0 + b0) >> ROW_SHIFT;
  67     row[1] = (a1 + b1) >> ROW_SHIFT;
  68     row[2] = (a2 + b2) >> ROW_SHIFT;
  69     row[3] = (a3 + b3) >> ROW_SHIFT;
  70     row[4] = (a3 - b3) >> ROW_SHIFT;
  71     row[5] = (a2 - b2) >> ROW_SHIFT;
  72     row[6] = (a1 - b1) >> ROW_SHIFT;
  73     row[7] = (a0 - b0) >> ROW_SHIFT;
  74 }
  75 #endif
  76
  77
  78 /* MMXEXT row IDCT */
  79
  80 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
  81                                                    c4,  c6,  c4,  c6,   \
  82                                                    c1,  c3, -c1, -c5,   \
  83                                                    c5,  c7,  c3, -c7,   \
  84                                                    c4, -c6,  c4, -c6,   \
  85                                                   -c4,  c2,  c4, -c2,   \
  86                                                    c5, -c1,  c3, -c1,   \
  87                                                    c7,  c3,  c7, -c5 }
  88
  89 static inline void mmxext_row_head (int16_t * const row, const int offset,
  90                                     const int16_t * const table)
  91 {
  92     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
  93
  94     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
  95     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
  96
  97     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
  98     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
  99
 100     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
 101     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
 102
 103     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 104 }
 105
 106 static inline void mmxext_row (const int16_t * const table,
 107                                const int32_t * const rounder)
 108 {
 109     movq_m2r (*(table+8), mm1);         /* mm1 = -C5 -C1 C3 C1 */
 110     pmaddwd_r2r (mm2, mm4);             /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
 111
 112     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
 113     pshufw_r2r (mm6, mm6, 0x4e);        /* mm6 = x3 x1 x7 x5 */
 114
 115     movq_m2r (*(table+12), mm7);        /* mm7 = -C7 C3 C7 C5 */
 116     pmaddwd_r2r (mm5, mm1);             /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
 117
 118     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 119     pmaddwd_r2r (mm6, mm7);             /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
 120
 121     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
 122     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 123
 124     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
 125     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 126
 127     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
 128     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 129
 130     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 131     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 132
 133     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 134     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 135
 136     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 137     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 138
 139     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 140     movq_r2r (mm0, mm4);                /* mm4 = a3 a2 + rounder */
 141
 142     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 143     psubd_r2r (mm5, mm4);               /* mm4 = a3-b3 a2-b2 + rounder */
 144 }
 145
 146 static inline void mmxext_row_tail (int16_t * const row, const int store)
 147 {
 148     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 149
 150     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 151
 152     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 153
 154     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 155
 156     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 157     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 158
 159     /* slot */
 160
 161     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 162 }
 163
 164 static inline void mmxext_row_mid (int16_t * const row, const int store,
 165                                    const int offset,
 166                                    const int16_t * const table)
 167 {
 168     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 169     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 170
 171     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 172     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 173
 174     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 175     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 176
 177     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 178     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 179
 180     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 181     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 182
 183     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
 184     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 185
 186     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
 187
 188     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
 189     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 190 }
 191
 192
 193 /* MMX row IDCT */
 194
 195 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
 196                                            c4,  c6, -c4, -c2,   \
 197                                            c1,  c3,  c3, -c7,   \
 198                                            c5,  c7, -c1, -c5,   \
 199                                            c4, -c6,  c4, -c2,   \
 200                                           -c4,  c2,  c4, -c6,   \
 201                                            c5, -c1,  c7, -c5,   \
 202                                            c7,  c3,  c3, -c1 }
 203
 204 static inline void mmx_row_head (int16_t * const row, const int offset,
 205                                  const int16_t * const table)
 206 {
 207     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 208
 209     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 210     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 211
 212     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 213     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 214
 215     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 216
 217     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 218     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 219
 220     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 221     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 222 }
 223
 224 static inline void mmx_row (const int16_t * const table,
 225                             const int32_t * const rounder)
 226 {
 227     pmaddwd_r2r (mm2, mm4);             /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
 228     punpckldq_r2r (mm5, mm5);           /* mm5 = x3 x1 x3 x1 */
 229
 230     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
 231     punpckhdq_r2r (mm6, mm6);           /* mm6 = x7 x5 x7 x5 */
 232
 233     movq_m2r (*(table+12), mm7);        /* mm7 = -C5 -C1 C7 C5 */
 234     pmaddwd_r2r (mm5, mm1);             /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
 235
 236     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 237     pmaddwd_r2r (mm6, mm7);             /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
 238
 239     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
 240     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 241
 242     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
 243     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 244
 245     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
 246     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 247
 248     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 249     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 250
 251     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 252     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 253
 254     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 255     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 256
 257     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 258     movq_r2r (mm0, mm7);                /* mm7 = a3 a2 + rounder */
 259
 260     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 261     psubd_r2r (mm5, mm7);               /* mm7 = a3-b3 a2-b2 + rounder */
 262 }
 263
 264 static inline void mmx_row_tail (int16_t * const row, const int store)
 265 {
 266     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 267
 268     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 269
 270     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 271
 272     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 273
 274     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 275     movq_r2r (mm7, mm4);                /* mm4 = y6 y7 y4 y5 */
 276
 277     pslld_i2r (16, mm7);                /* mm7 = y7 0 y5 0 */
 278
 279     psrld_i2r (16, mm4);                /* mm4 = 0 y6 0 y4 */
 280
 281     por_r2r (mm4, mm7);                 /* mm7 = y7 y6 y5 y4 */
 282
 283     /* slot */
 284
 285     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 286 }
 287
 288 static inline void mmx_row_mid (int16_t * const row, const int store,
 289                                 const int offset, const int16_t * const table)
 290 {
 291     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 292     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 293
 294     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 295     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 296
 297     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 298     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 299
 300     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 301     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 302
 303     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 304     movq_r2r (mm7, mm1);                /* mm1 = y6 y7 y4 y5 */
 305
 306     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 307     psrld_i2r (16, mm7);                /* mm7 = 0 y6 0 y4 */
 308
 309     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 310     pslld_i2r (16, mm1);                /* mm1 = y7 0 y5 0 */
 311
 312     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 313     por_r2r (mm1, mm7);                 /* mm7 = y7 y6 y5 y4 */
 314
 315     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 316     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 317
 318     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 319     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 320 }
 321
 322
 323 #if 0
 324 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
 325 static inline void idct_col (int16_t * col, int offset)
 326 {
 327 /* multiplication - as implemented on mmx */
 328 #define F(c,x) (((c) * (x)) >> 16)
 329
 330 /* saturation - it helps us handle torture test cases */
 331 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
 332
 333     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
 334     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
 335     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
 336     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
 337
 338     col += offset;
 339
 340     x0 = col[0*8];
 341     x1 = col[1*8];
 342     x2 = col[2*8];
 343     x3 = col[3*8];
 344     x4 = col[4*8];
 345     x5 = col[5*8];
 346     x6 = col[6*8];
 347     x7 = col[7*8];
 348
 349     u04 = S (x0 + x4);
 350     v04 = S (x0 - x4);
 351     u26 = S (F (T2, x6) + x2);
 352     v26 = S (F (T2, x2) - x6);
 353
 354     a0 = S (u04 + u26);
 355     a1 = S (v04 + v26);
 356     a2 = S (v04 - v26);
 357     a3 = S (u04 - u26);
 358
 359     u17 = S (F (T1, x7) + x1);
 360     v17 = S (F (T1, x1) - x7);
 361     u35 = S (F (T3, x5) + x3);
 362     v35 = S (F (T3, x3) - x5);
 363
 364     b0 = S (u17 + u35);
 365     b3 = S (v17 - v35);
 366     u12 = S (u17 - u35);
 367     v12 = S (v17 + v35);
 368     u12 = S (2 * F (C4, u12));
 369     v12 = S (2 * F (C4, v12));
 370     b1 = S (u12 + v12);
 371     b2 = S (u12 - v12);
 372
 373     y0 = S (a0 + b0) >> COL_SHIFT;
 374     y1 = S (a1 + b1) >> COL_SHIFT;
 375     y2 = S (a2 + b2) >> COL_SHIFT;
 376     y3 = S (a3 + b3) >> COL_SHIFT;
 377
 378     y4 = S (a3 - b3) >> COL_SHIFT;
 379     y5 = S (a2 - b2) >> COL_SHIFT;
 380     y6 = S (a1 - b1) >> COL_SHIFT;
 381     y7 = S (a0 - b0) >> COL_SHIFT;
 382
 383     col[0*8] = y0;
 384     col[1*8] = y1;
 385     col[2*8] = y2;
 386     col[3*8] = y3;
 387     col[4*8] = y4;
 388     col[5*8] = y5;
 389     col[6*8] = y6;
 390     col[7*8] = y7;
 391 }
 392 #endif
 393
 394
 395 /* MMX column IDCT */
 396 static inline void idct_col (int16_t * const col, const int offset)
 397 {
 398 #define T1 13036
 399 #define T2 27146
 400 #define T3 43790
 401 #define C4 23170
 402
 403     static const short t1_vector[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
 404     static const short t2_vector[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
 405     static const short t3_vector[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
 406     static const short c4_vector[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
 407
 408     /* column code adapted from Peter Gubanov */
 409     /* http://www.elecard.com/peter/idct.shtml */
 410
 411     movq_m2r (*t1_vector, mm0);         /* mm0 = T1 */
 412
 413     movq_m2r (*(col+offset+1*8), mm1);  /* mm1 = x1 */
 414     movq_r2r (mm0, mm2);                /* mm2 = T1 */
 415
 416     movq_m2r (*(col+offset+7*8), mm4);  /* mm4 = x7 */
 417     pmulhw_r2r (mm1, mm0);              /* mm0 = T1*x1 */
 418
 419     movq_m2r (*t3_vector, mm5);         /* mm5 = T3 */
 420     pmulhw_r2r (mm4, mm2);              /* mm2 = T1*x7 */
 421
 422     movq_m2r (*(col+offset+5*8), mm6);  /* mm6 = x5 */
 423     movq_r2r (mm5, mm7);                /* mm7 = T3-1 */
 424
 425     movq_m2r (*(col+offset+3*8), mm3);  /* mm3 = x3 */
 426     psubsw_r2r (mm4, mm0);              /* mm0 = v17 */
 427
 428     movq_m2r (*t2_vector, mm4);         /* mm4 = T2 */
 429     pmulhw_r2r (mm3, mm5);              /* mm5 = (T3-1)*x3 */
 430
 431     paddsw_r2r (mm2, mm1);              /* mm1 = u17 */
 432     pmulhw_r2r (mm6, mm7);              /* mm7 = (T3-1)*x5 */
 433
 434     /* slot */
 435
 436     movq_r2r (mm4, mm2);                /* mm2 = T2 */
 437     paddsw_r2r (mm3, mm5);              /* mm5 = T3*x3 */
 438
 439     pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
 440     paddsw_r2r (mm6, mm7);              /* mm7 = T3*x5 */
 441
 442     psubsw_r2r (mm6, mm5);              /* mm5 = v35 */
 443     paddsw_r2r (mm3, mm7);              /* mm7 = u35 */
 444
 445     movq_m2r (*(col+offset+6*8), mm3);  /* mm3 = x6 */
 446     movq_r2r (mm0, mm6);                /* mm6 = v17 */
 447
 448     pmulhw_r2r (mm3, mm2);              /* mm2 = T2*x6 */
 449     psubsw_r2r (mm5, mm0);              /* mm0 = b3 */
 450
 451     psubsw_r2r (mm3, mm4);              /* mm4 = v26 */
 452     paddsw_r2r (mm6, mm5);              /* mm5 = v12 */
 453
 454     movq_r2m (mm0, *(col+offset+3*8));  /* save b3 in scratch0 */
 455     movq_r2r (mm1, mm6);                /* mm6 = u17 */
 456
 457     paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
 458     paddsw_r2r (mm7, mm6);              /* mm6 = b0 */
 459
 460     psubsw_r2r (mm7, mm1);              /* mm1 = u12 */
 461     movq_r2r (mm1, mm7);                /* mm7 = u12 */
 462
 463     movq_m2r (*(col+offset+0*8), mm3);  /* mm3 = x0 */
 464     paddsw_r2r (mm5, mm1);              /* mm1 = u12+v12 */
 465
 466     movq_m2r (*c4_vector, mm0);         /* mm0 = C4/2 */
 467     psubsw_r2r (mm5, mm7);              /* mm7 = u12-v12 */
 468
 469     movq_r2m (mm6, *(col+offset+5*8));  /* save b0 in scratch1 */
 470     pmulhw_r2r (mm0, mm1);              /* mm1 = b1/2 */
 471
 472     movq_r2r (mm4, mm6);                /* mm6 = v26 */
 473     pmulhw_r2r (mm0, mm7);              /* mm7 = b2/2 */
 474
 475     movq_m2r (*(col+offset+4*8), mm5);  /* mm5 = x4 */
 476     movq_r2r (mm3, mm0);                /* mm0 = x0 */
 477
 478     psubsw_r2r (mm5, mm3);              /* mm3 = v04 */
 479     paddsw_r2r (mm5, mm0);              /* mm0 = u04 */
 480
 481     paddsw_r2r (mm3, mm4);              /* mm4 = a1 */
 482     movq_r2r (mm0, mm5);                /* mm5 = u04 */
 483
 484     psubsw_r2r (mm6, mm3);              /* mm3 = a2 */
 485     paddsw_r2r (mm2, mm5);              /* mm5 = a0 */
 486
 487     paddsw_r2r (mm1, mm1);              /* mm1 = b1 */
 488     psubsw_r2r (mm2, mm0);              /* mm0 = a3 */
 489
 490     paddsw_r2r (mm7, mm7);              /* mm7 = b2 */
 491     movq_r2r (mm3, mm2);                /* mm2 = a2 */
 492
 493     movq_r2r (mm4, mm6);                /* mm6 = a1 */
 494     paddsw_r2r (mm7, mm3);              /* mm3 = a2+b2 */
 495
 496     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y2 */
 497     paddsw_r2r (mm1, mm4);              /* mm4 = a1+b1 */
 498
 499     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y1 */
 500     psubsw_r2r (mm1, mm6);              /* mm6 = a1-b1 */
 501
 502     movq_m2r (*(col+offset+5*8), mm1);  /* mm1 = b0 */
 503     psubsw_r2r (mm7, mm2);              /* mm2 = a2-b2 */
 504
 505     psraw_i2r (COL_SHIFT, mm6);         /* mm6 = y6 */
 506     movq_r2r (mm5, mm7);                /* mm7 = a0 */
 507
 508     movq_r2m (mm4, *(col+offset+1*8));  /* save y1 */
 509     psraw_i2r (COL_SHIFT, mm2);         /* mm2 = y5 */
 510
 511     movq_r2m (mm3, *(col+offset+2*8));  /* save y2 */
 512     paddsw_r2r (mm1, mm5);              /* mm5 = a0+b0 */
 513
 514     movq_m2r (*(col+offset+3*8), mm4);  /* mm4 = b3 */
 515     psubsw_r2r (mm1, mm7);              /* mm7 = a0-b0 */
 516
 517     psraw_i2r (COL_SHIFT, mm5);         /* mm5 = y0 */
 518     movq_r2r (mm0, mm3);                /* mm3 = a3 */
 519
 520     movq_r2m (mm2, *(col+offset+5*8));  /* save y5 */
 521     psubsw_r2r (mm4, mm3);              /* mm3 = a3-b3 */
 522
 523     psraw_i2r (COL_SHIFT, mm7);         /* mm7 = y7 */
 524     paddsw_r2r (mm0, mm4);              /* mm4 = a3+b3 */
 525
 526     movq_r2m (mm5, *(col+offset+0*8));  /* save y0 */
 527     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y4 */
 528
 529     movq_r2m (mm6, *(col+offset+6*8));  /* save y6 */
 530     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y3 */
 531
 532     movq_r2m (mm7, *(col+offset+7*8));  /* save y7 */
 533
 534     movq_r2m (mm3, *(col+offset+4*8));  /* save y4 */
 535
 536     movq_r2m (mm4, *(col+offset+3*8));  /* save y3 */
 537
 538 #undef T1
 539 #undef T2
 540 #undef T3
 541 #undef C4
 542 }
 543
 544
 545 static const int32_t rounder0[] ATTR_ALIGN(8) =
 546     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
 547 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
 548 static const int32_t rounder1[] ATTR_ALIGN(8) =
 549     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
 550 static const int32_t rounder7[] ATTR_ALIGN(8) =
 551     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
 552 static const int32_t rounder2[] ATTR_ALIGN(8) =
 553     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
 554 static const int32_t rounder6[] ATTR_ALIGN(8) =
 555     rounder (-0.25);                /* C2 * (C6-C2)/2 */
 556 static const int32_t rounder3[] ATTR_ALIGN(8) =
 557     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
 558 static const int32_t rounder5[] ATTR_ALIGN(8) =
 559     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
 560
 561 #undef COL_SHIFT
 562 #undef ROW_SHIFT
 563
 564 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
 565 void idct (int16_t * const block)                                       \
 566 {                                                                       \
 567     static const int16_t table04[] ATTR_ALIGN(16) =                     \
 568         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
 569     static const int16_t table17[] ATTR_ALIGN(16) =                     \
 570         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
 571     static const int16_t table26[] ATTR_ALIGN(16) =                     \
 572         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
 573     static const int16_t table35[] ATTR_ALIGN(16) =                     \
 574         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
 575                                                                         \
 576     idct_row_head (block, 0*8, table04);                                \
 577     idct_row (table04, rounder0);                                       \
 578     idct_row_mid (block, 0*8, 4*8, table04);                            \
 579     idct_row (table04, rounder4);                                       \
 580     idct_row_mid (block, 4*8, 1*8, table17);                            \
 581     idct_row (table17, rounder1);                                       \
 582     idct_row_mid (block, 1*8, 7*8, table17);                            \
 583     idct_row (table17, rounder7);                                       \
 584     idct_row_mid (block, 7*8, 2*8, table26);                            \
 585     idct_row (table26, rounder2);                                       \
 586     idct_row_mid (block, 2*8, 6*8, table26);                            \
 587     idct_row (table26, rounder6);                                       \
 588     idct_row_mid (block, 6*8, 3*8, table35);                            \
 589     idct_row (table35, rounder3);                                       \
 590     idct_row_mid (block, 3*8, 5*8, table35);                            \
 591     idct_row (table35, rounder5);                                       \
 592     idct_row_tail (block, 5*8);                                         \
 593                                                                         \
 594     idct_col (block, 0);                                                \
 595     idct_col (block, 4);                                                \
 596 }
 597
 598 declare_idct (ff_mmxext_idct, mmxext_table,
 599               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
 600
 601 declare_idct (ff_mmx_idct, mmx_table,
 602               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
 603