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 "mmx.h"
  27
  28 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
  29
  30 #define ROW_SHIFT 11
  31 #define COL_SHIFT 6
  32
  33 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
  34 #define rounder(bias) {round (bias), round (bias)}
  35
  36
  37 #if 0
  38 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
  39 static inline void idct_row (int16_t * row, int offset,
  40                              int16_t * table, int32_t * rounder)
  41 {
  42     int C1, C2, C3, C4, C5, C6, C7;
  43     int a0, a1, a2, a3, b0, b1, b2, b3;
  44
  45     row += offset;
  46
  47     C1 = table[1];
  48     C2 = table[2];
  49     C3 = table[3];
  50     C4 = table[4];
  51     C5 = table[5];
  52     C6 = table[6];
  53     C7 = table[7];
  54
  55     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
  56     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
  57     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
  58     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
  59
  60     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
  61     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
  62     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
  63     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
  64
  65     row[0] = (a0 + b0) >> ROW_SHIFT;
  66     row[1] = (a1 + b1) >> ROW_SHIFT;
  67     row[2] = (a2 + b2) >> ROW_SHIFT;
  68     row[3] = (a3 + b3) >> ROW_SHIFT;
  69     row[4] = (a3 - b3) >> ROW_SHIFT;
  70     row[5] = (a2 - b2) >> ROW_SHIFT;
  71     row[6] = (a1 - b1) >> ROW_SHIFT;
  72     row[7] = (a0 - b0) >> ROW_SHIFT;
  73 }
  74 #endif
  75
  76
  77 /* MMXEXT row IDCT */
  78
  79 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
  80                                                    c4,  c6,  c4,  c6,   \
  81                                                    c1,  c3, -c1, -c5,   \
  82                                                    c5,  c7,  c3, -c7,   \
  83                                                    c4, -c6,  c4, -c6,   \
  84                                                   -c4,  c2,  c4, -c2,   \
  85                                                    c5, -c1,  c3, -c1,   \
  86                                                    c7,  c3,  c7, -c5 }
  87
  88 static inline void mmxext_row_head (int16_t * const row, const int offset,
  89                                     const int16_t * const table)
  90 {
  91     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
  92
  93     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
  94     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
  95
  96     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
  97     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
  98
  99     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
 100     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
 101
 102     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 103 }
 104
 105 static inline void mmxext_row (const int16_t * const table,
 106                                const int32_t * const rounder)
 107 {
 108     movq_m2r (*(table+8), mm1);         /* mm1 = -C5 -C1 C3 C1 */
 109     pmaddwd_r2r (mm2, mm4);             /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
 110
 111     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
 112     pshufw_r2r (mm6, mm6, 0x4e);        /* mm6 = x3 x1 x7 x5 */
 113
 114     movq_m2r (*(table+12), mm7);        /* mm7 = -C7 C3 C7 C5 */
 115     pmaddwd_r2r (mm5, mm1);             /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
 116
 117     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 118     pmaddwd_r2r (mm6, mm7);             /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
 119
 120     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
 121     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 122
 123     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
 124     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 125
 126     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
 127     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 128
 129     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 130     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 131
 132     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 133     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 134
 135     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 136     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 137
 138     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 139     movq_r2r (mm0, mm4);                /* mm4 = a3 a2 + rounder */
 140
 141     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 142     psubd_r2r (mm5, mm4);               /* mm4 = a3-b3 a2-b2 + rounder */
 143 }
 144
 145 static inline void mmxext_row_tail (int16_t * const row, const int store)
 146 {
 147     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 148
 149     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 150
 151     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 152
 153     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 154
 155     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 156     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 157
 158     /* slot */
 159
 160     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 161 }
 162
 163 static inline void mmxext_row_mid (int16_t * const row, const int store,
 164                                    const int offset,
 165                                    const int16_t * const table)
 166 {
 167     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 168     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 169
 170     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 171     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 172
 173     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 174     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 175
 176     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 177     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 178
 179     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 180     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 181
 182     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
 183     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 184
 185     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
 186
 187     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
 188     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 189 }
 190
 191
 192 /* MMX row IDCT */
 193
 194 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
 195                                            c4,  c6, -c4, -c2,   \
 196                                            c1,  c3,  c3, -c7,   \
 197                                            c5,  c7, -c1, -c5,   \
 198                                            c4, -c6,  c4, -c2,   \
 199                                           -c4,  c2,  c4, -c6,   \
 200                                            c5, -c1,  c7, -c5,   \
 201                                            c7,  c3,  c3, -c1 }
 202
 203 static inline void mmx_row_head (int16_t * const row, const int offset,
 204                                  const int16_t * const table)
 205 {
 206     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 207
 208     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 209     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 210
 211     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 212     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 213
 214     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 215
 216     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 217     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 218
 219     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 220     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 221 }
 222
 223 static inline void mmx_row (const int16_t * const table,
 224                             const int32_t * const rounder)
 225 {
 226     pmaddwd_r2r (mm2, mm4);             /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
 227     punpckldq_r2r (mm5, mm5);           /* mm5 = x3 x1 x3 x1 */
 228
 229     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
 230     punpckhdq_r2r (mm6, mm6);           /* mm6 = x7 x5 x7 x5 */
 231
 232     movq_m2r (*(table+12), mm7);        /* mm7 = -C5 -C1 C7 C5 */
 233     pmaddwd_r2r (mm5, mm1);             /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
 234
 235     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 236     pmaddwd_r2r (mm6, mm7);             /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
 237
 238     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
 239     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 240
 241     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
 242     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 243
 244     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
 245     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 246
 247     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 248     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 249
 250     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 251     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 252
 253     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 254     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 255
 256     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 257     movq_r2r (mm0, mm7);                /* mm7 = a3 a2 + rounder */
 258
 259     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 260     psubd_r2r (mm5, mm7);               /* mm7 = a3-b3 a2-b2 + rounder */
 261 }
 262
 263 static inline void mmx_row_tail (int16_t * const row, const int store)
 264 {
 265     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 266
 267     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 268
 269     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 270
 271     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 272
 273     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 274     movq_r2r (mm7, mm4);                /* mm4 = y6 y7 y4 y5 */
 275
 276     pslld_i2r (16, mm7);                /* mm7 = y7 0 y5 0 */
 277
 278     psrld_i2r (16, mm4);                /* mm4 = 0 y6 0 y4 */
 279
 280     por_r2r (mm4, mm7);                 /* mm7 = y7 y6 y5 y4 */
 281
 282     /* slot */
 283
 284     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 285 }
 286
 287 static inline void mmx_row_mid (int16_t * const row, const int store,
 288                                 const int offset, const int16_t * const table)
 289 {
 290     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 291     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 292
 293     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 294     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 295
 296     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 297     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 298
 299     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 300     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 301
 302     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 303     movq_r2r (mm7, mm1);                /* mm1 = y6 y7 y4 y5 */
 304
 305     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 306     psrld_i2r (16, mm7);                /* mm7 = 0 y6 0 y4 */
 307
 308     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 309     pslld_i2r (16, mm1);                /* mm1 = y7 0 y5 0 */
 310
 311     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 312     por_r2r (mm1, mm7);                 /* mm7 = y7 y6 y5 y4 */
 313
 314     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 315     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 316
 317     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 318     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 319 }
 320
 321
 322 #if 0
 323 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
 324 static inline void idct_col (int16_t * col, int offset)
 325 {
 326 /* multiplication - as implemented on mmx */
 327 #define F(c,x) (((c) * (x)) >> 16)
 328
 329 /* saturation - it helps us handle torture test cases */
 330 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
 331
 332     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
 333     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
 334     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
 335     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
 336
 337     col += offset;
 338
 339     x0 = col[0*8];
 340     x1 = col[1*8];
 341     x2 = col[2*8];
 342     x3 = col[3*8];
 343     x4 = col[4*8];
 344     x5 = col[5*8];
 345     x6 = col[6*8];
 346     x7 = col[7*8];
 347
 348     u04 = S (x0 + x4);
 349     v04 = S (x0 - x4);
 350     u26 = S (F (T2, x6) + x2);
 351     v26 = S (F (T2, x2) - x6);
 352
 353     a0 = S (u04 + u26);
 354     a1 = S (v04 + v26);
 355     a2 = S (v04 - v26);
 356     a3 = S (u04 - u26);
 357
 358     u17 = S (F (T1, x7) + x1);
 359     v17 = S (F (T1, x1) - x7);
 360     u35 = S (F (T3, x5) + x3);
 361     v35 = S (F (T3, x3) - x5);
 362
 363     b0 = S (u17 + u35);
 364     b3 = S (v17 - v35);
 365     u12 = S (u17 - u35);
 366     v12 = S (v17 + v35);
 367     u12 = S (2 * F (C4, u12));
 368     v12 = S (2 * F (C4, v12));
 369     b1 = S (u12 + v12);
 370     b2 = S (u12 - v12);
 371
 372     y0 = S (a0 + b0) >> COL_SHIFT;
 373     y1 = S (a1 + b1) >> COL_SHIFT;
 374     y2 = S (a2 + b2) >> COL_SHIFT;
 375     y3 = S (a3 + b3) >> COL_SHIFT;
 376
 377     y4 = S (a3 - b3) >> COL_SHIFT;
 378     y5 = S (a2 - b2) >> COL_SHIFT;
 379     y6 = S (a1 - b1) >> COL_SHIFT;
 380     y7 = S (a0 - b0) >> COL_SHIFT;
 381
 382     col[0*8] = y0;
 383     col[1*8] = y1;
 384     col[2*8] = y2;
 385     col[3*8] = y3;
 386     col[4*8] = y4;
 387     col[5*8] = y5;
 388     col[6*8] = y6;
 389     col[7*8] = y7;
 390 }
 391 #endif
 392
 393
 394 /* MMX column IDCT */
 395 static inline void idct_col (int16_t * const col, const int offset)
 396 {
 397 #define T1 13036
 398 #define T2 27146
 399 #define T3 43790
 400 #define C4 23170
 401
 402     static const short t1_vector[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
 403     static const short t2_vector[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
 404     static const short t3_vector[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
 405     static const short c4_vector[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
 406
 407     /* column code adapted from Peter Gubanov */
 408     /* http://www.elecard.com/peter/idct.shtml */
 409
 410     movq_m2r (*t1_vector, mm0);         /* mm0 = T1 */
 411
 412     movq_m2r (*(col+offset+1*8), mm1);  /* mm1 = x1 */
 413     movq_r2r (mm0, mm2);                /* mm2 = T1 */
 414
 415     movq_m2r (*(col+offset+7*8), mm4);  /* mm4 = x7 */
 416     pmulhw_r2r (mm1, mm0);              /* mm0 = T1*x1 */
 417
 418     movq_m2r (*t3_vector, mm5);         /* mm5 = T3 */
 419     pmulhw_r2r (mm4, mm2);              /* mm2 = T1*x7 */
 420
 421     movq_m2r (*(col+offset+5*8), mm6);  /* mm6 = x5 */
 422     movq_r2r (mm5, mm7);                /* mm7 = T3-1 */
 423
 424     movq_m2r (*(col+offset+3*8), mm3);  /* mm3 = x3 */
 425     psubsw_r2r (mm4, mm0);              /* mm0 = v17 */
 426
 427     movq_m2r (*t2_vector, mm4);         /* mm4 = T2 */
 428     pmulhw_r2r (mm3, mm5);              /* mm5 = (T3-1)*x3 */
 429
 430     paddsw_r2r (mm2, mm1);              /* mm1 = u17 */
 431     pmulhw_r2r (mm6, mm7);              /* mm7 = (T3-1)*x5 */
 432
 433     /* slot */
 434
 435     movq_r2r (mm4, mm2);                /* mm2 = T2 */
 436     paddsw_r2r (mm3, mm5);              /* mm5 = T3*x3 */
 437
 438     pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
 439     paddsw_r2r (mm6, mm7);              /* mm7 = T3*x5 */
 440
 441     psubsw_r2r (mm6, mm5);              /* mm5 = v35 */
 442     paddsw_r2r (mm3, mm7);              /* mm7 = u35 */
 443
 444     movq_m2r (*(col+offset+6*8), mm3);  /* mm3 = x6 */
 445     movq_r2r (mm0, mm6);                /* mm6 = v17 */
 446
 447     pmulhw_r2r (mm3, mm2);              /* mm2 = T2*x6 */
 448     psubsw_r2r (mm5, mm0);              /* mm0 = b3 */
 449
 450     psubsw_r2r (mm3, mm4);              /* mm4 = v26 */
 451     paddsw_r2r (mm6, mm5);              /* mm5 = v12 */
 452
 453     movq_r2m (mm0, *(col+offset+3*8));  /* save b3 in scratch0 */
 454     movq_r2r (mm1, mm6);                /* mm6 = u17 */
 455
 456     paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
 457     paddsw_r2r (mm7, mm6);              /* mm6 = b0 */
 458
 459     psubsw_r2r (mm7, mm1);              /* mm1 = u12 */
 460     movq_r2r (mm1, mm7);                /* mm7 = u12 */
 461
 462     movq_m2r (*(col+offset+0*8), mm3);  /* mm3 = x0 */
 463     paddsw_r2r (mm5, mm1);              /* mm1 = u12+v12 */
 464
 465     movq_m2r (*c4_vector, mm0);         /* mm0 = C4/2 */
 466     psubsw_r2r (mm5, mm7);              /* mm7 = u12-v12 */
 467
 468     movq_r2m (mm6, *(col+offset+5*8));  /* save b0 in scratch1 */
 469     pmulhw_r2r (mm0, mm1);              /* mm1 = b1/2 */
 470
 471     movq_r2r (mm4, mm6);                /* mm6 = v26 */
 472     pmulhw_r2r (mm0, mm7);              /* mm7 = b2/2 */
 473
 474     movq_m2r (*(col+offset+4*8), mm5);  /* mm5 = x4 */
 475     movq_r2r (mm3, mm0);                /* mm0 = x0 */
 476
 477     psubsw_r2r (mm5, mm3);              /* mm3 = v04 */
 478     paddsw_r2r (mm5, mm0);              /* mm0 = u04 */
 479
 480     paddsw_r2r (mm3, mm4);              /* mm4 = a1 */
 481     movq_r2r (mm0, mm5);                /* mm5 = u04 */
 482
 483     psubsw_r2r (mm6, mm3);              /* mm3 = a2 */
 484     paddsw_r2r (mm2, mm5);              /* mm5 = a0 */
 485
 486     paddsw_r2r (mm1, mm1);              /* mm1 = b1 */
 487     psubsw_r2r (mm2, mm0);              /* mm0 = a3 */
 488
 489     paddsw_r2r (mm7, mm7);              /* mm7 = b2 */
 490     movq_r2r (mm3, mm2);                /* mm2 = a2 */
 491
 492     movq_r2r (mm4, mm6);                /* mm6 = a1 */
 493     paddsw_r2r (mm7, mm3);              /* mm3 = a2+b2 */
 494
 495     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y2 */
 496     paddsw_r2r (mm1, mm4);              /* mm4 = a1+b1 */
 497
 498     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y1 */
 499     psubsw_r2r (mm1, mm6);              /* mm6 = a1-b1 */
 500
 501     movq_m2r (*(col+offset+5*8), mm1);  /* mm1 = b0 */
 502     psubsw_r2r (mm7, mm2);              /* mm2 = a2-b2 */
 503
 504     psraw_i2r (COL_SHIFT, mm6);         /* mm6 = y6 */
 505     movq_r2r (mm5, mm7);                /* mm7 = a0 */
 506
 507     movq_r2m (mm4, *(col+offset+1*8));  /* save y1 */
 508     psraw_i2r (COL_SHIFT, mm2);         /* mm2 = y5 */
 509
 510     movq_r2m (mm3, *(col+offset+2*8));  /* save y2 */
 511     paddsw_r2r (mm1, mm5);              /* mm5 = a0+b0 */
 512
 513     movq_m2r (*(col+offset+3*8), mm4);  /* mm4 = b3 */
 514     psubsw_r2r (mm1, mm7);              /* mm7 = a0-b0 */
 515
 516     psraw_i2r (COL_SHIFT, mm5);         /* mm5 = y0 */
 517     movq_r2r (mm0, mm3);                /* mm3 = a3 */
 518
 519     movq_r2m (mm2, *(col+offset+5*8));  /* save y5 */
 520     psubsw_r2r (mm4, mm3);              /* mm3 = a3-b3 */
 521
 522     psraw_i2r (COL_SHIFT, mm7);         /* mm7 = y7 */
 523     paddsw_r2r (mm0, mm4);              /* mm4 = a3+b3 */
 524
 525     movq_r2m (mm5, *(col+offset+0*8));  /* save y0 */
 526     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y4 */
 527
 528     movq_r2m (mm6, *(col+offset+6*8));  /* save y6 */
 529     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y3 */
 530
 531     movq_r2m (mm7, *(col+offset+7*8));  /* save y7 */
 532
 533     movq_r2m (mm3, *(col+offset+4*8));  /* save y4 */
 534
 535     movq_r2m (mm4, *(col+offset+3*8));  /* save y3 */
 536
 537 #undef T1
 538 #undef T2
 539 #undef T3
 540 #undef C4
 541 }
 542
 543
 544 static const int32_t rounder0[] ATTR_ALIGN(8) =
 545     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
 546 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
 547 static const int32_t rounder1[] ATTR_ALIGN(8) =
 548     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
 549 static const int32_t rounder7[] ATTR_ALIGN(8) =
 550     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
 551 static const int32_t rounder2[] ATTR_ALIGN(8) =
 552     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
 553 static const int32_t rounder6[] ATTR_ALIGN(8) =
 554     rounder (-0.25);                /* C2 * (C6-C2)/2 */
 555 static const int32_t rounder3[] ATTR_ALIGN(8) =
 556     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
 557 static const int32_t rounder5[] ATTR_ALIGN(8) =
 558     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
 559
 560 #undef COL_SHIFT
 561 #undef ROW_SHIFT
 562
 563 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
 564 void idct (int16_t * const block)                                       \
 565 {                                                                       \
 566     static const int16_t table04[] ATTR_ALIGN(16) =                     \
 567         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
 568     static const int16_t table17[] ATTR_ALIGN(16) =                     \
 569         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
 570     static const int16_t table26[] ATTR_ALIGN(16) =                     \
 571         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
 572     static const int16_t table35[] ATTR_ALIGN(16) =                     \
 573         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
 574                                                                         \
 575     idct_row_head (block, 0*8, table04);                                \
 576     idct_row (table04, rounder0);                                       \
 577     idct_row_mid (block, 0*8, 4*8, table04);                            \
 578     idct_row (table04, rounder4);                                       \
 579     idct_row_mid (block, 4*8, 1*8, table17);                            \
 580     idct_row (table17, rounder1);                                       \
 581     idct_row_mid (block, 1*8, 7*8, table17);                            \
 582     idct_row (table17, rounder7);                                       \
 583     idct_row_mid (block, 7*8, 2*8, table26);                            \
 584     idct_row (table26, rounder2);                                       \
 585     idct_row_mid (block, 2*8, 6*8, table26);                            \
 586     idct_row (table26, rounder6);                                       \
 587     idct_row_mid (block, 6*8, 3*8, table35);                            \
 588     idct_row (table35, rounder3);                                       \
 589     idct_row_mid (block, 3*8, 5*8, table35);                            \
 590     idct_row (table35, rounder5);                                       \
 591     idct_row_tail (block, 5*8);                                         \
 592                                                                         \
 593     idct_col (block, 0);                                                \
 594     idct_col (block, 4);                                                \
 595 }
 596
 597 void ff_mmx_idct(DCTELEM *block);
 598 void ff_mmxext_idct(DCTELEM *block);
 599
 600 declare_idct (ff_mmxext_idct, mmxext_table,
 601               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
 602
 603 declare_idct (ff_mmx_idct, mmx_table,
 604               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
 605