libavcodec/i386/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  *
   7  * mpeg2dec is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License as published by
   9  * the Free Software Foundation; either version 2 of the License, or
  10  * (at your option) any later version.
  11  *
  12  * mpeg2dec is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with mpeg2dec; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include "libavutil/common.h"
  23 #include "libavcodec/dsputil.h"
  24
  25 #include "mmx.h"
  26
  27 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
  28
  29 #define ROW_SHIFT 11
  30 #define COL_SHIFT 6
  31
  32 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
  33 #define rounder(bias) {round (bias), round (bias)}
  34
  35 #if 0
  36 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
  37 static inline void idct_row (int16_t * row, int offset,
  38                              int16_t * table, int32_t * rounder)
  39 {
  40     int C1, C2, C3, C4, C5, C6, C7;
  41     int a0, a1, a2, a3, b0, b1, b2, b3;
  42
  43     row += offset;
  44
  45     C1 = table[1];
  46     C2 = table[2];
  47     C3 = table[3];
  48     C4 = table[4];
  49     C5 = table[5];
  50     C6 = table[6];
  51     C7 = table[7];
  52
  53     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
  54     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
  55     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
  56     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
  57
  58     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
  59     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
  60     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
  61     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
  62
  63     row[0] = (a0 + b0) >> ROW_SHIFT;
  64     row[1] = (a1 + b1) >> ROW_SHIFT;
  65     row[2] = (a2 + b2) >> ROW_SHIFT;
  66     row[3] = (a3 + b3) >> ROW_SHIFT;
  67     row[4] = (a3 - b3) >> ROW_SHIFT;
  68     row[5] = (a2 - b2) >> ROW_SHIFT;
  69     row[6] = (a1 - b1) >> ROW_SHIFT;
  70     row[7] = (a0 - b0) >> ROW_SHIFT;
  71 }
  72 #endif
  73
  74
  75 /* MMXEXT row IDCT */
  76
  77 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
  78                                                    c4,  c6,  c4,  c6,   \
  79                                                    c1,  c3, -c1, -c5,   \
  80                                                    c5,  c7,  c3, -c7,   \
  81                                                    c4, -c6,  c4, -c6,   \
  82                                                   -c4,  c2,  c4, -c2,   \
  83                                                    c5, -c1,  c3, -c1,   \
  84                                                    c7,  c3,  c7, -c5 }
  85
  86 static inline void mmxext_row_head (int16_t * row, int offset, const int16_t * table)
  87 {
  88     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
  89
  90     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
  91     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
  92
  93     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
  94     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
  95
  96     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
  97     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
  98
  99     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 100 }
 101
 102 static inline void mmxext_row (const int16_t * table, const int32_t * rounder)
 103 {
 104     movq_m2r (*(table+8), mm1);         /* mm1 = -C5 -C1 C3 C1 */
 105     pmaddwd_r2r (mm2, mm4);             /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
 106
 107     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
 108     pshufw_r2r (mm6, mm6, 0x4e);        /* mm6 = x3 x1 x7 x5 */
 109
 110     movq_m2r (*(table+12), mm7);        /* mm7 = -C7 C3 C7 C5 */
 111     pmaddwd_r2r (mm5, mm1);             /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
 112
 113     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 114     pmaddwd_r2r (mm6, mm7);             /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
 115
 116     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
 117     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 118
 119     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
 120     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 121
 122     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
 123     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 124
 125     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 126     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 127
 128     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 129     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 130
 131     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 132     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 133
 134     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 135     movq_r2r (mm0, mm4);                /* mm4 = a3 a2 + rounder */
 136
 137     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 138     psubd_r2r (mm5, mm4);               /* mm4 = a3-b3 a2-b2 + rounder */
 139 }
 140
 141 static inline void mmxext_row_tail (int16_t * row, int store)
 142 {
 143     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 144
 145     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 146
 147     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 148
 149     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 150
 151     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 152     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 153
 154     /* slot */
 155
 156     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 157 }
 158
 159 static inline void mmxext_row_mid (int16_t * row, int store,
 160                                    int offset, const int16_t * table)
 161 {
 162     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 163     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 164
 165     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 166     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
 167
 168     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 169     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 170
 171     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
 172     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 173
 174     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 175     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
 176
 177     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
 178     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
 179
 180     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
 181
 182     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
 183     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
 184 }
 185
 186
 187 /* MMX row IDCT */
 188
 189 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
 190                                            c4,  c6, -c4, -c2,   \
 191                                            c1,  c3,  c3, -c7,   \
 192                                            c5,  c7, -c1, -c5,   \
 193                                            c4, -c6,  c4, -c2,   \
 194                                           -c4,  c2,  c4, -c6,   \
 195                                            c5, -c1,  c7, -c5,   \
 196                                            c7,  c3,  c3, -c1 }
 197
 198 static inline void mmx_row_head (int16_t * row, int offset, const int16_t * table)
 199 {
 200     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 201
 202     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 203     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 204
 205     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 206     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 207
 208     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 209
 210     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 211     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 212
 213     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 214     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 215 }
 216
 217 static inline void mmx_row (const int16_t * table, const int32_t * rounder)
 218 {
 219     pmaddwd_r2r (mm2, mm4);             /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
 220     punpckldq_r2r (mm5, mm5);           /* mm5 = x3 x1 x3 x1 */
 221
 222     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
 223     punpckhdq_r2r (mm6, mm6);           /* mm6 = x7 x5 x7 x5 */
 224
 225     movq_m2r (*(table+12), mm7);        /* mm7 = -C5 -C1 C7 C5 */
 226     pmaddwd_r2r (mm5, mm1);             /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
 227
 228     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
 229     pmaddwd_r2r (mm6, mm7);             /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
 230
 231     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
 232     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
 233
 234     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
 235     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
 236
 237     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
 238     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
 239
 240     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
 241     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
 242
 243     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
 244     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
 245
 246     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
 247     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
 248
 249     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
 250     movq_r2r (mm0, mm7);                /* mm7 = a3 a2 + rounder */
 251
 252     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
 253     psubd_r2r (mm5, mm7);               /* mm7 = a3-b3 a2-b2 + rounder */
 254 }
 255
 256 static inline void mmx_row_tail (int16_t * row, int store)
 257 {
 258     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 259
 260     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 261
 262     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 263
 264     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 265
 266     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 267     movq_r2r (mm7, mm4);                /* mm4 = y6 y7 y4 y5 */
 268
 269     pslld_i2r (16, mm7);                /* mm7 = y7 0 y5 0 */
 270
 271     psrld_i2r (16, mm4);                /* mm4 = 0 y6 0 y4 */
 272
 273     por_r2r (mm4, mm7);                 /* mm7 = y7 y6 y5 y4 */
 274
 275     /* slot */
 276
 277     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 278 }
 279
 280 static inline void mmx_row_mid (int16_t * row, int store,
 281                                 int offset, const int16_t * table)
 282 {
 283     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
 284     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
 285
 286     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
 287     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
 288
 289     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
 290     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
 291
 292     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
 293     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
 294
 295     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
 296     movq_r2r (mm7, mm1);                /* mm1 = y6 y7 y4 y5 */
 297
 298     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
 299     psrld_i2r (16, mm7);                /* mm7 = 0 y6 0 y4 */
 300
 301     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
 302     pslld_i2r (16, mm1);                /* mm1 = y7 0 y5 0 */
 303
 304     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
 305     por_r2r (mm1, mm7);                 /* mm7 = y7 y6 y5 y4 */
 306
 307     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
 308     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
 309
 310     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
 311     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
 312 }
 313
 314
 315 #if 0
 316 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
 317 static inline void idct_col (int16_t * col, int offset)
 318 {
 319 /* multiplication - as implemented on mmx */
 320 #define F(c,x) (((c) * (x)) >> 16)
 321
 322 /* saturation - it helps us handle torture test cases */
 323 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
 324
 325     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
 326     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
 327     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
 328     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
 329
 330     col += offset;
 331
 332     x0 = col[0*8];
 333     x1 = col[1*8];
 334     x2 = col[2*8];
 335     x3 = col[3*8];
 336     x4 = col[4*8];
 337     x5 = col[5*8];
 338     x6 = col[6*8];
 339     x7 = col[7*8];
 340
 341     u04 = S (x0 + x4);
 342     v04 = S (x0 - x4);
 343     u26 = S (F (T2, x6) + x2);
 344     v26 = S (F (T2, x2) - x6);
 345
 346     a0 = S (u04 + u26);
 347     a1 = S (v04 + v26);
 348     a2 = S (v04 - v26);
 349     a3 = S (u04 - u26);
 350
 351     u17 = S (F (T1, x7) + x1);
 352     v17 = S (F (T1, x1) - x7);
 353     u35 = S (F (T3, x5) + x3);
 354     v35 = S (F (T3, x3) - x5);
 355
 356     b0 = S (u17 + u35);
 357     b3 = S (v17 - v35);
 358     u12 = S (u17 - u35);
 359     v12 = S (v17 + v35);
 360     u12 = S (2 * F (C4, u12));
 361     v12 = S (2 * F (C4, v12));
 362     b1 = S (u12 + v12);
 363     b2 = S (u12 - v12);
 364
 365     y0 = S (a0 + b0) >> COL_SHIFT;
 366     y1 = S (a1 + b1) >> COL_SHIFT;
 367     y2 = S (a2 + b2) >> COL_SHIFT;
 368     y3 = S (a3 + b3) >> COL_SHIFT;
 369
 370     y4 = S (a3 - b3) >> COL_SHIFT;
 371     y5 = S (a2 - b2) >> COL_SHIFT;
 372     y6 = S (a1 - b1) >> COL_SHIFT;
 373     y7 = S (a0 - b0) >> COL_SHIFT;
 374
 375     col[0*8] = y0;
 376     col[1*8] = y1;
 377     col[2*8] = y2;
 378     col[3*8] = y3;
 379     col[4*8] = y4;
 380     col[5*8] = y5;
 381     col[6*8] = y6;
 382     col[7*8] = y7;
 383 }
 384 #endif
 385
 386
 387 /* MMX column IDCT */
 388 static inline void idct_col (int16_t * col, int offset)
 389 {
 390 #define T1 13036
 391 #define T2 27146
 392 #define T3 43790
 393 #define C4 23170
 394
 395     static const short t1_vector[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
 396     static const short t2_vector[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
 397     static const short t3_vector[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
 398     static const short c4_vector[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
 399
 400     /* column code adapted from Peter Gubanov */
 401     /* http://www.elecard.com/peter/idct.shtml */
 402
 403     movq_m2r (*t1_vector, mm0);         /* mm0 = T1 */
 404
 405     movq_m2r (*(col+offset+1*8), mm1);  /* mm1 = x1 */
 406     movq_r2r (mm0, mm2);                /* mm2 = T1 */
 407
 408     movq_m2r (*(col+offset+7*8), mm4);  /* mm4 = x7 */
 409     pmulhw_r2r (mm1, mm0);              /* mm0 = T1*x1 */
 410
 411     movq_m2r (*t3_vector, mm5);         /* mm5 = T3 */
 412     pmulhw_r2r (mm4, mm2);              /* mm2 = T1*x7 */
 413
 414     movq_m2r (*(col+offset+5*8), mm6);  /* mm6 = x5 */
 415     movq_r2r (mm5, mm7);                /* mm7 = T3-1 */
 416
 417     movq_m2r (*(col+offset+3*8), mm3);  /* mm3 = x3 */
 418     psubsw_r2r (mm4, mm0);              /* mm0 = v17 */
 419
 420     movq_m2r (*t2_vector, mm4);         /* mm4 = T2 */
 421     pmulhw_r2r (mm3, mm5);              /* mm5 = (T3-1)*x3 */
 422
 423     paddsw_r2r (mm2, mm1);              /* mm1 = u17 */
 424     pmulhw_r2r (mm6, mm7);              /* mm7 = (T3-1)*x5 */
 425
 426     /* slot */
 427
 428     movq_r2r (mm4, mm2);                /* mm2 = T2 */
 429     paddsw_r2r (mm3, mm5);              /* mm5 = T3*x3 */
 430
 431     pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
 432     paddsw_r2r (mm6, mm7);              /* mm7 = T3*x5 */
 433
 434     psubsw_r2r (mm6, mm5);              /* mm5 = v35 */
 435     paddsw_r2r (mm3, mm7);              /* mm7 = u35 */
 436
 437     movq_m2r (*(col+offset+6*8), mm3);  /* mm3 = x6 */
 438     movq_r2r (mm0, mm6);                /* mm6 = v17 */
 439
 440     pmulhw_r2r (mm3, mm2);              /* mm2 = T2*x6 */
 441     psubsw_r2r (mm5, mm0);              /* mm0 = b3 */
 442
 443     psubsw_r2r (mm3, mm4);              /* mm4 = v26 */
 444     paddsw_r2r (mm6, mm5);              /* mm5 = v12 */
 445
 446     movq_r2m (mm0, *(col+offset+3*8));  /* save b3 in scratch0 */
 447     movq_r2r (mm1, mm6);                /* mm6 = u17 */
 448
 449     paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
 450     paddsw_r2r (mm7, mm6);              /* mm6 = b0 */
 451
 452     psubsw_r2r (mm7, mm1);              /* mm1 = u12 */
 453     movq_r2r (mm1, mm7);                /* mm7 = u12 */
 454
 455     movq_m2r (*(col+offset+0*8), mm3);  /* mm3 = x0 */
 456     paddsw_r2r (mm5, mm1);              /* mm1 = u12+v12 */
 457
 458     movq_m2r (*c4_vector, mm0);         /* mm0 = C4/2 */
 459     psubsw_r2r (mm5, mm7);              /* mm7 = u12-v12 */
 460
 461     movq_r2m (mm6, *(col+offset+5*8));  /* save b0 in scratch1 */
 462     pmulhw_r2r (mm0, mm1);              /* mm1 = b1/2 */
 463
 464     movq_r2r (mm4, mm6);                /* mm6 = v26 */
 465     pmulhw_r2r (mm0, mm7);              /* mm7 = b2/2 */
 466
 467     movq_m2r (*(col+offset+4*8), mm5);  /* mm5 = x4 */
 468     movq_r2r (mm3, mm0);                /* mm0 = x0 */
 469
 470     psubsw_r2r (mm5, mm3);              /* mm3 = v04 */
 471     paddsw_r2r (mm5, mm0);              /* mm0 = u04 */
 472
 473     paddsw_r2r (mm3, mm4);              /* mm4 = a1 */
 474     movq_r2r (mm0, mm5);                /* mm5 = u04 */
 475
 476     psubsw_r2r (mm6, mm3);              /* mm3 = a2 */
 477     paddsw_r2r (mm2, mm5);              /* mm5 = a0 */
 478
 479     paddsw_r2r (mm1, mm1);              /* mm1 = b1 */
 480     psubsw_r2r (mm2, mm0);              /* mm0 = a3 */
 481
 482     paddsw_r2r (mm7, mm7);              /* mm7 = b2 */
 483     movq_r2r (mm3, mm2);                /* mm2 = a2 */
 484
 485     movq_r2r (mm4, mm6);                /* mm6 = a1 */
 486     paddsw_r2r (mm7, mm3);              /* mm3 = a2+b2 */
 487
 488     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y2 */
 489     paddsw_r2r (mm1, mm4);              /* mm4 = a1+b1 */
 490
 491     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y1 */
 492     psubsw_r2r (mm1, mm6);              /* mm6 = a1-b1 */
 493
 494     movq_m2r (*(col+offset+5*8), mm1);  /* mm1 = b0 */
 495     psubsw_r2r (mm7, mm2);              /* mm2 = a2-b2 */
 496
 497     psraw_i2r (COL_SHIFT, mm6);         /* mm6 = y6 */
 498     movq_r2r (mm5, mm7);                /* mm7 = a0 */
 499
 500     movq_r2m (mm4, *(col+offset+1*8));  /* save y1 */
 501     psraw_i2r (COL_SHIFT, mm2);         /* mm2 = y5 */
 502
 503     movq_r2m (mm3, *(col+offset+2*8));  /* save y2 */
 504     paddsw_r2r (mm1, mm5);              /* mm5 = a0+b0 */
 505
 506     movq_m2r (*(col+offset+3*8), mm4);  /* mm4 = b3 */
 507     psubsw_r2r (mm1, mm7);              /* mm7 = a0-b0 */
 508
 509     psraw_i2r (COL_SHIFT, mm5);         /* mm5 = y0 */
 510     movq_r2r (mm0, mm3);                /* mm3 = a3 */
 511
 512     movq_r2m (mm2, *(col+offset+5*8));  /* save y5 */
 513     psubsw_r2r (mm4, mm3);              /* mm3 = a3-b3 */
 514
 515     psraw_i2r (COL_SHIFT, mm7);         /* mm7 = y7 */
 516     paddsw_r2r (mm0, mm4);              /* mm4 = a3+b3 */
 517
 518     movq_r2m (mm5, *(col+offset+0*8));  /* save y0 */
 519     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y4 */
 520
 521     movq_r2m (mm6, *(col+offset+6*8));  /* save y6 */
 522     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y3 */
 523
 524     movq_r2m (mm7, *(col+offset+7*8));  /* save y7 */
 525
 526     movq_r2m (mm3, *(col+offset+4*8));  /* save y4 */
 527
 528     movq_r2m (mm4, *(col+offset+3*8));  /* save y3 */
 529
 530 #undef T1
 531 #undef T2
 532 #undef T3
 533 #undef C4
 534 }
 535
 536 static const int32_t rounder0[] ATTR_ALIGN(8) =
 537     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
 538 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
 539 static const int32_t rounder1[] ATTR_ALIGN(8) =
 540     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
 541 static const int32_t rounder7[] ATTR_ALIGN(8) =
 542     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
 543 static const int32_t rounder2[] ATTR_ALIGN(8) =
 544     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
 545 static const int32_t rounder6[] ATTR_ALIGN(8) =
 546     rounder (-0.25);                /* C2 * (C6-C2)/2 */
 547 static const int32_t rounder3[] ATTR_ALIGN(8) =
 548     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
 549 static const int32_t rounder5[] ATTR_ALIGN(8) =
 550     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
 551
 552 #undef COL_SHIFT
 553 #undef ROW_SHIFT
 554
 555 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
 556 void idct (int16_t * block)                                             \
 557 {                                                                       \
 558     static const int16_t table04[] ATTR_ALIGN(16) =                     \
 559         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
 560     static const int16_t table17[] ATTR_ALIGN(16) =                     \
 561         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
 562     static const int16_t table26[] ATTR_ALIGN(16) =                     \
 563         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
 564     static const int16_t table35[] ATTR_ALIGN(16) =                     \
 565         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
 566                                                                         \
 567     idct_row_head (block, 0*8, table04);                                \
 568     idct_row (table04, rounder0);                                       \
 569     idct_row_mid (block, 0*8, 4*8, table04);                            \
 570     idct_row (table04, rounder4);                                       \
 571     idct_row_mid (block, 4*8, 1*8, table17);                            \
 572     idct_row (table17, rounder1);                                       \
 573     idct_row_mid (block, 1*8, 7*8, table17);                            \
 574     idct_row (table17, rounder7);                                       \
 575     idct_row_mid (block, 7*8, 2*8, table26);                            \
 576     idct_row (table26, rounder2);                                       \
 577     idct_row_mid (block, 2*8, 6*8, table26);                            \
 578     idct_row (table26, rounder6);                                       \
 579     idct_row_mid (block, 6*8, 3*8, table35);                            \
 580     idct_row (table35, rounder3);                                       \
 581     idct_row_mid (block, 3*8, 5*8, table35);                            \
 582     idct_row (table35, rounder5);                                       \
 583     idct_row_tail (block, 5*8);                                         \
 584                                                                         \
 585     idct_col (block, 0);                                                \
 586     idct_col (block, 4);                                                \
 587 }
 588
 589 void ff_mmx_idct(DCTELEM *block);
 590 void ff_mmxext_idct(DCTELEM *block);
 591
 592 declare_idct (ff_mmxext_idct, mmxext_table,
 593               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
 594
 595 declare_idct (ff_mmx_idct, mmx_table,
 596               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
 597