libavcodec/i386/fdct_mmx.c

   1 /*
   2  * MMX optimized forward DCT
   3  * The gcc porting is Copyright (c) 2001 Fabrice Bellard.
   4  * cleanup/optimizations are Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  * SSE2 optimization is Copyright (c) 2004 Denes Balatoni.
   6  *
   7  * from  fdctam32.c - AP922 MMX(3D-Now) forward-DCT
   8  *
   9  *  Intel Application Note AP-922 - fast, precise implementation of DCT
  10  *        http://developer.intel.com/vtune/cbts/appnotes.htm
  11  *
  12  * Also of inspiration:
  13  * a page about fdct at http://www.geocities.com/ssavekar/dct.htm
  14  * Skal's fdct at http://skal.planet-d.net/coding/dct.html
  15  *
  16  * This file is part of FFmpeg.
  17  *
  18  * FFmpeg is free software; you can redistribute it and/or
  19  * modify it under the terms of the GNU Lesser General Public
  20  * License as published by the Free Software Foundation; either
  21  * version 2.1 of the License, or (at your option) any later version.
  22  *
  23  * FFmpeg is distributed in the hope that it will be useful,
  24  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  25  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  26  * Lesser General Public License for more details.
  27  *
  28  * You should have received a copy of the GNU Lesser General Public
  29  * License along with FFmpeg; if not, write to the Free Software
  30  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  31  */
  32
  33 #include "libavutil/common.h"
  34 #include "libavcodec/dsputil.h"
  35 #include "mmx.h"
  36
  37 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
  38
  39 //////////////////////////////////////////////////////////////////////
  40 //
  41 // constants for the forward DCT
  42 // -----------------------------
  43 //
  44 // Be sure to check that your compiler is aligning all constants to QWORD
  45 // (8-byte) memory boundaries!  Otherwise the unaligned memory access will
  46 // severely stall MMX execution.
  47 //
  48 //////////////////////////////////////////////////////////////////////
  49
  50 #define BITS_FRW_ACC   3 //; 2 or 3 for accuracy
  51 #define SHIFT_FRW_COL  BITS_FRW_ACC
  52 #define SHIFT_FRW_ROW  (BITS_FRW_ACC + 17 - 3)
  53 #define RND_FRW_ROW    (1 << (SHIFT_FRW_ROW-1))
  54 //#define RND_FRW_COL    (1 << (SHIFT_FRW_COL-1))
  55
  56 #define X8(x) x,x,x,x,x,x,x,x
  57
  58 //concatenated table, for forward DCT transformation
  59 static const int16_t fdct_tg_all_16[24] ATTR_ALIGN(16) = {
  60     X8(13036),  // tg * (2<<16) + 0.5
  61     X8(27146),  // tg * (2<<16) + 0.5
  62     X8(-21746)  // tg * (2<<16) + 0.5
  63 };
  64
  65 static const int16_t ocos_4_16[8] ATTR_ALIGN(16) = {
  66     X8(23170)   //cos * (2<<15) + 0.5
  67 };
  68
  69 static const int16_t fdct_one_corr[8] ATTR_ALIGN(16) = { X8(1) };
  70
  71 static const int32_t fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW };
  72
  73 static struct
  74 {
  75  const int32_t fdct_r_row_sse2[4] ATTR_ALIGN(16);
  76 } fdct_r_row_sse2 ATTR_ALIGN(16)=
  77 {{
  78  RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
  79 }};
  80 //static const long fdct_r_row_sse2[4] ATTR_ALIGN(16) = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
  81
  82 static const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = {  // forward_dct coeff table
  83   16384,   16384,   22725,   19266,
  84   16384,   16384,   12873,    4520,
  85   21407,    8867,   19266,   -4520,
  86   -8867,  -21407,  -22725,  -12873,
  87   16384,  -16384,   12873,  -22725,
  88  -16384,   16384,    4520,   19266,
  89    8867,  -21407,    4520,  -12873,
  90   21407,   -8867,   19266,  -22725,
  91
  92   22725,   22725,   31521,   26722,
  93   22725,   22725,   17855,    6270,
  94   29692,   12299,   26722,   -6270,
  95  -12299,  -29692,  -31521,  -17855,
  96   22725,  -22725,   17855,  -31521,
  97  -22725,   22725,    6270,   26722,
  98   12299,  -29692,    6270,  -17855,
  99   29692,  -12299,   26722,  -31521,
 100
 101   21407,   21407,   29692,   25172,
 102   21407,   21407,   16819,    5906,
 103   27969,   11585,   25172,   -5906,
 104  -11585,  -27969,  -29692,  -16819,
 105   21407,  -21407,   16819,  -29692,
 106  -21407,   21407,    5906,   25172,
 107   11585,  -27969,    5906,  -16819,
 108   27969,  -11585,   25172,  -29692,
 109
 110   19266,   19266,   26722,   22654,
 111   19266,   19266,   15137,    5315,
 112   25172,   10426,   22654,   -5315,
 113  -10426,  -25172,  -26722,  -15137,
 114   19266,  -19266,   15137,  -26722,
 115  -19266,   19266,    5315,   22654,
 116   10426,  -25172,    5315,  -15137,
 117   25172,  -10426,   22654,  -26722,
 118
 119   16384,   16384,   22725,   19266,
 120   16384,   16384,   12873,    4520,
 121   21407,    8867,   19266,   -4520,
 122   -8867,  -21407,  -22725,  -12873,
 123   16384,  -16384,   12873,  -22725,
 124  -16384,   16384,    4520,   19266,
 125    8867,  -21407,    4520,  -12873,
 126   21407,   -8867,   19266,  -22725,
 127
 128   19266,   19266,   26722,   22654,
 129   19266,   19266,   15137,    5315,
 130   25172,   10426,   22654,   -5315,
 131  -10426,  -25172,  -26722,  -15137,
 132   19266,  -19266,   15137,  -26722,
 133  -19266,   19266,    5315,   22654,
 134   10426,  -25172,    5315,  -15137,
 135   25172,  -10426,   22654,  -26722,
 136
 137   21407,   21407,   29692,   25172,
 138   21407,   21407,   16819,    5906,
 139   27969,   11585,   25172,   -5906,
 140  -11585,  -27969,  -29692,  -16819,
 141   21407,  -21407,   16819,  -29692,
 142  -21407,   21407,    5906,   25172,
 143   11585,  -27969,    5906,  -16819,
 144   27969,  -11585,   25172,  -29692,
 145
 146   22725,   22725,   31521,   26722,
 147   22725,   22725,   17855,    6270,
 148   29692,   12299,   26722,   -6270,
 149  -12299,  -29692,  -31521,  -17855,
 150   22725,  -22725,   17855,  -31521,
 151  -22725,   22725,    6270,   26722,
 152   12299,  -29692,    6270,  -17855,
 153   29692,  -12299,   26722,  -31521,
 154 };
 155
 156 static struct
 157 {
 158  const int16_t tab_frw_01234567_sse2[256] ATTR_ALIGN(16);
 159 } tab_frw_01234567_sse2 ATTR_ALIGN(16) =
 160 {{
 161 //static const int16_t tab_frw_01234567_sse2[] ATTR_ALIGN(16) = {  // forward_dct coeff table
 162 #define TABLE_SSE2 C4,  C4,  C1,  C3, -C6, -C2, -C1, -C5, \
 163                    C4,  C4,  C5,  C7,  C2,  C6,  C3, -C7, \
 164                   -C4,  C4,  C7,  C3,  C6, -C2,  C7, -C5, \
 165                    C4, -C4,  C5, -C1,  C2, -C6,  C3, -C1,
 166 // c1..c7 * cos(pi/4) * 2^15
 167 #define C1 22725
 168 #define C2 21407
 169 #define C3 19266
 170 #define C4 16384
 171 #define C5 12873
 172 #define C6 8867
 173 #define C7 4520
 174 TABLE_SSE2
 175
 176 #undef C1
 177 #undef C2
 178 #undef C3
 179 #undef C4
 180 #undef C5
 181 #undef C6
 182 #undef C7
 183 #define C1 31521
 184 #define C2 29692
 185 #define C3 26722
 186 #define C4 22725
 187 #define C5 17855
 188 #define C6 12299
 189 #define C7 6270
 190 TABLE_SSE2
 191
 192 #undef C1
 193 #undef C2
 194 #undef C3
 195 #undef C4
 196 #undef C5
 197 #undef C6
 198 #undef C7
 199 #define C1 29692
 200 #define C2 27969
 201 #define C3 25172
 202 #define C4 21407
 203 #define C5 16819
 204 #define C6 11585
 205 #define C7 5906
 206 TABLE_SSE2
 207
 208 #undef C1
 209 #undef C2
 210 #undef C3
 211 #undef C4
 212 #undef C5
 213 #undef C6
 214 #undef C7
 215 #define C1 26722
 216 #define C2 25172
 217 #define C3 22654
 218 #define C4 19266
 219 #define C5 15137
 220 #define C6 10426
 221 #define C7 5315
 222 TABLE_SSE2
 223
 224 #undef C1
 225 #undef C2
 226 #undef C3
 227 #undef C4
 228 #undef C5
 229 #undef C6
 230 #undef C7
 231 #define C1 22725
 232 #define C2 21407
 233 #define C3 19266
 234 #define C4 16384
 235 #define C5 12873
 236 #define C6 8867
 237 #define C7 4520
 238 TABLE_SSE2
 239
 240 #undef C1
 241 #undef C2
 242 #undef C3
 243 #undef C4
 244 #undef C5
 245 #undef C6
 246 #undef C7
 247 #define C1 26722
 248 #define C2 25172
 249 #define C3 22654
 250 #define C4 19266
 251 #define C5 15137
 252 #define C6 10426
 253 #define C7 5315
 254 TABLE_SSE2
 255
 256 #undef C1
 257 #undef C2
 258 #undef C3
 259 #undef C4
 260 #undef C5
 261 #undef C6
 262 #undef C7
 263 #define C1 29692
 264 #define C2 27969
 265 #define C3 25172
 266 #define C4 21407
 267 #define C5 16819
 268 #define C6 11585
 269 #define C7 5906
 270 TABLE_SSE2
 271
 272 #undef C1
 273 #undef C2
 274 #undef C3
 275 #undef C4
 276 #undef C5
 277 #undef C6
 278 #undef C7
 279 #define C1 31521
 280 #define C2 29692
 281 #define C3 26722
 282 #define C4 22725
 283 #define C5 17855
 284 #define C6 12299
 285 #define C7 6270
 286 TABLE_SSE2
 287 }};
 288
 289 #define FDCT_COL(cpu, mm, mov)\
 290 static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\
 291 {\
 292     mov##_m2r(*(in + offset + 1 * 8), mm##0);\
 293     mov##_m2r(*(in + offset + 6 * 8), mm##1);\
 294     mov##_r2r(mm##0, mm##2);\
 295     mov##_m2r(*(in + offset + 2 * 8), mm##3);\
 296     paddsw_r2r(mm##1, mm##0);\
 297     mov##_m2r(*(in + offset + 5 * 8), mm##4);\
 298     psllw_i2r(SHIFT_FRW_COL, mm##0);\
 299     mov##_m2r(*(in + offset + 0 * 8), mm##5);\
 300     paddsw_r2r(mm##3, mm##4);\
 301     paddsw_m2r(*(in + offset + 7 * 8), mm##5);\
 302     psllw_i2r(SHIFT_FRW_COL, mm##4);\
 303     mov##_r2r(mm##0, mm##6);\
 304     psubsw_r2r(mm##1, mm##2);\
 305     mov##_m2r(*(fdct_tg_all_16 + 8), mm##1);\
 306     psubsw_r2r(mm##4, mm##0);\
 307     mov##_m2r(*(in + offset + 3 * 8), mm##7);\
 308     pmulhw_r2r(mm##0, mm##1);\
 309     paddsw_m2r(*(in + offset + 4 * 8), mm##7);\
 310     psllw_i2r(SHIFT_FRW_COL, mm##5);\
 311     paddsw_r2r(mm##4, mm##6);\
 312     psllw_i2r(SHIFT_FRW_COL, mm##7);\
 313     mov##_r2r(mm##5, mm##4);\
 314     psubsw_r2r(mm##7, mm##5);\
 315     paddsw_r2r(mm##5, mm##1);\
 316     paddsw_r2r(mm##7, mm##4);\
 317     por_m2r(*fdct_one_corr, mm##1);\
 318     psllw_i2r(SHIFT_FRW_COL + 1, mm##2);\
 319     pmulhw_m2r(*(fdct_tg_all_16 + 8), mm##5);\
 320     mov##_r2r(mm##4, mm##7);\
 321     psubsw_m2r(*(in + offset + 5 * 8), mm##3);\
 322     psubsw_r2r(mm##6, mm##4);\
 323     mov##_r2m(mm##1, *(out + offset + 2 * 8));\
 324     paddsw_r2r(mm##6, mm##7);\
 325     mov##_m2r(*(in + offset + 3 * 8), mm##1);\
 326     psllw_i2r(SHIFT_FRW_COL + 1, mm##3);\
 327     psubsw_m2r(*(in + offset + 4 * 8), mm##1);\
 328     mov##_r2r(mm##2, mm##6);\
 329     mov##_r2m(mm##4, *(out + offset + 4 * 8));\
 330     paddsw_r2r(mm##3, mm##2);\
 331     pmulhw_m2r(*ocos_4_16, mm##2);\
 332     psubsw_r2r(mm##3, mm##6);\
 333     pmulhw_m2r(*ocos_4_16, mm##6);\
 334     psubsw_r2r(mm##0, mm##5);\
 335     por_m2r(*fdct_one_corr, mm##5);\
 336     psllw_i2r(SHIFT_FRW_COL, mm##1);\
 337     por_m2r(*fdct_one_corr, mm##2);\
 338     mov##_r2r(mm##1, mm##4);\
 339     mov##_m2r(*(in + offset + 0 * 8), mm##3);\
 340     paddsw_r2r(mm##6, mm##1);\
 341     psubsw_m2r(*(in + offset + 7 * 8), mm##3);\
 342     psubsw_r2r(mm##6, mm##4);\
 343     mov##_m2r(*(fdct_tg_all_16 + 0), mm##0);\
 344     psllw_i2r(SHIFT_FRW_COL, mm##3);\
 345     mov##_m2r(*(fdct_tg_all_16 + 16), mm##6);\
 346     pmulhw_r2r(mm##1, mm##0);\
 347     mov##_r2m(mm##7, *(out + offset + 0 * 8));\
 348     pmulhw_r2r(mm##4, mm##6);\
 349     mov##_r2m(mm##5, *(out + offset + 6 * 8));\
 350     mov##_r2r(mm##3, mm##7);\
 351     mov##_m2r(*(fdct_tg_all_16 + 16), mm##5);\
 352     psubsw_r2r(mm##2, mm##7);\
 353     paddsw_r2r(mm##2, mm##3);\
 354     pmulhw_r2r(mm##7, mm##5);\
 355     paddsw_r2r(mm##3, mm##0);\
 356     paddsw_r2r(mm##4, mm##6);\
 357     pmulhw_m2r(*(fdct_tg_all_16 + 0), mm##3);\
 358     por_m2r(*fdct_one_corr, mm##0);\
 359     paddsw_r2r(mm##7, mm##5);\
 360     psubsw_r2r(mm##6, mm##7);\
 361     mov##_r2m(mm##0, *(out + offset + 1 * 8));\
 362     paddsw_r2r(mm##4, mm##5);\
 363     mov##_r2m(mm##7, *(out + offset + 3 * 8));\
 364     psubsw_r2r(mm##1, mm##3);\
 365     mov##_r2m(mm##5, *(out + offset + 5 * 8));\
 366     mov##_r2m(mm##3, *(out + offset + 7 * 8));\
 367 }
 368
 369 FDCT_COL(mmx, mm, movq)
 370 FDCT_COL(sse2, xmm, movdqa)
 371
 372 static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
 373 {
 374     asm volatile(
 375 #define FDCT_ROW_SSE2_H1(i,t)                    \
 376         "movq      " #i "(%0), %%xmm2      \n\t" \
 377         "movq      " #i "+8(%0), %%xmm0    \n\t" \
 378         "movdqa    " #t "+32(%1), %%xmm3   \n\t" \
 379         "movdqa    " #t "+48(%1), %%xmm7   \n\t" \
 380         "movdqa    " #t "(%1), %%xmm4      \n\t" \
 381         "movdqa    " #t "+16(%1), %%xmm5   \n\t"
 382
 383 #define FDCT_ROW_SSE2_H2(i,t)                    \
 384         "movq      " #i "(%0), %%xmm2      \n\t" \
 385         "movq      " #i "+8(%0), %%xmm0    \n\t" \
 386         "movdqa    " #t "+32(%1), %%xmm3   \n\t" \
 387         "movdqa    " #t "+48(%1), %%xmm7   \n\t"
 388
 389 #define FDCT_ROW_SSE2(i)                      \
 390         "movq      %%xmm2, %%xmm1       \n\t" \
 391         "pshuflw   $27, %%xmm0, %%xmm0  \n\t" \
 392         "paddsw    %%xmm0, %%xmm1       \n\t" \
 393         "psubsw    %%xmm0, %%xmm2       \n\t" \
 394         "punpckldq %%xmm2, %%xmm1       \n\t" \
 395         "pshufd    $78, %%xmm1, %%xmm2  \n\t" \
 396         "pmaddwd   %%xmm2, %%xmm3       \n\t" \
 397         "pmaddwd   %%xmm1, %%xmm7       \n\t" \
 398         "pmaddwd   %%xmm5, %%xmm2       \n\t" \
 399         "pmaddwd   %%xmm4, %%xmm1       \n\t" \
 400         "paddd     %%xmm7, %%xmm3       \n\t" \
 401         "paddd     %%xmm2, %%xmm1       \n\t" \
 402         "paddd     %%xmm6, %%xmm3       \n\t" \
 403         "paddd     %%xmm6, %%xmm1       \n\t" \
 404         "psrad     %3, %%xmm3           \n\t" \
 405         "psrad     %3, %%xmm1           \n\t" \
 406         "packssdw  %%xmm3, %%xmm1       \n\t" \
 407         "movdqa    %%xmm1, " #i "(%4)   \n\t"
 408
 409         "movdqa    (%2), %%xmm6         \n\t"
 410         FDCT_ROW_SSE2_H1(0,0)
 411         FDCT_ROW_SSE2(0)
 412         FDCT_ROW_SSE2_H2(64,0)
 413         FDCT_ROW_SSE2(64)
 414
 415         FDCT_ROW_SSE2_H1(16,64)
 416         FDCT_ROW_SSE2(16)
 417         FDCT_ROW_SSE2_H2(112,64)
 418         FDCT_ROW_SSE2(112)
 419
 420         FDCT_ROW_SSE2_H1(32,128)
 421         FDCT_ROW_SSE2(32)
 422         FDCT_ROW_SSE2_H2(96,128)
 423         FDCT_ROW_SSE2(96)
 424
 425         FDCT_ROW_SSE2_H1(48,192)
 426         FDCT_ROW_SSE2(48)
 427         FDCT_ROW_SSE2_H2(80,192)
 428         FDCT_ROW_SSE2(80)
 429         :
 430         : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2), "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out)
 431     );
 432 }
 433
 434 static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
 435 {
 436     pshufw_m2r(*(in + 4), mm5, 0x1B);
 437     movq_m2r(*(in + 0), mm0);
 438     movq_r2r(mm0, mm1);
 439     paddsw_r2r(mm5, mm0);
 440     psubsw_r2r(mm5, mm1);
 441     movq_r2r(mm0, mm2);
 442     punpckldq_r2r(mm1, mm0);
 443     punpckhdq_r2r(mm1, mm2);
 444     movq_m2r(*(table + 0), mm1);
 445     movq_m2r(*(table + 4), mm3);
 446     movq_m2r(*(table + 8), mm4);
 447     movq_m2r(*(table + 12), mm5);
 448     movq_m2r(*(table + 16), mm6);
 449     movq_m2r(*(table + 20), mm7);
 450     pmaddwd_r2r(mm0, mm1);
 451     pmaddwd_r2r(mm2, mm3);
 452     pmaddwd_r2r(mm0, mm4);
 453     pmaddwd_r2r(mm2, mm5);
 454     pmaddwd_r2r(mm0, mm6);
 455     pmaddwd_r2r(mm2, mm7);
 456     pmaddwd_m2r(*(table + 24), mm0);
 457     pmaddwd_m2r(*(table + 28), mm2);
 458     paddd_r2r(mm1, mm3);
 459     paddd_r2r(mm4, mm5);
 460     paddd_r2r(mm6, mm7);
 461     paddd_r2r(mm0, mm2);
 462     movq_m2r(*fdct_r_row, mm0);
 463     paddd_r2r(mm0, mm3);
 464     paddd_r2r(mm0, mm5);
 465     paddd_r2r(mm0, mm7);
 466     paddd_r2r(mm0, mm2);
 467     psrad_i2r(SHIFT_FRW_ROW, mm3);
 468     psrad_i2r(SHIFT_FRW_ROW, mm5);
 469     psrad_i2r(SHIFT_FRW_ROW, mm7);
 470     psrad_i2r(SHIFT_FRW_ROW, mm2);
 471     packssdw_r2r(mm5, mm3);
 472     packssdw_r2r(mm2, mm7);
 473     movq_r2m(mm3, *(out + 0));
 474     movq_r2m(mm7, *(out + 4));
 475 }
 476
 477 static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
 478 {
 479 //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
 480     movd_m2r(*(in + 6), mm1);
 481     punpcklwd_m2r(*(in + 4), mm1);
 482     movq_r2r(mm1, mm2);
 483     psrlq_i2r(0x20, mm1);
 484     movq_m2r(*(in + 0), mm0);
 485     punpcklwd_r2r(mm2, mm1);
 486     movq_r2r(mm0, mm5);
 487     paddsw_r2r(mm1, mm0);
 488     psubsw_r2r(mm1, mm5);
 489     movq_r2r(mm0, mm2);
 490     punpckldq_r2r(mm5, mm0);
 491     punpckhdq_r2r(mm5, mm2);
 492     movq_m2r(*(table + 0), mm1);
 493     movq_m2r(*(table + 4), mm3);
 494     movq_m2r(*(table + 8), mm4);
 495     movq_m2r(*(table + 12), mm5);
 496     movq_m2r(*(table + 16), mm6);
 497     movq_m2r(*(table + 20), mm7);
 498     pmaddwd_r2r(mm0, mm1);
 499     pmaddwd_r2r(mm2, mm3);
 500     pmaddwd_r2r(mm0, mm4);
 501     pmaddwd_r2r(mm2, mm5);
 502     pmaddwd_r2r(mm0, mm6);
 503     pmaddwd_r2r(mm2, mm7);
 504     pmaddwd_m2r(*(table + 24), mm0);
 505     pmaddwd_m2r(*(table + 28), mm2);
 506     paddd_r2r(mm1, mm3);
 507     paddd_r2r(mm4, mm5);
 508     paddd_r2r(mm6, mm7);
 509     paddd_r2r(mm0, mm2);
 510     movq_m2r(*fdct_r_row, mm0);
 511     paddd_r2r(mm0, mm3);
 512     paddd_r2r(mm0, mm5);
 513     paddd_r2r(mm0, mm7);
 514     paddd_r2r(mm0, mm2);
 515     psrad_i2r(SHIFT_FRW_ROW, mm3);
 516     psrad_i2r(SHIFT_FRW_ROW, mm5);
 517     psrad_i2r(SHIFT_FRW_ROW, mm7);
 518     psrad_i2r(SHIFT_FRW_ROW, mm2);
 519     packssdw_r2r(mm5, mm3);
 520     packssdw_r2r(mm2, mm7);
 521     movq_r2m(mm3, *(out + 0));
 522     movq_r2m(mm7, *(out + 4));
 523 }
 524
 525 void ff_fdct_mmx(int16_t *block)
 526 {
 527     int64_t align_tmp[16] ATTR_ALIGN(8);
 528     int16_t * block1= (int16_t*)align_tmp;
 529     const int16_t *table= tab_frw_01234567;
 530     int i;
 531
 532     fdct_col_mmx(block, block1, 0);
 533     fdct_col_mmx(block, block1, 4);
 534
 535     for(i=8;i>0;i--) {
 536         fdct_row_mmx(block1, block, table);
 537         block1 += 8;
 538         table += 32;
 539         block += 8;
 540     }
 541 }
 542
 543 void ff_fdct_mmx2(int16_t *block)
 544 {
 545     int64_t align_tmp[16] ATTR_ALIGN(8);
 546     int16_t *block1= (int16_t*)align_tmp;
 547     const int16_t *table= tab_frw_01234567;
 548     int i;
 549
 550     fdct_col_mmx(block, block1, 0);
 551     fdct_col_mmx(block, block1, 4);
 552
 553     for(i=8;i>0;i--) {
 554         fdct_row_mmx2(block1, block, table);
 555         block1 += 8;
 556         table += 32;
 557         block += 8;
 558     }
 559 }
 560
 561 void ff_fdct_sse2(int16_t *block)
 562 {
 563     int64_t align_tmp[16] ATTR_ALIGN(16);
 564     int16_t * const block1= (int16_t*)align_tmp;
 565
 566     fdct_col_sse2(block, block1, 0);
 567     fdct_row_sse2(block1, block);
 568 }
 569