git.sesse.net Git - ffmpeg/blob - libavcodec/mips/simple_idct_mmi.c

   1 /*
   2  * Loongson SIMD optimized simple idct
   3  *
   4  * Copyright (c) 2015 Loongson Technology Corporation Limited
   5  * Copyright (c) 2015 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
   6  *                    Zhang Shuangshuang <zhangshuangshuang@ict.ac.cn>
   7  *
   8  * This file is part of FFmpeg.
   9  *
  10  * FFmpeg is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * FFmpeg is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with FFmpeg; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 #include "idctdsp_mips.h"
  26 #include "constants.h"
  27 #include "libavutil/mips/asmdefs.h"
  28 #include "libavutil/mips/mmiutils.h"
  29
  30 #define W1  22725  //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  31 #define W2  21407  //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  32 #define W3  19266  //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  33 #define W4  16383  //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  34 #define W5  12873  //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  35 #define W6  8867   //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  36 #define W7  4520   //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
  37
  38 #define ROW_SHIFT 11
  39 #define COL_SHIFT 20
  40 #define DC_SHIFT 3
  41
  42 DECLARE_ALIGNED(8, const int16_t, W_arr)[46] = {
  43     W4,  W2,  W4,  W6,
  44     W1,  W3,  W5,  W7,
  45     W4,  W6, -W4, -W2,
  46     W3, -W7, -W1, -W5,
  47     W4, -W6, -W4,  W2,
  48     W5, -W1,  W7,  W3,
  49     W4, -W2,  W4, -W6,
  50     W7, -W5,  W3, -W1,
  51     1024, 0,  1024, 0, //ff_p32_1024 = 0x0000040000000400ULL
  52     0,   -1,  -1,  -1, //mask = 0xffffffffffff0000ULL
  53     32,  32,  32,  32  //ff_p16_32 = 0x0020002000200020ULL
  54 };
  55
  56 void ff_simple_idct_8_mmi(int16_t *block)
  57 {
  58     BACKUP_REG
  59     __asm__ volatile (
  60
  61 #define IDCT_ROW_COND_DC(src1, src2)                                  \
  62         "dmfc1        $11,      "#src1"                         \n\t" \
  63         "dmfc1        $12,      "#src2"                         \n\t" \
  64         "and          $11,       $11,       $9                  \n\t" \
  65         "or           $10,       $11,       $12                 \n\t" \
  66         "beqz         $10,       1f                             \n\t" \
  67                                                                       \
  68         "punpcklhw    $f30,     "#src1",   "#src2"              \n\t" \
  69         "punpckhhw    $f31,     "#src1",   "#src2"              \n\t" \
  70         /* s6, s4, s2, s0 */                                          \
  71         "punpcklhw   "#src1",    $f30,      $f31                \n\t" \
  72         /* s7, s5, s3, s1 */                                          \
  73         "punpckhhw   "#src2",    $f30,      $f31                \n\t" \
  74                                                                       \
  75         "pmaddhw      $f30,     "#src1",    $f18                \n\t" \
  76         "pmaddhw      $f31,     "#src2",    $f19                \n\t" \
  77         "paddw        $f28,      $f30,      $f31                \n\t" \
  78         "psubw        $f29,      $f30,      $f31                \n\t" \
  79         "punpcklwd    $f30,      $f28,      $f29                \n\t" \
  80         "punpckhwd    $f31,      $f28,      $f29                \n\t" \
  81         "paddw        $f26,      $f30,      $f31                \n\t" \
  82         "paddw        $f26,      $f26,      $f16                \n\t" \
  83         /* $f26: src[7], src[0] */                                    \
  84         "psraw        $f26,      $f26,      $f17                \n\t" \
  85                                                                       \
  86         "pmaddhw      $f30,     "#src1",    $f20                \n\t" \
  87         "pmaddhw      $f31,     "#src2",    $f21                \n\t" \
  88         "paddw        $f28,      $f30,      $f31                \n\t" \
  89         "psubw        $f29,      $f30,      $f31                \n\t" \
  90         "punpcklwd    $f30,      $f28,      $f29                \n\t" \
  91         "punpckhwd    $f31,      $f28,      $f29                \n\t" \
  92         "paddw        $f27,      $f30,      $f31                \n\t" \
  93         "paddw        $f27,      $f27,      $f16                \n\t" \
  94         /* $f27: src[6], src[1] */                                    \
  95         "psraw        $f27,      $f27,      $f17                \n\t" \
  96                                                                       \
  97         "pmaddhw      $f30,     "#src1",    $f22                \n\t" \
  98         "pmaddhw      $f31,     "#src2",    $f23                \n\t" \
  99         "paddw        $f28,      $f30,      $f31                \n\t" \
 100         "psubw        $f29,      $f30,      $f31                \n\t" \
 101         "punpcklwd    $f30,      $f28,      $f29                \n\t" \
 102         "punpckhwd    $f31,      $f28,      $f29                \n\t" \
 103         "paddw        $f28,      $f30,      $f31                \n\t" \
 104         "paddw        $f28,      $f28,      $f16                \n\t" \
 105         /* $f28: src[5], src[2] */                                    \
 106         "psraw        $f28,      $f28,      $f17                \n\t" \
 107                                                                       \
 108         "pmaddhw      $f30,     "#src1",    $f24                \n\t" \
 109         "pmaddhw      $f31,     "#src2",    $f25                \n\t" \
 110         "paddw       "#src1",    $f30,      $f31                \n\t" \
 111         "psubw       "#src2",    $f30,      $f31                \n\t" \
 112         "punpcklwd    $f30,     "#src1",   "#src2"              \n\t" \
 113         "punpckhwd    $f31,     "#src1",   "#src2"              \n\t" \
 114         "paddw        $f29,      $f30,      $f31                \n\t" \
 115         "paddw        $f29,      $f29,      $f16                \n\t" \
 116         /* $f29: src[4], src[3] */                                    \
 117         "psraw        $f29,      $f29,      $f17                \n\t" \
 118                                                                       \
 119         "punpcklhw   "#src1",    $f26,      $f27                \n\t" \
 120         "punpckhhw    $f30,      $f27,      $f26                \n\t" \
 121         "punpcklhw    $f31,      $f28,      $f29                \n\t" \
 122         "punpckhhw   "#src2",    $f29,      $f28                \n\t" \
 123         /* src[3], src[2], src[1], src[0] */                          \
 124         "punpcklwd   "#src1",   "#src1",    $f31                \n\t" \
 125         /* src[7], src[6], src[5], src[4] */                          \
 126         "punpcklwd   "#src2",   "#src2",    $f30                \n\t" \
 127         "j                       2f                             \n\t" \
 128                                                                       \
 129         "1:                                                     \n\t" \
 130         "li           $10,       3                              \n\t" \
 131         "dmtc1        $10,       $f30                           \n\t" \
 132         "psllh        $f28,     "#src1",    $f30                \n\t" \
 133         "dmtc1        $9,        $f31                           \n\t" \
 134         "punpcklhw    $f29,      $f28,      $f28                \n\t" \
 135         "and          $f29,      $f29,      $f31                \n\t" \
 136         "paddw        $f28,      $f28,      $f29                \n\t" \
 137         "punpcklwd   "#src1",    $f28,      $f28                \n\t" \
 138         "punpcklwd   "#src2",    $f28,      $f28                \n\t" \
 139         "2:                                                     \n\t" \
 140
 141         /* idctRowCondDC row0~8 */
 142
 143         /* load W */
 144         "gslqc1       $f19,      $f18,      0x00(%[w_arr])      \n\t"
 145         "gslqc1       $f21,      $f20,      0x10(%[w_arr])      \n\t"
 146         "gslqc1       $f23,      $f22,      0x20(%[w_arr])      \n\t"
 147         "gslqc1       $f25,      $f24,      0x30(%[w_arr])      \n\t"
 148         "gslqc1       $f17,      $f16,      0x40(%[w_arr])      \n\t"
 149         /* load source in block */
 150         "gslqc1       $f1,       $f0,       0x00(%[block])      \n\t"
 151         "gslqc1       $f3,       $f2,       0x10(%[block])      \n\t"
 152         "gslqc1       $f5,       $f4,       0x20(%[block])      \n\t"
 153         "gslqc1       $f7,       $f6,       0x30(%[block])      \n\t"
 154         "gslqc1       $f9,       $f8,       0x40(%[block])      \n\t"
 155         "gslqc1       $f11,      $f10,      0x50(%[block])      \n\t"
 156         "gslqc1       $f13,      $f12,      0x60(%[block])      \n\t"
 157         "gslqc1       $f15,      $f14,      0x70(%[block])      \n\t"
 158
 159         /* $9: mask ; $f17: ROW_SHIFT */
 160         "dmfc1        $9,        $f17                           \n\t"
 161         "li           $10,       11                             \n\t"
 162         "mtc1         $10,       $f17                           \n\t"
 163         IDCT_ROW_COND_DC($f0,$f1)
 164         IDCT_ROW_COND_DC($f2,$f3)
 165         IDCT_ROW_COND_DC($f4,$f5)
 166         IDCT_ROW_COND_DC($f6,$f7)
 167         IDCT_ROW_COND_DC($f8,$f9)
 168         IDCT_ROW_COND_DC($f10,$f11)
 169         IDCT_ROW_COND_DC($f12,$f13)
 170         IDCT_ROW_COND_DC($f14,$f15)
 171
 172 #define IDCT_COL_CASE1(src, out1, out2)                               \
 173         "pmaddhw      $f26,     "#src",     $f18                \n\t" \
 174         "pmaddhw      $f27,     "#src",     $f20                \n\t" \
 175         "pmaddhw      $f28,     "#src",     $f22                \n\t" \
 176         "pmaddhw      $f29,     "#src",     $f24                \n\t" \
 177                                                                       \
 178         "punpcklwd    $f30,      $f26,      $f26                \n\t" \
 179         "punpckhwd    $f31,      $f26,      $f26                \n\t" \
 180         /* $f26: src[0], src[56] */                                   \
 181         "paddw        $f26,      $f30,      $f31                \n\t" \
 182         "punpcklwd    $f30,      $f27,      $f27                \n\t" \
 183         "punpckhwd    $f31,      $f27,      $f27                \n\t" \
 184         /* $f27: src[8], src[48] */                                   \
 185         "paddw        $f27,      $f30,      $f31                \n\t" \
 186         "punpcklwd    $f30,      $f28,      $f28                \n\t" \
 187         "punpckhwd    $f31,      $f28,      $f28                \n\t" \
 188         /* $f28: src[16], src[40] */                                  \
 189         "paddw        $f28,      $f30,      $f31                \n\t" \
 190         "punpcklwd    $f30,      $f29,      $f29                \n\t" \
 191         "punpckhwd    $f31,      $f29,      $f29                \n\t" \
 192         /* $f29: src[24], src[32] */                                  \
 193         "paddw        $f29,      $f30,      $f31                \n\t" \
 194                                                                       \
 195         /* out1: src[24], src[16], src[8], src[0] */                  \
 196         /* out2: src[56], src[48], src[40], src[32] */                \
 197         "punpcklhw    $f30,      $f26,      $f27                \n\t" \
 198         "punpcklhw    $f31,      $f28,      $f29                \n\t" \
 199         "punpckhwd   "#out1",    $f30,      $f31                \n\t" \
 200         "psrah       "#out1",   "#out1",    $f16                \n\t" \
 201         "punpcklhw    $f30,      $f27,      $f26                \n\t" \
 202         "punpcklhw    $f31,      $f29,      $f28                \n\t" \
 203         "punpckhwd   "#out2",    $f31,      $f30                \n\t" \
 204         "psrah       "#out2",   "#out2",    $f16                \n\t"
 205
 206 #define IDCT_COL_CASE2(src1, src2, out1, out2)                        \
 207         "pmaddhw      $f28,     "#src1",    $f18                \n\t" \
 208         "pmaddhw      $f29,     "#src2",    $f19                \n\t" \
 209         "paddw        $f30,      $f28,      $f29                \n\t" \
 210         "psubw        $f31,      $f28,      $f29                \n\t" \
 211         "punpcklwd    $f28,      $f30,      $f31                \n\t" \
 212         "punpckhwd    $f29,      $f30,      $f31                \n\t" \
 213         "pmaddhw      $f30,     "#src1",    $f20                \n\t" \
 214         "pmaddhw      $f31,     "#src2",    $f21                \n\t" \
 215         /* $f26: src[0], src[56] */                                   \
 216         "paddw        $f26,      $f28,      $f29                \n\t" \
 217         "paddw        $f28,      $f30,      $f31                \n\t" \
 218         "psubw        $f29,      $f30,      $f31                \n\t" \
 219         "punpcklwd    $f30,      $f28,      $f29                \n\t" \
 220         "punpckhwd    $f31,      $f28,      $f29                \n\t" \
 221         "pmaddhw      $f28,     "#src1",    $f22                \n\t" \
 222         "pmaddhw      $f29,     "#src2",    $f23                \n\t" \
 223         /* $f27: src[8], src[48] */                                   \
 224         "paddw        $f27,      $f30,      $f31                \n\t" \
 225         "paddw        $f30,      $f28,      $f29                \n\t" \
 226         "psubw        $f31,      $f28,      $f29                \n\t" \
 227         "punpcklwd    $f28,      $f30,      $f31                \n\t" \
 228         "punpckhwd    $f29,      $f30,      $f31                \n\t" \
 229         "pmaddhw      $f30,     "#src1",    $f24                \n\t" \
 230         "pmaddhw      $f31,     "#src2",    $f25                \n\t" \
 231         /* $f28: src[16], src[40] */                                  \
 232         "paddw        $f28,      $f28,      $f29                \n\t" \
 233         "paddw       "#out1",    $f30,      $f31                \n\t" \
 234         "psubw       "#out2",    $f30,      $f31                \n\t" \
 235         "punpcklwd    $f30,     "#out1",   "#out2"              \n\t" \
 236         "punpckhwd    $f31,     "#out1",   "#out2"              \n\t" \
 237         /* $f29: src[24], src[32] */                                  \
 238         "paddw        $f29,      $f30,      $f31                \n\t" \
 239                                                                       \
 240         /* out1: src[24], src[16], src[8], src[0] */                  \
 241         /* out2: src[56], src[48], src[40], src[32] */                \
 242         "punpcklhw   "#out1",    $f26,      $f27                \n\t" \
 243         "punpckhhw   "#out2",    $f27,      $f26                \n\t" \
 244         "punpcklhw    $f30,      $f28,      $f29                \n\t" \
 245         "punpckhhw    $f31,      $f29,      $f28                \n\t" \
 246         "punpckhwd   "#out1",   "#out1",    $f30                \n\t" \
 247         "punpckhwd   "#out2",    $f31,     "#out2"              \n\t" \
 248         "psrah       "#out1",   "#out1",    $f16                \n\t" \
 249         "psrah       "#out2",   "#out2",    $f16                \n\t"
 250
 251
 252         /* idctSparseCol col0~3 */
 253
 254         /* $f17: ff_p16_32; $f16: COL_SHIFT-16 */
 255         "gsldlc1      $f17,      0x57(%[w_arr])                 \n\t"
 256         "gsldrc1      $f17,      0x50(%[w_arr])                 \n\t"
 257         "li           $10,       4                              \n\t"
 258         "dmtc1        $10,       $f16                           \n\t"
 259         "paddh        $f0,       $f0,       $f17                \n\t"
 260         /* Transpose row[0,2,4,6] */
 261         "punpcklhw    $f26,      $f0,       $f4                 \n\t"
 262         "punpckhhw    $f27,      $f0,       $f4                 \n\t"
 263         "punpcklhw    $f28,      $f8,       $f12                \n\t"
 264         "punpckhhw    $f29,      $f8,       $f12                \n\t"
 265         "punpcklwd    $f0,       $f26,      $f28                \n\t"
 266         "punpckhwd    $f4,       $f26,      $f28                \n\t"
 267         "punpcklwd    $f8,       $f27,      $f29                \n\t"
 268         "punpckhwd    $f12,      $f27,      $f29                \n\t"
 269
 270         "or           $f26,      $f2,       $f6                 \n\t"
 271         "or           $f26,      $f26,      $f10                \n\t"
 272         "or           $f26,      $f26,      $f14                \n\t"
 273         "dmfc1        $10,       $f26                           \n\t"
 274         "bnez         $10,       1f                             \n\t"
 275         /* case1: In this case, row[1,3,5,7] are all zero */
 276         /* col0: $f0: col[24,16,8,0]; $f2: col[56,48,40,32] */
 277         IDCT_COL_CASE1($f0, $f0, $f2)
 278         /* col1: $f4: col[25,17,9,1]; $f6: col[57,49,41,33] */
 279         IDCT_COL_CASE1($f4, $f4, $f6)
 280         /* col2: $f8: col[26,18,10,2]; $f10: col[58,50,42,34] */
 281         IDCT_COL_CASE1($f8, $f8, $f10)
 282         /* col3: $f12: col[27,19,11,3]; $f14: col[59,51,43,35] */
 283         IDCT_COL_CASE1($f12, $f12, $f14)
 284         "j                                  2f                  \n\t"
 285
 286         "1:                                                     \n\t"
 287         /* case2: row[1,3,5,7] are not all zero */
 288         /* Transpose */
 289         "punpcklhw    $f26,      $f2,       $f6                 \n\t"
 290         "punpckhhw    $f27,      $f2,       $f6                 \n\t"
 291         "punpcklhw    $f28,      $f10,      $f14                \n\t"
 292         "punpckhhw    $f29,      $f10,      $f14                \n\t"
 293         "punpcklwd    $f2,       $f26,      $f28                \n\t"
 294         "punpckhwd    $f6,       $f26,      $f28                \n\t"
 295         "punpcklwd    $f10,      $f27,      $f29                \n\t"
 296         "punpckhwd    $f14,      $f27,      $f29                \n\t"
 297
 298         /* col0: $f0: col[24,16,8,0]; $f2: col[56,48,40,32] */
 299         IDCT_COL_CASE2($f0, $f2, $f0, $f2)
 300         /* col1: $f4: col[25,17,9,1]; $f6: col[57,49,41,33] */
 301         IDCT_COL_CASE2($f4, $f6, $f4, $f6)
 302         /* col2: $f8: col[26,18,10,2]; $f10: col[58,50,42,34] */
 303         IDCT_COL_CASE2($f8, $f10, $f8, $f10)
 304         /* col3: $f12: col[27,19,11,3]; $f14: col[59,51,43,35] */
 305         IDCT_COL_CASE2($f12, $f14, $f12, $f14)
 306
 307         "2:                                                     \n\t"
 308         /* Transpose */
 309         "punpcklhw    $f26,      $f0,       $f4                 \n\t"
 310         "punpckhhw    $f27,      $f0,       $f4                 \n\t"
 311         "punpcklhw    $f28,      $f8,       $f12                \n\t"
 312         "punpckhhw    $f29,      $f8,       $f12                \n\t"
 313         "punpcklwd    $f0,       $f26,      $f28                \n\t"
 314         "punpckhwd    $f4,       $f26,      $f28                \n\t"
 315         "punpcklwd    $f8,       $f27,      $f29                \n\t"
 316         "punpckhwd    $f12,      $f27,      $f29                \n\t"
 317         /* Transpose */
 318         "punpcklhw    $f26,      $f2,       $f6                 \n\t"
 319         "punpckhhw    $f27,      $f2,       $f6                 \n\t"
 320         "punpcklhw    $f28,      $f10,      $f14                \n\t"
 321         "punpckhhw    $f29,      $f10,      $f14                \n\t"
 322         "punpcklwd    $f2,       $f26,      $f28                \n\t"
 323         "punpckhwd    $f6,       $f26,      $f28                \n\t"
 324         "punpcklwd    $f10,      $f27,      $f29                \n\t"
 325         "punpckhwd    $f14,      $f27,      $f29                \n\t"
 326
 327         /* idctSparseCol col4~7 */
 328
 329         "paddh        $f1,       $f1,       $f17                \n\t"
 330         /* Transpose */
 331         "punpcklhw    $f26,      $f1,       $f5                 \n\t"
 332         "punpckhhw    $f27,      $f1,       $f5                 \n\t"
 333         "punpcklhw    $f28,      $f9,       $f13                \n\t"
 334         "punpckhhw    $f29,      $f9,       $f13                \n\t"
 335         "punpcklwd    $f1,       $f26,      $f28                \n\t"
 336         "punpckhwd    $f5,       $f26,      $f28                \n\t"
 337         "punpcklwd    $f9,       $f27,      $f29                \n\t"
 338         "punpckhwd    $f13,      $f27,      $f29                \n\t"
 339
 340         "or           $f26,      $f3,       $f7                 \n\t"
 341         "or           $f26,      $f26,      $f11                \n\t"
 342         "or           $f26,      $f26,      $f15                \n\t"
 343         "dmfc1        $10,       $f26                           \n\t"
 344         "bnez         $10,       1f                             \n\t"
 345         /* case1: In this case, row[1,3,5,7] are all zero */
 346         /* col4: $f1: col[24,16,8,0]; $f3: col[56,48,40,32] */
 347         IDCT_COL_CASE1($f1, $f1, $f3)
 348         /* col5: $f5: col[25,17,9,1]; $f7: col[57,49,41,33] */
 349         IDCT_COL_CASE1($f5, $f5, $f7)
 350         /* col6: $f9: col[26,18,10,2]; $f11: col[58,50,42,34] */
 351         IDCT_COL_CASE1($f9, $f9, $f11)
 352         /* col7: $f13: col[27,19,11,3]; $f15: col[59,51,43,35] */
 353         IDCT_COL_CASE1($f13, $f13, $f15)
 354         "j                                  2f                  \n\t"
 355
 356         "1:                                                     \n\t"
 357         /* case2: row[1,3,5,7] are not all zero */
 358         /* Transpose */
 359         "punpcklhw    $f26,      $f3,       $f7                 \n\t"
 360         "punpckhhw    $f27,      $f3,       $f7                 \n\t"
 361         "punpcklhw    $f28,      $f11,      $f15                \n\t"
 362         "punpckhhw    $f29,      $f11,      $f15                \n\t"
 363         "punpcklwd    $f3,       $f26,      $f28                \n\t"
 364         "punpckhwd    $f7,       $f26,      $f28                \n\t"
 365         "punpcklwd    $f11,      $f27,      $f29                \n\t"
 366         "punpckhwd    $f15,      $f27,      $f29                \n\t"
 367
 368         /* col4: $f1: col[24,16,8,0]; $f3: col[56,48,40,32] */
 369         IDCT_COL_CASE2($f1, $f3, $f1, $f3)
 370         /* col5: $f5: col[25,17,9,1]; $f7: col[57,49,41,33] */
 371         IDCT_COL_CASE2($f5, $f7, $f5, $f7)
 372         /* col6: $f9: col[26,18,10,2]; $f11: col[58,50,42,34] */
 373         IDCT_COL_CASE2($f9, $f11, $f9, $f11)
 374         /* col7: $f13: col[27,19,11,3]; $f15: col[59,51,43,35] */
 375         IDCT_COL_CASE2($f13, $f15, $f13, $f15)
 376
 377         "2:                                                     \n\t"
 378         /* Transpose */
 379         "punpcklhw    $f26,      $f1,       $f5                 \n\t"
 380         "punpckhhw    $f27,      $f1,       $f5                 \n\t"
 381         "punpcklhw    $f28,      $f9,       $f13                \n\t"
 382         "punpckhhw    $f29,      $f9,       $f13                \n\t"
 383         "punpcklwd    $f1,       $f26,      $f28                \n\t"
 384         "punpckhwd    $f5,       $f26,      $f28                \n\t"
 385         "punpcklwd    $f9,       $f27,      $f29                \n\t"
 386         "punpckhwd    $f13,      $f27,      $f29                \n\t"
 387         /* Transpose */
 388         "punpcklhw    $f26,      $f3,       $f7                 \n\t"
 389         "punpckhhw    $f27,      $f3,       $f7                 \n\t"
 390         "punpcklhw    $f28,      $f11,      $f15                \n\t"
 391         "punpckhhw    $f29,      $f11,      $f15                \n\t"
 392         "punpcklwd    $f3,       $f26,      $f28                \n\t"
 393         "punpckhwd    $f7,       $f26,      $f28                \n\t"
 394         "punpcklwd    $f11,      $f27,      $f29                \n\t"
 395         "punpckhwd    $f15,      $f27,      $f29                \n\t"
 396         /* Store */
 397         "gssqc1       $f1,       $f0,       0x00(%[block])      \n\t"
 398         "gssqc1       $f5,       $f4,       0x10(%[block])      \n\t"
 399         "gssqc1       $f9,       $f8,       0x20(%[block])      \n\t"
 400         "gssqc1       $f13,      $f12,      0x30(%[block])      \n\t"
 401         "gssqc1       $f3,       $f2,       0x40(%[block])      \n\t"
 402         "gssqc1       $f7,       $f6,       0x50(%[block])      \n\t"
 403         "gssqc1       $f11,      $f10,      0x60(%[block])      \n\t"
 404         "gssqc1       $f15,      $f14,      0x70(%[block])      \n\t"
 405
 406         : [block]"+&r"(block)
 407         : [w_arr]"r"(W_arr)
 408         : "memory"
 409     );
 410
 411     RECOVER_REG
 412 }
 413
 414 void ff_simple_idct_put_8_mmi(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
 415 {
 416     ff_simple_idct_8_mmi(block);
 417     ff_put_pixels_clamped_mmi(block, dest, line_size);
 418 }
 419 void ff_simple_idct_add_8_mmi(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
 420 {
 421     ff_simple_idct_8_mmi(block);
 422     ff_add_pixels_clamped_mmi(block, dest, line_size);
 423 }