git.sesse.net Git - ffmpeg/blob - libavcodec/mips/h264chroma_mmi.c

   1 /*
   2  * Loongson SIMD optimized h264chroma
   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 "h264chroma_mips.h"
  26 #include "constants.h"
  27 #include "libavutil/mips/mmiutils.h"
  28
  29 void ff_put_h264_chroma_mc8_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
  30         int h, int x, int y)
  31 {
  32     int A = 64, B, C, D, E;
  33     double ftmp[10];
  34     uint64_t tmp[1];
  35
  36     if (!(x || y)) {
  37         /* x=0, y=0, A=64 */
  38         __asm__ volatile (
  39             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]           \n\t"
  40             "dli        %[tmp0],    0x06                               \n\t"
  41             "mtc1       %[tmp0],    %[ftmp4]                           \n\t"
  42
  43             "1:                                                        \n\t"
  44             MMI_ULDC1(%[ftmp1], %[src], 0x00)
  45             "addi       %[h],       %[h],           -0x04              \n\t"
  46             PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
  47             MMI_ULDC1(%[ftmp5], %[src], 0x00)
  48             PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
  49             MMI_ULDC1(%[ftmp6], %[src], 0x00)
  50             PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
  51             MMI_ULDC1(%[ftmp7], %[src], 0x00)
  52
  53             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]           \n\t"
  54             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]           \n\t"
  55             "psllh      %[ftmp1],   %[ftmp2],       %[ftmp4]           \n\t"
  56             "psllh      %[ftmp2],   %[ftmp3],       %[ftmp4]           \n\t"
  57             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
  58             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
  59             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]           \n\t"
  60             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
  61             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
  62             MMI_SDC1(%[ftmp1], %[dst], 0x00)
  63
  64             "punpcklbh  %[ftmp2],   %[ftmp5],       %[ftmp0]           \n\t"
  65             "punpckhbh  %[ftmp3],   %[ftmp5],       %[ftmp0]           \n\t"
  66             "psllh      %[ftmp1],   %[ftmp2],       %[ftmp4]           \n\t"
  67             "psllh      %[ftmp2],   %[ftmp3],       %[ftmp4]           \n\t"
  68             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
  69             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
  70             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]           \n\t"
  71             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
  72             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
  73             PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
  74             MMI_SDC1(%[ftmp1], %[dst], 0x00)
  75
  76             "punpcklbh  %[ftmp2],   %[ftmp6],       %[ftmp0]           \n\t"
  77             "punpckhbh  %[ftmp3],   %[ftmp6],       %[ftmp0]           \n\t"
  78             "psllh      %[ftmp1],   %[ftmp2],       %[ftmp4]           \n\t"
  79             "psllh      %[ftmp2],   %[ftmp3],       %[ftmp4]           \n\t"
  80             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
  81             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
  82             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]           \n\t"
  83             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
  84             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
  85             PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
  86             MMI_SDC1(%[ftmp1], %[dst], 0x00)
  87
  88             "punpcklbh  %[ftmp2],   %[ftmp7],       %[ftmp0]           \n\t"
  89             "punpckhbh  %[ftmp3],   %[ftmp7],       %[ftmp0]           \n\t"
  90             "psllh      %[ftmp1],   %[ftmp2],       %[ftmp4]           \n\t"
  91             "psllh      %[ftmp2],   %[ftmp3],       %[ftmp4]           \n\t"
  92             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
  93             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
  94             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]           \n\t"
  95             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
  96             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
  97             PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
  98             MMI_SDC1(%[ftmp1], %[dst], 0x00)
  99
 100             PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
 101             PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
 102             "bnez       %[h],       1b                                 \n\t"
 103             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 104               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 105               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 106               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 107               [tmp0]"=&r"(tmp[0]),
 108               [dst]"+&r"(dst),              [src]"+&r"(src),
 109               [h]"+&r"(h)
 110             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32)
 111             : "memory"
 112         );
 113     } else {
 114         if (x && y) {
 115             /* x!=0, y!=0 */
 116             D = x * y;
 117             B = (x << 3) - D;
 118             C = (y << 3) - D;
 119             A = 64 - D - B - C;
 120
 121             __asm__ volatile (
 122                 "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]           \n\t"
 123                 "dli        %[tmp0],    0x06                               \n\t"
 124                 "pshufh     %[A],       %[A],           %[ftmp0]           \n\t"
 125                 "pshufh     %[B],       %[B],           %[ftmp0]           \n\t"
 126                 "mtc1       %[tmp0],    %[ftmp9]                           \n\t"
 127                 "pshufh     %[C],       %[C],           %[ftmp0]           \n\t"
 128                 "pshufh     %[D],       %[D],           %[ftmp0]           \n\t"
 129
 130                 "1:                                                        \n\t"
 131                 MMI_ULDC1(%[ftmp1], %[src], 0x00)
 132                 MMI_ULDC1(%[ftmp2], %[src], 0x01)
 133                 PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
 134                 MMI_ULDC1(%[ftmp3], %[src], 0x00)
 135                 MMI_ULDC1(%[ftmp4], %[src], 0x01)
 136                 "addi       %[h],       %[h],           -0x02              \n\t"
 137
 138                 "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]           \n\t"
 139                 "punpckhbh  %[ftmp6],   %[ftmp1],       %[ftmp0]           \n\t"
 140                 "punpcklbh  %[ftmp7],   %[ftmp2],       %[ftmp0]           \n\t"
 141                 "punpckhbh  %[ftmp8],   %[ftmp2],       %[ftmp0]           \n\t"
 142                 "pmullh     %[ftmp5],   %[ftmp5],       %[A]               \n\t"
 143                 "pmullh     %[ftmp7],   %[ftmp7],       %[B]               \n\t"
 144                 "paddh      %[ftmp1],   %[ftmp5],       %[ftmp7]           \n\t"
 145                 "pmullh     %[ftmp6],   %[ftmp6],       %[A]               \n\t"
 146                 "pmullh     %[ftmp8],   %[ftmp8],       %[B]               \n\t"
 147                 "paddh      %[ftmp2],   %[ftmp6],       %[ftmp8]           \n\t"
 148
 149                 "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]           \n\t"
 150                 "punpckhbh  %[ftmp6],   %[ftmp3],       %[ftmp0]           \n\t"
 151                 "punpcklbh  %[ftmp7],   %[ftmp4],       %[ftmp0]           \n\t"
 152                 "punpckhbh  %[ftmp8],   %[ftmp4],       %[ftmp0]           \n\t"
 153                 "pmullh     %[ftmp5],   %[ftmp5],       %[C]               \n\t"
 154                 "pmullh     %[ftmp7],   %[ftmp7],       %[D]               \n\t"
 155                 "paddh      %[ftmp3],   %[ftmp5],       %[ftmp7]           \n\t"
 156                 "pmullh     %[ftmp6],   %[ftmp6],       %[C]               \n\t"
 157                 "pmullh     %[ftmp8],   %[ftmp8],       %[D]               \n\t"
 158                 "paddh      %[ftmp4],   %[ftmp6],       %[ftmp8]           \n\t"
 159
 160                 "paddh      %[ftmp1],   %[ftmp1],       %[ftmp3]           \n\t"
 161                 "paddh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
 162                 "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
 163                 "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
 164                 "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp9]           \n\t"
 165                 "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp9]           \n\t"
 166                 "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
 167                 MMI_SDC1(%[ftmp1], %[dst], 0x00)
 168                 PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
 169
 170                 MMI_ULDC1(%[ftmp1], %[src], 0x00)
 171                 MMI_ULDC1(%[ftmp2], %[src], 0x01)
 172                 PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
 173                 MMI_ULDC1(%[ftmp3], %[src], 0x00)
 174                 MMI_ULDC1(%[ftmp4], %[src], 0x01)
 175
 176                 "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]           \n\t"
 177                 "punpckhbh  %[ftmp6],   %[ftmp1],       %[ftmp0]           \n\t"
 178                 "punpcklbh  %[ftmp7],   %[ftmp2],       %[ftmp0]           \n\t"
 179                 "punpckhbh  %[ftmp8],   %[ftmp2],       %[ftmp0]           \n\t"
 180                 "pmullh     %[ftmp5],   %[ftmp5],       %[A]               \n\t"
 181                 "pmullh     %[ftmp7],   %[ftmp7],       %[B]               \n\t"
 182                 "paddh      %[ftmp1],   %[ftmp5],       %[ftmp7]           \n\t"
 183                 "pmullh     %[ftmp6],   %[ftmp6],       %[A]               \n\t"
 184                 "pmullh     %[ftmp8],   %[ftmp8],       %[B]               \n\t"
 185                 "paddh      %[ftmp2],   %[ftmp6],       %[ftmp8]           \n\t"
 186
 187                 "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]           \n\t"
 188                 "punpckhbh  %[ftmp6],   %[ftmp3],       %[ftmp0]           \n\t"
 189                 "punpcklbh  %[ftmp7],   %[ftmp4],       %[ftmp0]           \n\t"
 190                 "punpckhbh  %[ftmp8],   %[ftmp4],       %[ftmp0]           \n\t"
 191                 "pmullh     %[ftmp5],   %[ftmp5],       %[C]               \n\t"
 192                 "pmullh     %[ftmp7],   %[ftmp7],       %[D]               \n\t"
 193                 "paddh      %[ftmp3],   %[ftmp5],       %[ftmp7]           \n\t"
 194                 "pmullh     %[ftmp6],   %[ftmp6],       %[C]               \n\t"
 195                 "pmullh     %[ftmp8],   %[ftmp8],       %[D]               \n\t"
 196                 "paddh      %[ftmp4],   %[ftmp6],       %[ftmp8]           \n\t"
 197
 198                 "paddh      %[ftmp1],   %[ftmp1],       %[ftmp3]           \n\t"
 199                 "paddh      %[ftmp2],   %[ftmp2],       %[ftmp4]           \n\t"
 200                 "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
 201                 "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
 202                 "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp9]           \n\t"
 203                 "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp9]           \n\t"
 204                 "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
 205                 MMI_SDC1(%[ftmp1], %[dst], 0x00)
 206                 PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
 207
 208                 "bnez       %[h],       1b                                 \n\t"
 209                 : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 210                   [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 211                   [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 212                   [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 213                   [ftmp8]"=&f"(ftmp[8]),        [ftmp9]"=&f"(ftmp[9]),
 214                   [tmp0]"=&r"(tmp[0]),
 215                   [dst]"+&r"(dst),              [src]"+&r"(src),
 216                   [h]"+&r"(h)
 217                 : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 218                   [A]"f"(A),                    [B]"f"(B),
 219                   [C]"f"(C),                    [D]"f"(D)
 220                 : "memory"
 221             );
 222         } else {
 223             if (x) {
 224                 /* x!=0, y==0 */
 225                 E = x << 3;
 226                 A = 64 - E;
 227
 228                 __asm__ volatile (
 229                     "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]           \n\t"
 230                     "dli        %[tmp0],    0x06                               \n\t"
 231                     "pshufh     %[A],       %[A],           %[ftmp0]           \n\t"
 232                     "pshufh     %[E],       %[E],           %[ftmp0]           \n\t"
 233                     "mtc1       %[tmp0],    %[ftmp7]                           \n\t"
 234
 235                     "1:                                                        \n\t"
 236                     MMI_ULDC1(%[ftmp1], %[src], 0x00)
 237                     MMI_ULDC1(%[ftmp2], %[src], 0x01)
 238                     "addi       %[h],       %[h],           -0x01              \n\t"
 239                     PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
 240
 241                     "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]           \n\t"
 242                     "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]           \n\t"
 243                     "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]           \n\t"
 244                     "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]           \n\t"
 245                     "pmullh     %[ftmp3],   %[ftmp3],       %[A]               \n\t"
 246                     "pmullh     %[ftmp5],   %[ftmp5],       %[E]               \n\t"
 247                     "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]           \n\t"
 248                     "pmullh     %[ftmp4],   %[ftmp4],       %[A]               \n\t"
 249                     "pmullh     %[ftmp6],   %[ftmp6],       %[E]               \n\t"
 250                     "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]           \n\t"
 251
 252                     "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
 253                     "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
 254                     "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]           \n\t"
 255                     "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]           \n\t"
 256                     "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
 257                     MMI_SDC1(%[ftmp1], %[dst], 0x00)
 258                     PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
 259                     "bnez       %[h],       1b                                 \n\t"
 260                     : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 261                       [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 262                       [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 263                       [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 264                       [tmp0]"=&r"(tmp[0]),
 265                       [dst]"+&r"(dst),              [src]"+&r"(src),
 266                       [h]"+&r"(h)
 267                     : [stride]"r"((mips_reg)stride),
 268                       [ff_pw_32]"f"(ff_pw_32),
 269                       [A]"f"(A),                    [E]"f"(E)
 270                     : "memory"
 271                 );
 272             } else {
 273                 /* x==0, y!=0 */
 274                 E = y << 3;
 275                 A = 64 - E;
 276
 277                 __asm__ volatile (
 278                     "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]           \n\t"
 279                     "dli        %[tmp0],    0x06                               \n\t"
 280                     "pshufh     %[A],       %[A],           %[ftmp0]           \n\t"
 281                     "pshufh     %[E],       %[E],           %[ftmp0]           \n\t"
 282                     "mtc1       %[tmp0],    %[ftmp7]                           \n\t"
 283
 284                     "1:                                                        \n\t"
 285                     MMI_ULDC1(%[ftmp1], %[src], 0x00)
 286                     PTR_ADDU   "%[src],     %[src],         %[stride]          \n\t"
 287                     MMI_ULDC1(%[ftmp2], %[src], 0x00)
 288                     "addi       %[h],       %[h],           -0x01              \n\t"
 289
 290                     "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]           \n\t"
 291                     "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]           \n\t"
 292                     "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]           \n\t"
 293                     "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]           \n\t"
 294                     "pmullh     %[ftmp3],   %[ftmp3],       %[A]               \n\t"
 295                     "pmullh     %[ftmp5],   %[ftmp5],       %[E]               \n\t"
 296                     "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]           \n\t"
 297                     "pmullh     %[ftmp4],   %[ftmp4],       %[A]               \n\t"
 298                     "pmullh     %[ftmp6],   %[ftmp6],       %[E]               \n\t"
 299                     "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]           \n\t"
 300
 301                     "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]        \n\t"
 302                     "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]        \n\t"
 303                     "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]           \n\t"
 304                     "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]           \n\t"
 305                     "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]           \n\t"
 306                     MMI_SDC1(%[ftmp1], %[dst], 0x00)
 307
 308                     PTR_ADDU   "%[dst],     %[dst],         %[stride]          \n\t"
 309                     "bnez       %[h],       1b                                 \n\t"
 310                     : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 311                       [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 312                       [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 313                       [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 314                       [tmp0]"=&r"(tmp[0]),
 315                       [dst]"+&r"(dst),              [src]"+&r"(src),
 316                       [h]"+&r"(h)
 317                     : [stride]"r"((mips_reg)stride),
 318                       [ff_pw_32]"f"(ff_pw_32),
 319                       [A]"f"(A),                    [E]"f"(E)
 320                     : "memory"
 321                 );
 322             }
 323         }
 324     }
 325 }
 326
 327 void ff_avg_h264_chroma_mc8_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 328         int h, int x, int y)
 329 {
 330     int A = 64, B, C, D, E;
 331     double ftmp[10];
 332     uint64_t tmp[1];
 333
 334     if(!(x || y)){
 335         /* x=0, y=0, A=64 */
 336         __asm__ volatile (
 337             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 338             "dli        %[tmp0],    0x06                                \n\t"
 339             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 340             "mtc1       %[tmp0],    %[ftmp4]                            \n\t"
 341
 342             "1:                                                         \n\t"
 343             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 344             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 345             MMI_ULDC1(%[ftmp5], %[src], 0x00)
 346             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 347
 348             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 349             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 350             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 351             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 352             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 353             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 354             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 355             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 356             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 357             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 358             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 359             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 360             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 361
 362             "punpcklbh  %[ftmp2],   %[ftmp5],       %[ftmp0]            \n\t"
 363             "punpckhbh  %[ftmp3],   %[ftmp5],       %[ftmp0]            \n\t"
 364             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 365             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 366             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 367             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 368             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 369             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 370             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 371             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 372             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 373             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 374             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 375
 376             "addi       %[h],       %[h],           -0x02               \n\t"
 377             "bnez       %[h],       1b                                  \n\t"
 378             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 379               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 380               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 381               [tmp0]"=&r"(tmp[0]),
 382               [dst]"+&r"(dst),              [src]"+&r"(src),
 383               [h]"+&r"(h)
 384             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 385               [A]"f"(A)
 386             : "memory"
 387         );
 388     } else {
 389         if(x && y) {
 390             /* x!=0, y!=0 */
 391             D = x * y;
 392             B = (x << 3) - D;
 393             C = (y << 3) - D;
 394             A = 64 - D - B - C;
 395             __asm__ volatile (
 396                 "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]       \n\t"
 397                 "dli        %[tmp0],    0x06                           \n\t"
 398                 "pshufh     %[A],       %[A],           %[ftmp0]       \n\t"
 399                 "pshufh     %[B],       %[B],           %[ftmp0]       \n\t"
 400                 "mtc1       %[tmp0],    %[ftmp9]                       \n\t"
 401                 "pshufh     %[C],       %[C],           %[ftmp0]       \n\t"
 402                 "pshufh     %[D],       %[D],           %[ftmp0]       \n\t"
 403
 404                 "1:                                                    \n\t"
 405                 MMI_ULDC1(%[ftmp1], %[src], 0x00)
 406                 MMI_ULDC1(%[ftmp2], %[src], 0x01)
 407                 PTR_ADDU   "%[src],     %[src],         %[stride]      \n\t"
 408                 MMI_ULDC1(%[ftmp3], %[src], 0x00)
 409                 MMI_ULDC1(%[ftmp4], %[src], 0x01)
 410                 "addi       %[h],       %[h],           -0x01          \n\t"
 411
 412                 "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]       \n\t"
 413                 "punpckhbh  %[ftmp6],   %[ftmp1],       %[ftmp0]       \n\t"
 414                 "punpcklbh  %[ftmp7],   %[ftmp2],       %[ftmp0]       \n\t"
 415                 "punpckhbh  %[ftmp8],   %[ftmp2],       %[ftmp0]       \n\t"
 416                 "pmullh     %[ftmp5],   %[ftmp5],       %[A]           \n\t"
 417                 "pmullh     %[ftmp7],   %[ftmp7],       %[B]           \n\t"
 418                 "paddh      %[ftmp1],   %[ftmp5],       %[ftmp7]       \n\t"
 419                 "pmullh     %[ftmp6],   %[ftmp6],       %[A]           \n\t"
 420                 "pmullh     %[ftmp8],   %[ftmp8],       %[B]           \n\t"
 421                 "paddh      %[ftmp2],   %[ftmp6],       %[ftmp8]       \n\t"
 422
 423                 "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]       \n\t"
 424                 "punpckhbh  %[ftmp6],   %[ftmp3],       %[ftmp0]       \n\t"
 425                 "punpcklbh  %[ftmp7],   %[ftmp4],       %[ftmp0]       \n\t"
 426                 "punpckhbh  %[ftmp8],   %[ftmp4],       %[ftmp0]       \n\t"
 427                 "pmullh     %[ftmp5],   %[ftmp5],       %[C]           \n\t"
 428                 "pmullh     %[ftmp7],   %[ftmp7],       %[D]           \n\t"
 429                 "paddh      %[ftmp3],   %[ftmp5],       %[ftmp7]       \n\t"
 430                 "pmullh     %[ftmp6],   %[ftmp6],       %[C]           \n\t"
 431                 "pmullh     %[ftmp8],   %[ftmp8],       %[D]           \n\t"
 432                 "paddh      %[ftmp4],   %[ftmp6],       %[ftmp8]       \n\t"
 433
 434                 "paddh      %[ftmp1],   %[ftmp1],       %[ftmp3]       \n\t"
 435                 "paddh      %[ftmp2],   %[ftmp2],       %[ftmp4]       \n\t"
 436                 "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]    \n\t"
 437                 "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]    \n\t"
 438                 "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp9]       \n\t"
 439                 "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp9]       \n\t"
 440                 "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 441                 MMI_LDC1(%[ftmp2], %[dst], 0x00)
 442                 "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 443                 MMI_SDC1(%[ftmp1], %[dst], 0x00)
 444                 PTR_ADDU   "%[dst],     %[dst],         %[stride]      \n\t"
 445                 "bnez       %[h],       1b                             \n\t"
 446                 : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 447                   [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 448                   [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 449                   [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 450                   [ftmp8]"=&f"(ftmp[8]),        [ftmp9]"=&f"(ftmp[9]),
 451                   [tmp0]"=&r"(tmp[0]),
 452                   [dst]"+&r"(dst),              [src]"+&r"(src),
 453                   [h]"+&r"(h)
 454                 : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 455                   [A]"f"(A),                    [B]"f"(B),
 456                   [C]"f"(C),                    [D]"f"(D)
 457                 : "memory"
 458             );
 459         } else {
 460             if(x) {
 461                 /* x!=0, y==0 */
 462                 E = x << 3;
 463                 A = 64 - E;
 464                 __asm__ volatile (
 465                     "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]       \n\t"
 466                     "dli        %[tmp0],    0x06                           \n\t"
 467                     "pshufh     %[A],       %[A],           %[ftmp0]       \n\t"
 468                     "pshufh     %[E],       %[E],           %[ftmp0]       \n\t"
 469                     "mtc1       %[tmp0],    %[ftmp7]                       \n\t"
 470
 471                     "1:                                                    \n\t"
 472                     MMI_ULDC1(%[ftmp1], %[src], 0x00)
 473                     MMI_ULDC1(%[ftmp2], %[src], 0x01)
 474                     PTR_ADDU   "%[src],     %[src],         %[stride]      \n\t"
 475                     "addi       %[h],       %[h],           -0x01          \n\t"
 476
 477                     "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]       \n\t"
 478                     "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]       \n\t"
 479                     "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]       \n\t"
 480                     "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]       \n\t"
 481                     "pmullh     %[ftmp3],   %[ftmp3],       %[A]           \n\t"
 482                     "pmullh     %[ftmp5],   %[ftmp5],       %[E]           \n\t"
 483                     "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]       \n\t"
 484                     "pmullh     %[ftmp4],   %[ftmp4],       %[A]           \n\t"
 485                     "pmullh     %[ftmp6],   %[ftmp6],       %[E]           \n\t"
 486                     "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]       \n\t"
 487
 488                     "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]    \n\t"
 489                     "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]    \n\t"
 490                     "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]       \n\t"
 491                     "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]       \n\t"
 492                     "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 493                     MMI_LDC1(%[ftmp2], %[dst], 0x00)
 494                     "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 495                     MMI_SDC1(%[ftmp1], %[dst], 0x00)
 496                     PTR_ADDU   "%[dst],     %[dst],         %[stride]      \n\t"
 497                     "bnez       %[h],       1b                             \n\t"
 498                     : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 499                       [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 500                       [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 501                       [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 502                       [tmp0]"=&r"(tmp[0]),
 503                       [dst]"+&r"(dst),              [src]"+&r"(src),
 504                       [h]"+&r"(h)
 505                     : [stride]"r"((mips_reg)stride),
 506                       [ff_pw_32]"f"(ff_pw_32),
 507                       [A]"f"(A),                    [E]"f"(E)
 508                     : "memory"
 509                 );
 510             } else {
 511                 /* x==0, y!=0 */
 512                 E = y << 3;
 513                 A = 64 - E;
 514                 __asm__ volatile (
 515                     "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]       \n\t"
 516                     "dli        %[tmp0],    0x06                           \n\t"
 517                     "pshufh     %[A],       %[A],           %[ftmp0]       \n\t"
 518                     "pshufh     %[E],       %[E],           %[ftmp0]       \n\t"
 519                     "mtc1       %[tmp0],    %[ftmp7]                       \n\t"
 520
 521                     "1:                                                    \n\t"
 522                     MMI_ULDC1(%[ftmp1], %[src], 0x00)
 523                     PTR_ADDU   "%[src],     %[src],         %[stride]      \n\t"
 524                     MMI_ULDC1(%[ftmp2], %[src], 0x00)
 525                     "addi       %[h],       %[h],           -0x01          \n\t"
 526
 527                     "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]       \n\t"
 528                     "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]       \n\t"
 529                     "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]       \n\t"
 530                     "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]       \n\t"
 531                     "pmullh     %[ftmp3],   %[ftmp3],       %[A]           \n\t"
 532                     "pmullh     %[ftmp5],   %[ftmp5],       %[E]           \n\t"
 533                     "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]       \n\t"
 534                     "pmullh     %[ftmp4],   %[ftmp4],       %[A]           \n\t"
 535                     "pmullh     %[ftmp6],   %[ftmp6],       %[E]           \n\t"
 536                     "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]       \n\t"
 537
 538                     "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]    \n\t"
 539                     "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]    \n\t"
 540                     "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]       \n\t"
 541                     "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]       \n\t"
 542                     "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 543                     MMI_LDC1(%[ftmp2], %[dst], 0x00)
 544                     "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]       \n\t"
 545                     MMI_SDC1(%[ftmp1], %[dst], 0x00)
 546                     PTR_ADDU   "%[dst],     %[dst],         %[stride]      \n\t"
 547                     "bnez       %[h],       1b                             \n\t"
 548                     : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 549                       [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 550                       [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 551                       [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 552                       [tmp0]"=&r"(tmp[0]),
 553                       [dst]"+&r"(dst),              [src]"+&r"(src),
 554                       [h]"+&r"(h)
 555                     : [stride]"r"((mips_reg)stride),
 556                       [ff_pw_32]"f"(ff_pw_32),
 557                       [A]"f"(A),                    [E]"f"(E)
 558                     : "memory"
 559                 );
 560             }
 561         }
 562     }
 563 }
 564
 565 void ff_put_h264_chroma_mc4_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 566         int h, int x, int y)
 567 {
 568     const int A = (8 - x) * (8 - y);
 569     const int B = x * (8 - y);
 570     const int C = (8 - x) *  y;
 571     const int D = x *  y;
 572     const int E = B + C;
 573     double ftmp[8];
 574     uint64_t tmp[1];
 575     mips_reg addr[1];
 576     DECLARE_VAR_LOW32;
 577
 578     if (D) {
 579         __asm__ volatile (
 580             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 581             "dli        %[tmp0],    0x06                                \n\t"
 582             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 583             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
 584             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 585             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
 586             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
 587
 588             "1:                                                         \n\t"
 589             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
 590             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 591             MMI_ULWC1(%[ftmp2], %[src], 0x01)
 592             MMI_ULWC1(%[ftmp3], %[addr0], 0x00)
 593             MMI_ULWC1(%[ftmp4], %[addr0], 0x01)
 594
 595             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
 596             "punpcklbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 597             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
 598             "pmullh     %[ftmp6],   %[ftmp6],       %[B]                \n\t"
 599             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp6]            \n\t"
 600
 601             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
 602             "punpcklbh  %[ftmp6],   %[ftmp4],       %[ftmp0]            \n\t"
 603             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
 604             "pmullh     %[ftmp6],   %[ftmp6],       %[D]                \n\t"
 605             "paddh      %[ftmp2],   %[ftmp5],       %[ftmp6]            \n\t"
 606
 607             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 608             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 609             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 610             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 611             "addi       %[h],       %[h],           -0x01               \n\t"
 612             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 613             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 614             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 615             "bnez       %[h],       1b                                  \n\t"
 616             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 617               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 618               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 619               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 620               [tmp0]"=&r"(tmp[0]),
 621               RESTRICT_ASM_LOW32
 622               [addr0]"=&r"(addr[0]),
 623               [dst]"+&r"(dst),              [src]"+&r"(src),
 624               [h]"+&r"(h)
 625             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 626               [A]"f"(A),                    [B]"f"(B),
 627               [C]"f"(C),                    [D]"f"(D)
 628             : "memory"
 629         );
 630     } else if (E) {
 631         const int step = C ? stride : 1;
 632
 633         __asm__ volatile (
 634             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 635             "dli        %[tmp0],    0x06                                \n\t"
 636             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 637             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 638             "mtc1       %[tmp0],    %[ftmp5]                            \n\t"
 639
 640             "1:                                                         \n\t"
 641             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 642             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 643             MMI_ULWC1(%[ftmp2], %[addr0], 0x00)
 644
 645             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 646             "punpcklbh  %[ftmp4],   %[ftmp2],       %[ftmp0]            \n\t"
 647             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 648             "pmullh     %[ftmp4],   %[ftmp4],       %[E]                \n\t"
 649             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp4]            \n\t"
 650
 651             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 652             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp5]            \n\t"
 653             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 654             "addi       %[h],       %[h],           -0x01               \n\t"
 655             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 656             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 657             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 658             "bnez       %[h],       1b                                  \n\t"
 659             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 660               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 661               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 662               [tmp0]"=&r"(tmp[0]),
 663               RESTRICT_ASM_LOW32
 664               [addr0]"=&r"(addr[0]),
 665               [dst]"+&r"(dst),              [src]"+&r"(src),
 666               [h]"+&r"(h)
 667             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 668               [ff_pw_32]"f"(ff_pw_32),
 669               [A]"f"(A),                    [E]"f"(E)
 670             : "memory"
 671         );
 672     } else {
 673         __asm__ volatile (
 674             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 675             "dli        %[tmp0],    0x06                                \n\t"
 676             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 677             "mtc1       %[tmp0],    %[ftmp3]                            \n\t"
 678
 679             "1:                                                         \n\t"
 680             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 681             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 682             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 683             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 684             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 685             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 686             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 687             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 688             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 689
 690             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 691             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 692             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 693             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 694             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 695             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 696             "addi       %[h],       %[h],           -0x02               \n\t"
 697             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 698
 699             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 700             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 701             "bnez       %[h],       1b                                  \n\t"
 702             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 703               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 704               [tmp0]"=&r"(tmp[0]),
 705               RESTRICT_ASM_LOW32
 706               [dst]"+&r"(dst),              [src]"+&r"(src),
 707               [h]"+&r"(h)
 708             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 709               [A]"f"(A)
 710             : "memory"
 711         );
 712     }
 713 }
 714
 715 void ff_avg_h264_chroma_mc4_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 716         int h, int x, int y)
 717 {
 718     const int A = (8 - x) *(8 - y);
 719     const int B = x * (8 - y);
 720     const int C = (8 - x) * y;
 721     const int D = x * y;
 722     const int E = B + C;
 723     double ftmp[8];
 724     uint64_t tmp[1];
 725     mips_reg addr[1];
 726     DECLARE_VAR_LOW32;
 727
 728     if (D) {
 729         __asm__ volatile (
 730             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 731             "dli        %[tmp0],    0x06                                \n\t"
 732             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 733             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
 734             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 735             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
 736             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
 737
 738             "1:                                                         \n\t"
 739             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
 740             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 741             MMI_ULWC1(%[ftmp2], %[src], 0x01)
 742             MMI_ULWC1(%[ftmp3], %[addr0], 0x00)
 743             MMI_ULWC1(%[ftmp4], %[addr0], 0x01)
 744
 745             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
 746             "punpcklbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 747             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
 748             "pmullh     %[ftmp6],   %[ftmp6],       %[B]                \n\t"
 749             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp6]            \n\t"
 750
 751             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
 752             "punpcklbh  %[ftmp6],   %[ftmp4],       %[ftmp0]            \n\t"
 753             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
 754             "pmullh     %[ftmp6],   %[ftmp6],       %[D]                \n\t"
 755             "paddh      %[ftmp2],   %[ftmp5],       %[ftmp6]            \n\t"
 756
 757             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 758             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 759             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 760             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 761             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 762             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 763             "addi       %[h],       %[h],           -0x01               \n\t"
 764             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 765             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 766             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 767             "bnez       %[h],       1b                                  \n\t"
 768             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 769               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 770               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 771               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 772               [tmp0]"=&r"(tmp[0]),
 773               RESTRICT_ASM_LOW32
 774               [addr0]"=&r"(addr[0]),
 775               [dst]"+&r"(dst),              [src]"+&r"(src),
 776               [h]"+&r"(h)
 777             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 778               [A]"f"(A),                    [B]"f"(B),
 779               [C]"f"(C),                    [D]"f"(D)
 780             : "memory"
 781         );
 782     } else if (E) {
 783         const int step = C ? stride : 1;
 784
 785         __asm__ volatile (
 786             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 787             "dli        %[tmp0],    0x06                                \n\t"
 788             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 789             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 790             "mtc1       %[tmp0],    %[ftmp5]                            \n\t"
 791             "1:                                                         \n\t"
 792             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 793             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 794             MMI_ULWC1(%[ftmp2], %[addr0], 0x00)
 795
 796             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 797             "punpcklbh  %[ftmp4],   %[ftmp2],       %[ftmp0]            \n\t"
 798             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 799             "pmullh     %[ftmp4],   %[ftmp4],       %[E]                \n\t"
 800             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp4]            \n\t"
 801
 802             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 803             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp5]            \n\t"
 804             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 805             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 806             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 807             "addi       %[h],       %[h],           -0x01               \n\t"
 808             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 809             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 810             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 811             "bnez       %[h],       1b                                  \n\t"
 812             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 813               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 814               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 815               [tmp0]"=&r"(tmp[0]),
 816               RESTRICT_ASM_LOW32
 817               [addr0]"=&r"(addr[0]),
 818               [dst]"+&r"(dst),              [src]"+&r"(src),
 819               [h]"+&r"(h)
 820             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 821               [ff_pw_32]"f"(ff_pw_32),
 822               [A]"f"(A),                    [E]"f"(E)
 823             : "memory"
 824         );
 825     } else {
 826         __asm__ volatile (
 827             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 828             "dli        %[tmp0],    0x06                                \n\t"
 829             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 830             "mtc1       %[tmp0],    %[ftmp3]                            \n\t"
 831
 832             "1:                                                         \n\t"
 833             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 834             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 835             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 836             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 837             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 838             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 839             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 840             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 841             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 842             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 843             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 844
 845             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 846             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 847             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 848             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 849             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 850             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 851             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 852             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 853             "addi       %[h],       %[h],           -0x02               \n\t"
 854             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 855
 856             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 857             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 858             "bnez       %[h],       1b                                  \n\t"
 859             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 860               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 861               [tmp0]"=&r"(tmp[0]),
 862               RESTRICT_ASM_LOW32
 863               [dst]"+&r"(dst),              [src]"+&r"(src),
 864               [h]"+&r"(h)
 865             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 866               [A]"f"(A)
 867             : "memory"
 868         );
 869     }
 870 }