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     const int A = (8 - x) * (8 - y);
  33     const int B = x * (8 - y);
  34     const int C = (8 - x) * y;
  35     const int D = x * y;
  36     const int E = B + C;
  37     double ftmp[10];
  38     uint64_t tmp[1];
  39     mips_reg addr[1];
  40     DECLARE_VAR_ALL64;
  41
  42     if (D) {
  43         __asm__ volatile (
  44             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
  45             "dli        %[tmp0],    0x06                                \n\t"
  46             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
  47             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
  48             "mtc1       %[tmp0],    %[ftmp9]                            \n\t"
  49             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
  50             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
  51
  52             "1:                                                         \n\t"
  53             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
  54             MMI_ULDC1(%[ftmp1], %[src], 0x00)
  55             MMI_ULDC1(%[ftmp2], %[src], 0x01)
  56             MMI_ULDC1(%[ftmp3], %[addr0], 0x00)
  57             MMI_ULDC1(%[ftmp4], %[addr0], 0x01)
  58
  59             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
  60             "punpckhbh  %[ftmp6],   %[ftmp1],       %[ftmp0]            \n\t"
  61             "punpcklbh  %[ftmp7],   %[ftmp2],       %[ftmp0]            \n\t"
  62             "punpckhbh  %[ftmp8],   %[ftmp2],       %[ftmp0]            \n\t"
  63             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
  64             "pmullh     %[ftmp7],   %[ftmp7],       %[B]                \n\t"
  65             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp7]            \n\t"
  66             "pmullh     %[ftmp6],   %[ftmp6],       %[A]                \n\t"
  67             "pmullh     %[ftmp8],   %[ftmp8],       %[B]                \n\t"
  68             "paddh      %[ftmp2],   %[ftmp6],       %[ftmp8]            \n\t"
  69
  70             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
  71             "punpckhbh  %[ftmp6],   %[ftmp3],       %[ftmp0]            \n\t"
  72             "punpcklbh  %[ftmp7],   %[ftmp4],       %[ftmp0]            \n\t"
  73             "punpckhbh  %[ftmp8],   %[ftmp4],       %[ftmp0]            \n\t"
  74             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
  75             "pmullh     %[ftmp7],   %[ftmp7],       %[D]                \n\t"
  76             "paddh      %[ftmp3],   %[ftmp5],       %[ftmp7]            \n\t"
  77             "pmullh     %[ftmp6],   %[ftmp6],       %[C]                \n\t"
  78             "pmullh     %[ftmp8],   %[ftmp8],       %[D]                \n\t"
  79             "paddh      %[ftmp4],   %[ftmp6],       %[ftmp8]            \n\t"
  80
  81             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
  82             "paddh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
  83             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
  84             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
  85             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp9]            \n\t"
  86             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp9]            \n\t"
  87             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
  88             "addi       %[h],       %[h],           -0x01               \n\t"
  89             MMI_SDC1(%[ftmp1], %[dst], 0x00)
  90             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
  91             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
  92             "bnez       %[h],       1b                                  \n\t"
  93             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
  94               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
  95               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
  96               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
  97               [ftmp8]"=&f"(ftmp[8]),        [ftmp9]"=&f"(ftmp[9]),
  98               [tmp0]"=&r"(tmp[0]),
  99               RESTRICT_ASM_ALL64
 100               [addr0]"=&r"(addr[0]),
 101               [dst]"+&r"(dst),              [src]"+&r"(src),
 102               [h]"+&r"(h)
 103             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 104               [A]"f"(A),                    [B]"f"(B),
 105               [C]"f"(C),                    [D]"f"(D)
 106             : "memory"
 107         );
 108     } else if (E) {
 109         const int step = C ? stride : 1;
 110
 111         __asm__ volatile (
 112             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 113             "dli        %[tmp0],    0x06                                \n\t"
 114             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 115             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 116             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 117
 118             "1:                                                         \n\t"
 119             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 120             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 121             MMI_ULDC1(%[ftmp2], %[addr0], 0x00)
 122
 123             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 124             "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]            \n\t"
 125             "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]            \n\t"
 126             "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 127             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 128             "pmullh     %[ftmp5],   %[ftmp5],       %[E]                \n\t"
 129             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]            \n\t"
 130             "pmullh     %[ftmp4],   %[ftmp4],       %[A]                \n\t"
 131             "pmullh     %[ftmp6],   %[ftmp6],       %[E]                \n\t"
 132             "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]            \n\t"
 133
 134             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 135             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 136             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 137             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]            \n\t"
 138             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 139             "addi       %[h],       %[h],           -0x01               \n\t"
 140             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 141             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 142             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 143             "bnez       %[h],       1b                                  \n\t"
 144             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 145               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 146               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 147               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 148               [tmp0]"=&r"(tmp[0]),
 149               RESTRICT_ASM_ALL64
 150               [addr0]"=&r"(addr[0]),
 151               [dst]"+&r"(dst),              [src]"+&r"(src),
 152               [h]"+&r"(h)
 153             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 154               [ff_pw_32]"f"(ff_pw_32),
 155               [A]"f"(A),                    [E]"f"(E)
 156             : "memory"
 157         );
 158     } else {
 159         __asm__ volatile (
 160             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 161             "dli        %[tmp0],    0x06                                \n\t"
 162             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 163             "mtc1       %[tmp0],    %[ftmp4]                            \n\t"
 164
 165             "1:                                                         \n\t"
 166             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 167             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 168             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 169             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 170             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 171             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 172             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 173             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 174             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 175             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 176             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 177             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 178
 179             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 180             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 181             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 182             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 183             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 184             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 185             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 186             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 187             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 188             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 189             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 190             "addi       %[h],       %[h],           -0x02               \n\t"
 191             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 192
 193             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 194             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 195             "bnez       %[h],       1b                                  \n\t"
 196             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 197               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 198               [ftmp4]"=&f"(ftmp[4]),
 199               [tmp0]"=&r"(tmp[0]),
 200               RESTRICT_ASM_ALL64
 201               [dst]"+&r"(dst),              [src]"+&r"(src),
 202               [h]"+&r"(h)
 203             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 204               [A]"f"(A)
 205             : "memory"
 206         );
 207     }
 208 }
 209
 210 void ff_avg_h264_chroma_mc8_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 211         int h, int x, int y)
 212 {
 213     const int A = (8 - x) * (8 - y);
 214     const int B = x * (8 - y);
 215     const int C = (8 - x) * y;
 216     const int D = x * y;
 217     const int E = B + C;
 218     double ftmp[10];
 219     uint64_t tmp[1];
 220     mips_reg addr[1];
 221     DECLARE_VAR_ALL64;
 222
 223     if (D) {
 224         __asm__ volatile (
 225             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 226             "dli        %[tmp0],    0x06                                \n\t"
 227             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 228             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
 229             "mtc1       %[tmp0],    %[ftmp9]                            \n\t"
 230             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
 231             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
 232
 233             "1:                                                         \n\t"
 234             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
 235             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 236             MMI_ULDC1(%[ftmp2], %[src], 0x01)
 237             MMI_ULDC1(%[ftmp3], %[addr0], 0x00)
 238             MMI_ULDC1(%[ftmp4], %[addr0], 0x01)
 239
 240             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
 241             "punpckhbh  %[ftmp6],   %[ftmp1],       %[ftmp0]            \n\t"
 242             "punpcklbh  %[ftmp7],   %[ftmp2],       %[ftmp0]            \n\t"
 243             "punpckhbh  %[ftmp8],   %[ftmp2],       %[ftmp0]            \n\t"
 244             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
 245             "pmullh     %[ftmp7],   %[ftmp7],       %[B]                \n\t"
 246             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp7]            \n\t"
 247             "pmullh     %[ftmp6],   %[ftmp6],       %[A]                \n\t"
 248             "pmullh     %[ftmp8],   %[ftmp8],       %[B]                \n\t"
 249             "paddh      %[ftmp2],   %[ftmp6],       %[ftmp8]            \n\t"
 250
 251             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
 252             "punpckhbh  %[ftmp6],   %[ftmp3],       %[ftmp0]            \n\t"
 253             "punpcklbh  %[ftmp7],   %[ftmp4],       %[ftmp0]            \n\t"
 254             "punpckhbh  %[ftmp8],   %[ftmp4],       %[ftmp0]            \n\t"
 255             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
 256             "pmullh     %[ftmp7],   %[ftmp7],       %[D]                \n\t"
 257             "paddh      %[ftmp3],   %[ftmp5],       %[ftmp7]            \n\t"
 258             "pmullh     %[ftmp6],   %[ftmp6],       %[C]                \n\t"
 259             "pmullh     %[ftmp8],   %[ftmp8],       %[D]                \n\t"
 260             "paddh      %[ftmp4],   %[ftmp6],       %[ftmp8]            \n\t"
 261
 262             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 263             "paddh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 264             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 265             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 266             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp9]            \n\t"
 267             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp9]            \n\t"
 268             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 269             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 270             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 271             "addi       %[h],       %[h],           -0x01               \n\t"
 272             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 273             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 274             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 275             "bnez       %[h],       1b                                  \n\t"
 276             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 277               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 278               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 279               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 280               [ftmp8]"=&f"(ftmp[8]),        [ftmp9]"=&f"(ftmp[9]),
 281               [tmp0]"=&r"(tmp[0]),
 282               RESTRICT_ASM_ALL64
 283               [addr0]"=&r"(addr[0]),
 284               [dst]"+&r"(dst),              [src]"+&r"(src),
 285               [h]"+&r"(h)
 286             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 287               [A]"f"(A),                    [B]"f"(B),
 288               [C]"f"(C),                    [D]"f"(D)
 289             : "memory"
 290         );
 291     } else if (E) {
 292         const int step = C ? stride : 1;
 293
 294         __asm__ volatile (
 295             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 296             "dli        %[tmp0],    0x06                                \n\t"
 297             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 298             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 299             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 300
 301             "1:                                                         \n\t"
 302             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 303             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 304             MMI_ULDC1(%[ftmp2], %[addr0], 0x00)
 305
 306             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 307             "punpckhbh  %[ftmp4],   %[ftmp1],       %[ftmp0]            \n\t"
 308             "punpcklbh  %[ftmp5],   %[ftmp2],       %[ftmp0]            \n\t"
 309             "punpckhbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 310             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 311             "pmullh     %[ftmp5],   %[ftmp5],       %[E]                \n\t"
 312             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp5]            \n\t"
 313             "pmullh     %[ftmp4],   %[ftmp4],       %[A]                \n\t"
 314             "pmullh     %[ftmp6],   %[ftmp6],       %[E]                \n\t"
 315             "paddh      %[ftmp2],   %[ftmp4],       %[ftmp6]            \n\t"
 316
 317             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 318             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 319             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 320             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp7]            \n\t"
 321             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 322             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 323             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 324             "addi       %[h],       %[h],           -0x01               \n\t"
 325             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 326             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 327             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 328             "bnez       %[h],       1b                                  \n\t"
 329             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 330               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 331               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 332               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 333               [tmp0]"=&r"(tmp[0]),
 334               RESTRICT_ASM_ALL64
 335               [addr0]"=&r"(addr[0]),
 336               [dst]"+&r"(dst),              [src]"+&r"(src),
 337               [h]"+&r"(h)
 338             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 339               [ff_pw_32]"f"(ff_pw_32),
 340               [A]"f"(A),                    [E]"f"(E)
 341             : "memory"
 342         );
 343     } else {
 344         __asm__ volatile (
 345             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 346             "dli        %[tmp0],    0x06                                \n\t"
 347             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 348             "mtc1       %[tmp0],    %[ftmp4]                            \n\t"
 349
 350             "1:                                                         \n\t"
 351             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 352             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 353             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 354             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 355             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 356             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 357             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 358             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 359             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 360             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 361             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 362             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 363             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 364             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 365             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 366
 367             MMI_ULDC1(%[ftmp1], %[src], 0x00)
 368             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 369             "punpckhbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 370             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 371             "pmullh     %[ftmp2],   %[ftmp3],       %[A]                \n\t"
 372             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 373             "paddh      %[ftmp2],   %[ftmp2],       %[ff_pw_32]         \n\t"
 374             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp4]            \n\t"
 375             "psrlh      %[ftmp2],   %[ftmp2],       %[ftmp4]            \n\t"
 376             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 377             MMI_LDC1(%[ftmp2], %[dst], 0x00)
 378             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 379             "addi       %[h],       %[h],           -0x02               \n\t"
 380             MMI_SDC1(%[ftmp1], %[dst], 0x00)
 381
 382             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 383             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 384             "bnez       %[h],       1b                                  \n\t"
 385             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 386               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 387               [ftmp4]"=&f"(ftmp[4]),
 388               [tmp0]"=&r"(tmp[0]),
 389               RESTRICT_ASM_ALL64
 390               [dst]"+&r"(dst),              [src]"+&r"(src),
 391               [h]"+&r"(h)
 392             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 393               [A]"f"(A)
 394             : "memory"
 395         );
 396     }
 397 }
 398
 399 void ff_put_h264_chroma_mc4_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 400         int h, int x, int y)
 401 {
 402     const int A = (8 - x) * (8 - y);
 403     const int B = x * (8 - y);
 404     const int C = (8 - x) *  y;
 405     const int D = x *  y;
 406     const int E = B + C;
 407     double ftmp[8];
 408     uint64_t tmp[1];
 409     mips_reg addr[1];
 410     DECLARE_VAR_LOW32;
 411
 412     if (D) {
 413         __asm__ volatile (
 414             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 415             "dli        %[tmp0],    0x06                                \n\t"
 416             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 417             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
 418             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 419             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
 420             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
 421
 422             "1:                                                         \n\t"
 423             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
 424             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 425             MMI_ULWC1(%[ftmp2], %[src], 0x01)
 426             MMI_ULWC1(%[ftmp3], %[addr0], 0x00)
 427             MMI_ULWC1(%[ftmp4], %[addr0], 0x01)
 428
 429             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
 430             "punpcklbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 431             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
 432             "pmullh     %[ftmp6],   %[ftmp6],       %[B]                \n\t"
 433             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp6]            \n\t"
 434
 435             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
 436             "punpcklbh  %[ftmp6],   %[ftmp4],       %[ftmp0]            \n\t"
 437             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
 438             "pmullh     %[ftmp6],   %[ftmp6],       %[D]                \n\t"
 439             "paddh      %[ftmp2],   %[ftmp5],       %[ftmp6]            \n\t"
 440
 441             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 442             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 443             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 444             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 445             "addi       %[h],       %[h],           -0x01               \n\t"
 446             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 447             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 448             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 449             "bnez       %[h],       1b                                  \n\t"
 450             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 451               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 452               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 453               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 454               [tmp0]"=&r"(tmp[0]),
 455               RESTRICT_ASM_LOW32
 456               [addr0]"=&r"(addr[0]),
 457               [dst]"+&r"(dst),              [src]"+&r"(src),
 458               [h]"+&r"(h)
 459             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 460               [A]"f"(A),                    [B]"f"(B),
 461               [C]"f"(C),                    [D]"f"(D)
 462             : "memory"
 463         );
 464     } else if (E) {
 465         const int step = C ? stride : 1;
 466
 467         __asm__ volatile (
 468             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 469             "dli        %[tmp0],    0x06                                \n\t"
 470             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 471             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 472             "mtc1       %[tmp0],    %[ftmp5]                            \n\t"
 473
 474             "1:                                                         \n\t"
 475             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 476             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 477             MMI_ULWC1(%[ftmp2], %[addr0], 0x00)
 478
 479             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 480             "punpcklbh  %[ftmp4],   %[ftmp2],       %[ftmp0]            \n\t"
 481             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 482             "pmullh     %[ftmp4],   %[ftmp4],       %[E]                \n\t"
 483             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp4]            \n\t"
 484
 485             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 486             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp5]            \n\t"
 487             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 488             "addi       %[h],       %[h],           -0x01               \n\t"
 489             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 490             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 491             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 492             "bnez       %[h],       1b                                  \n\t"
 493             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 494               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 495               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 496               [tmp0]"=&r"(tmp[0]),
 497               RESTRICT_ASM_LOW32
 498               [addr0]"=&r"(addr[0]),
 499               [dst]"+&r"(dst),              [src]"+&r"(src),
 500               [h]"+&r"(h)
 501             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 502               [ff_pw_32]"f"(ff_pw_32),
 503               [A]"f"(A),                    [E]"f"(E)
 504             : "memory"
 505         );
 506     } else {
 507         __asm__ volatile (
 508             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 509             "dli        %[tmp0],    0x06                                \n\t"
 510             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 511             "mtc1       %[tmp0],    %[ftmp3]                            \n\t"
 512
 513             "1:                                                         \n\t"
 514             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 515             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 516             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 517             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 518             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 519             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 520             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 521             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 522             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 523
 524             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 525             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 526             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 527             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 528             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 529             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 530             "addi       %[h],       %[h],           -0x02               \n\t"
 531             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 532
 533             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 534             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 535             "bnez       %[h],       1b                                  \n\t"
 536             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 537               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 538               [tmp0]"=&r"(tmp[0]),
 539               RESTRICT_ASM_LOW32
 540               [dst]"+&r"(dst),              [src]"+&r"(src),
 541               [h]"+&r"(h)
 542             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 543               [A]"f"(A)
 544             : "memory"
 545         );
 546     }
 547 }
 548
 549 void ff_avg_h264_chroma_mc4_mmi(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
 550         int h, int x, int y)
 551 {
 552     const int A = (8 - x) *(8 - y);
 553     const int B = x * (8 - y);
 554     const int C = (8 - x) * y;
 555     const int D = x * y;
 556     const int E = B + C;
 557     double ftmp[8];
 558     uint64_t tmp[1];
 559     mips_reg addr[1];
 560     DECLARE_VAR_LOW32;
 561
 562     if (D) {
 563         __asm__ volatile (
 564             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 565             "dli        %[tmp0],    0x06                                \n\t"
 566             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 567             "pshufh     %[B],       %[B],           %[ftmp0]            \n\t"
 568             "mtc1       %[tmp0],    %[ftmp7]                            \n\t"
 569             "pshufh     %[C],       %[C],           %[ftmp0]            \n\t"
 570             "pshufh     %[D],       %[D],           %[ftmp0]            \n\t"
 571
 572             "1:                                                         \n\t"
 573             PTR_ADDU   "%[addr0],   %[src],         %[stride]           \n\t"
 574             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 575             MMI_ULWC1(%[ftmp2], %[src], 0x01)
 576             MMI_ULWC1(%[ftmp3], %[addr0], 0x00)
 577             MMI_ULWC1(%[ftmp4], %[addr0], 0x01)
 578
 579             "punpcklbh  %[ftmp5],   %[ftmp1],       %[ftmp0]            \n\t"
 580             "punpcklbh  %[ftmp6],   %[ftmp2],       %[ftmp0]            \n\t"
 581             "pmullh     %[ftmp5],   %[ftmp5],       %[A]                \n\t"
 582             "pmullh     %[ftmp6],   %[ftmp6],       %[B]                \n\t"
 583             "paddh      %[ftmp1],   %[ftmp5],       %[ftmp6]            \n\t"
 584
 585             "punpcklbh  %[ftmp5],   %[ftmp3],       %[ftmp0]            \n\t"
 586             "punpcklbh  %[ftmp6],   %[ftmp4],       %[ftmp0]            \n\t"
 587             "pmullh     %[ftmp5],   %[ftmp5],       %[C]                \n\t"
 588             "pmullh     %[ftmp6],   %[ftmp6],       %[D]                \n\t"
 589             "paddh      %[ftmp2],   %[ftmp5],       %[ftmp6]            \n\t"
 590
 591             "paddh      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 592             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 593             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp7]            \n\t"
 594             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 595             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 596             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 597             "addi       %[h],       %[h],           -0x01               \n\t"
 598             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 599             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 600             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 601             "bnez       %[h],       1b                                  \n\t"
 602             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 603               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 604               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 605               [ftmp6]"=&f"(ftmp[6]),        [ftmp7]"=&f"(ftmp[7]),
 606               [tmp0]"=&r"(tmp[0]),
 607               RESTRICT_ASM_LOW32
 608               [addr0]"=&r"(addr[0]),
 609               [dst]"+&r"(dst),              [src]"+&r"(src),
 610               [h]"+&r"(h)
 611             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 612               [A]"f"(A),                    [B]"f"(B),
 613               [C]"f"(C),                    [D]"f"(D)
 614             : "memory"
 615         );
 616     } else if (E) {
 617         const int step = C ? stride : 1;
 618
 619         __asm__ volatile (
 620             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 621             "dli        %[tmp0],    0x06                                \n\t"
 622             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 623             "pshufh     %[E],       %[E],           %[ftmp0]            \n\t"
 624             "mtc1       %[tmp0],    %[ftmp5]                            \n\t"
 625             "1:                                                         \n\t"
 626             PTR_ADDU   "%[addr0],   %[src],         %[step]             \n\t"
 627             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 628             MMI_ULWC1(%[ftmp2], %[addr0], 0x00)
 629
 630             "punpcklbh  %[ftmp3],   %[ftmp1],       %[ftmp0]            \n\t"
 631             "punpcklbh  %[ftmp4],   %[ftmp2],       %[ftmp0]            \n\t"
 632             "pmullh     %[ftmp3],   %[ftmp3],       %[A]                \n\t"
 633             "pmullh     %[ftmp4],   %[ftmp4],       %[E]                \n\t"
 634             "paddh      %[ftmp1],   %[ftmp3],       %[ftmp4]            \n\t"
 635
 636             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 637             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp5]            \n\t"
 638             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 639             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 640             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 641             "addi       %[h],       %[h],           -0x01               \n\t"
 642             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 643             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 644             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 645             "bnez       %[h],       1b                                  \n\t"
 646             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 647               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 648               [ftmp4]"=&f"(ftmp[4]),        [ftmp5]"=&f"(ftmp[5]),
 649               [tmp0]"=&r"(tmp[0]),
 650               RESTRICT_ASM_LOW32
 651               [addr0]"=&r"(addr[0]),
 652               [dst]"+&r"(dst),              [src]"+&r"(src),
 653               [h]"+&r"(h)
 654             : [stride]"r"((mips_reg)stride),[step]"r"((mips_reg)step),
 655               [ff_pw_32]"f"(ff_pw_32),
 656               [A]"f"(A),                    [E]"f"(E)
 657             : "memory"
 658         );
 659     } else {
 660         __asm__ volatile (
 661             "xor        %[ftmp0],   %[ftmp0],       %[ftmp0]            \n\t"
 662             "dli        %[tmp0],    0x06                                \n\t"
 663             "pshufh     %[A],       %[A],           %[ftmp0]            \n\t"
 664             "mtc1       %[tmp0],    %[ftmp3]                            \n\t"
 665
 666             "1:                                                         \n\t"
 667             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 668             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 669             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 670             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 671             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 672             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 673             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 674             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 675             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 676             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 677             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 678
 679             MMI_ULWC1(%[ftmp1], %[src], 0x00)
 680             "punpcklbh  %[ftmp2],   %[ftmp1],       %[ftmp0]            \n\t"
 681             "pmullh     %[ftmp1],   %[ftmp2],       %[A]                \n\t"
 682             "paddh      %[ftmp1],   %[ftmp1],       %[ff_pw_32]         \n\t"
 683             "psrlh      %[ftmp1],   %[ftmp1],       %[ftmp3]            \n\t"
 684             "packushb   %[ftmp1],   %[ftmp1],       %[ftmp0]            \n\t"
 685             MMI_LWC1(%[ftmp2], %[dst], 0x00)
 686             "pavgb      %[ftmp1],   %[ftmp1],       %[ftmp2]            \n\t"
 687             "addi       %[h],       %[h],           -0x02               \n\t"
 688             MMI_SWC1(%[ftmp1], %[dst], 0x00)
 689
 690             PTR_ADDU   "%[src],     %[src],         %[stride]           \n\t"
 691             PTR_ADDU   "%[dst],     %[dst],         %[stride]           \n\t"
 692             "bnez       %[h],       1b                                  \n\t"
 693             : [ftmp0]"=&f"(ftmp[0]),        [ftmp1]"=&f"(ftmp[1]),
 694               [ftmp2]"=&f"(ftmp[2]),        [ftmp3]"=&f"(ftmp[3]),
 695               [tmp0]"=&r"(tmp[0]),
 696               RESTRICT_ASM_LOW32
 697               [dst]"+&r"(dst),              [src]"+&r"(src),
 698               [h]"+&r"(h)
 699             : [stride]"r"((mips_reg)stride),[ff_pw_32]"f"(ff_pw_32),
 700               [A]"f"(A)
 701             : "memory"
 702         );
 703     }
 704 }