git.sesse.net Git - ffmpeg/blob - libavcodec/x86/cabac.h

   1 /*
   2  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   3  *
   4  * This file is part of Libav.
   5  *
   6  * Libav is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * Libav is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with Libav; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #ifndef AVCODEC_X86_CABAC_H
  22 #define AVCODEC_X86_CABAC_H
  23
  24 #include "libavcodec/cabac.h"
  25 #include "libavutil/attributes.h"
  26 #include "libavutil/x86_cpu.h"
  27 #include "config.h"
  28
  29 #if HAVE_FAST_CMOV
  30 #define BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, range, tmp)\
  31         "mov    "tmp"       , %%ecx     \n\t"\
  32         "shl    $17         , "tmp"     \n\t"\
  33         "cmp    "low"       , "tmp"     \n\t"\
  34         "cmova  %%ecx       , "range"   \n\t"\
  35         "sbb    %%ecx       , %%ecx     \n\t"\
  36         "and    %%ecx       , "tmp"     \n\t"\
  37         "xor    %%ecx       , "ret"     \n\t"\
  38         "sub    "tmp"       , "low"     \n\t"
  39 #else /* HAVE_FAST_CMOV */
  40 #define BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, range, tmp)\
  41         "mov    "tmp"       , %%ecx     \n\t"\
  42         "shl    $17         , "tmp"     \n\t"\
  43         "sub    "low"       , "tmp"     \n\t"\
  44         "sar    $31         , "tmp"     \n\t" /*lps_mask*/\
  45         "sub    %%ecx       , "range"   \n\t" /*RangeLPS - range*/\
  46         "and    "tmp"       , "range"   \n\t" /*(RangeLPS - range)&lps_mask*/\
  47         "add    %%ecx       , "range"   \n\t" /*new range*/\
  48         "shl    $17         , %%ecx     \n\t"\
  49         "and    "tmp"       , %%ecx     \n\t"\
  50         "sub    %%ecx       , "low"     \n\t"\
  51         "xor    "tmp"       , "ret"     \n\t"
  52 #endif /* HAVE_FAST_CMOV */
  53
  54 #define BRANCHLESS_GET_CABAC(ret, statep, low, lowword, range, tmp, tmpbyte, byte, end) \
  55         "movzbl "statep"    , "ret"                                     \n\t"\
  56         "mov    "range"     , "tmp"                                     \n\t"\
  57         "and    $0xC0       , "range"                                   \n\t"\
  58         "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
  59         "sub    "range"     , "tmp"                                     \n\t"\
  60         BRANCHLESS_GET_CABAC_UPDATE(ret, statep, low, range, tmp)   \
  61         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
  62         "shl    %%cl        , "range"                                   \n\t"\
  63         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
  64         "shl    %%cl        , "low"                                     \n\t"\
  65         "mov    "tmpbyte"   , "statep"                                  \n\t"\
  66         "test   "lowword"   , "lowword"                                 \n\t"\
  67         " jnz   2f                                                      \n\t"\
  68         "mov    "byte"      , %%"REG_c"                                 \n\t"\
  69         "cmp    "end"       , %%"REG_c"                                 \n\t"\
  70         "jge    1f                                                      \n\t"\
  71         "add"OPSIZE" $2     , "byte"                                    \n\t"\
  72         "1:                                                             \n\t"\
  73         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
  74         "lea    -1("low")   , %%ecx                                     \n\t"\
  75         "xor    "low"       , %%ecx                                     \n\t"\
  76         "shr    $15         , %%ecx                                     \n\t"\
  77         "bswap  "tmp"                                                   \n\t"\
  78         "shr    $15         , "tmp"                                     \n\t"\
  79         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
  80         "sub    $0xFFFF     , "tmp"                                     \n\t"\
  81         "neg    %%ecx                                                   \n\t"\
  82         "add    $7          , %%ecx                                     \n\t"\
  83         "shl    %%cl        , "tmp"                                     \n\t"\
  84         "add    "tmp"       , "low"                                     \n\t"\
  85         "2:                                                             \n\t"
  86
  87 #if HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
  88 #define get_cabac_inline get_cabac_inline_x86
  89 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
  90                                                  uint8_t *const state)
  91 {
  92     int bit, tmp;
  93
  94     __asm__ volatile(
  95         BRANCHLESS_GET_CABAC("%0", "(%4)", "%1", "%w1",
  96                              "%2", "%3", "%b3",
  97                              "%a6(%5)", "%a7(%5)")
  98         : "=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp)
  99         : "r"(state), "r"(c),
 100           "i"(offsetof(CABACContext, bytestream)),
 101           "i"(offsetof(CABACContext, bytestream_end))
 102         : "%"REG_c, "memory"
 103     );
 104     return bit & 1;
 105 }
 106 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 107
 108 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
 109 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
 110 {
 111     x86_reg tmp;
 112     __asm__ volatile(
 113         "movl        %a6(%2), %k1       \n\t"
 114         "movl        %a3(%2), %%eax     \n\t"
 115         "shl             $17, %k1       \n\t"
 116         "add           %%eax, %%eax     \n\t"
 117         "sub             %k1, %%eax     \n\t"
 118         "cltd                           \n\t"
 119         "and           %%edx, %k1       \n\t"
 120         "add             %k1, %%eax     \n\t"
 121         "xor           %%edx, %%ecx     \n\t"
 122         "sub           %%edx, %%ecx     \n\t"
 123         "test           %%ax, %%ax      \n\t"
 124         "jnz              1f            \n\t"
 125         "mov         %a4(%2), %1        \n\t"
 126         "subl        $0xFFFF, %%eax     \n\t"
 127         "movzwl         (%1), %%edx     \n\t"
 128         "bswap         %%edx            \n\t"
 129         "shrl            $15, %%edx     \n\t"
 130         "addl          %%edx, %%eax     \n\t"
 131         "cmp         %a5(%2), %1        \n\t"
 132         "jge              1f            \n\t"
 133         "add"OPSIZE"      $2, %a4(%2)   \n\t"
 134         "1:                             \n\t"
 135         "movl          %%eax, %a3(%2)   \n\t"
 136
 137         : "+c"(val), "=&r"(tmp)
 138         : "r"(c),
 139           "i"(offsetof(CABACContext, low)),
 140           "i"(offsetof(CABACContext, bytestream)),
 141           "i"(offsetof(CABACContext, bytestream_end)),
 142           "i"(offsetof(CABACContext, range))
 143         : "%eax", "%edx", "memory"
 144     );
 145     return val;
 146 }
 147
 148 #endif /* AVCODEC_X86_CABAC_H */