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, cabac, statep, low, lowword, 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, cabac, statep, low, lowword, 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, cabac, statep, low, lowword, range, tmp, tmpbyte, byte) \
  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, cabac, statep, low, lowword,        \
  61                                     range, tmp)                              \
  62         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
  63         "shl    %%cl        , "range"                                   \n\t"\
  64         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
  65         "shl    %%cl        , "low"                                     \n\t"\
  66         "mov    "tmpbyte"   , "statep"                                  \n\t"\
  67         "test   "lowword"   , "lowword"                                 \n\t"\
  68         " jnz   1f                                                      \n\t"\
  69         "mov "byte"("cabac"), %%"REG_c"                                 \n\t"\
  70         "add"OPSIZE" $2     , "byte    "("cabac")                       \n\t"\
  71         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
  72         "lea    -1("low")   , %%ecx                                     \n\t"\
  73         "xor    "low"       , %%ecx                                     \n\t"\
  74         "shr    $15         , %%ecx                                     \n\t"\
  75         "bswap  "tmp"                                                   \n\t"\
  76         "shr    $15         , "tmp"                                     \n\t"\
  77         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
  78         "sub    $0xFFFF     , "tmp"                                     \n\t"\
  79         "neg    %%ecx                                                   \n\t"\
  80         "add    $7          , %%ecx                                     \n\t"\
  81         "shl    %%cl        , "tmp"                                     \n\t"\
  82         "add    "tmp"       , "low"                                     \n\t"\
  83         "1:                                                             \n\t"
  84
  85 #if ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
  86 #define get_cabac_inline get_cabac_inline_x86
  87 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
  88                                                  uint8_t *const state)
  89 {
  90     int bit, tmp;
  91
  92     __asm__ volatile(
  93         BRANCHLESS_GET_CABAC("%0", "%5", "(%4)", "%1", "%w1", "%2",
  94                              "%3", "%b3", "%a6")
  95         :"=&r"(bit), "+&r"(c->low), "+&r"(c->range), "=&q"(tmp)
  96         :"r"(state), "r"(c),
  97          "i"(offsetof(CABACContext, bytestream))
  98         : "%"REG_c, "memory"
  99     );
 100     return bit & 1;
 101 }
 102 #endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 103
 104 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
 105 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
 106 {
 107     x86_reg tmp;
 108     __asm__ volatile(
 109         "movl %a3(%2), %k1                      \n\t"
 110         "movl %a4(%2), %%eax                    \n\t"
 111         "shl $17, %k1                           \n\t"
 112         "add %%eax, %%eax                       \n\t"
 113         "sub %k1, %%eax                         \n\t"
 114         "cltd                                   \n\t"
 115         "and %%edx, %k1                         \n\t"
 116         "add %k1, %%eax                         \n\t"
 117         "xor %%edx, %%ecx                       \n\t"
 118         "sub %%edx, %%ecx                       \n\t"
 119         "test %%ax, %%ax                        \n\t"
 120         " jnz 1f                                \n\t"
 121         "mov  %a5(%2), %1                       \n\t"
 122         "subl $0xFFFF, %%eax                    \n\t"
 123         "movzwl (%1), %%edx                     \n\t"
 124         "bswap %%edx                            \n\t"
 125         "shrl $15, %%edx                        \n\t"
 126         "add  $2, %1                            \n\t"
 127         "addl %%edx, %%eax                      \n\t"
 128         "mov  %1, %a5(%2)                       \n\t"
 129         "1:                                     \n\t"
 130         "movl %%eax, %a4(%2)                    \n\t"
 131
 132         :"+c"(val), "=&r"(tmp)
 133         :"r"(c),
 134          "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
 135          "i"(offsetof(CABACContext, bytestream))
 136         : "%eax", "%edx", "memory"
 137     );
 138     return val;
 139 }
 140
 141 #endif /* AVCODEC_X86_CABAC_H */