Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / libavcodec / x86 / h264_i386.h
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * H.264 / AVC / MPEG4 part10 codec.
25  * non-MMX i386-specific optimizations for H.264
26  * @author Michael Niedermayer <michaelni@gmx.at>
27  */
28
29 #ifndef AVCODEC_X86_H264_I386_H
30 #define AVCODEC_X86_H264_I386_H
31
32 #include <stddef.h>
33
34 #include "libavcodec/cabac.h"
35 #include "cabac.h"
36
37 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
38 //as that would make optimization work hard)
39 #if ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS)
40 static int decode_significance_x86(CABACContext *c, int max_coeff,
41                                    uint8_t *significant_coeff_ctx_base,
42                                    int *index, x86_reg last_off){
43     void *end= significant_coeff_ctx_base + max_coeff - 1;
44     int minusstart= -(intptr_t)significant_coeff_ctx_base;
45     int minusindex= 4-(intptr_t)index;
46     int bit;
47     x86_reg coeff_count;
48     __asm__ volatile(
49         "2:                                     \n\t"
50
51         BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
52                              "%w3", "%5", "%k0", "%b0", "%a11")
53
54         "test $1, %4                            \n\t"
55         " jz 3f                                 \n\t"
56         "add  %10, %1                           \n\t"
57
58         BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
59                              "%w3", "%5", "%k0", "%b0", "%a11")
60
61         "sub  %10, %1                           \n\t"
62         "mov  %2, %0                            \n\t"
63         "movl %7, %%ecx                         \n\t"
64         "add  %1, %%"REG_c"                     \n\t"
65         "movl %%ecx, (%0)                       \n\t"
66
67         "test $1, %4                            \n\t"
68         " jnz 4f                                \n\t"
69
70         "add"OPSIZE"  $4, %2                    \n\t"
71
72         "3:                                     \n\t"
73         "add  $1, %1                            \n\t"
74         "cmp  %8, %1                            \n\t"
75         " jb 2b                                 \n\t"
76         "mov  %2, %0                            \n\t"
77         "movl %7, %%ecx                         \n\t"
78         "add  %1, %%"REG_c"                     \n\t"
79         "movl %%ecx, (%0)                       \n\t"
80         "4:                                     \n\t"
81         "add  %9, %k0                           \n\t"
82         "shr $2, %k0                            \n\t"
83         :"=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
84          "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
85         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
86          "i"(offsetof(CABACContext, bytestream))
87         : "%"REG_c, "memory"
88     );
89     return coeff_count;
90 }
91
92 static int decode_significance_8x8_x86(CABACContext *c,
93                                        uint8_t *significant_coeff_ctx_base,
94                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
95     int minusindex= 4-(intptr_t)index;
96     int bit;
97     x86_reg coeff_count;
98     x86_reg last=0;
99     x86_reg state;
100     __asm__ volatile(
101         "mov %1, %6                             \n\t"
102         "2:                                     \n\t"
103
104         "mov %10, %0                            \n\t"
105         "movzbl (%0, %6), %k6                   \n\t"
106         "add %9, %6                             \n\t"
107
108         BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
109                              "%w3", "%5", "%k0", "%b0", "%a12")
110
111         "mov %1, %k6                            \n\t"
112         "test $1, %4                            \n\t"
113         " jz 3f                                 \n\t"
114
115         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%k6), %k6\n\t"
116         "add %11, %6                            \n\t"
117
118         BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
119                              "%w3", "%5", "%k0", "%b0", "%a12")
120
121         "mov %2, %0                             \n\t"
122         "mov %1, %k6                            \n\t"
123         "movl %k6, (%0)                         \n\t"
124
125         "test $1, %4                            \n\t"
126         " jnz 4f                                \n\t"
127
128         "add"OPSIZE"  $4, %2                    \n\t"
129
130         "3:                                     \n\t"
131         "addl $1, %k6                           \n\t"
132         "mov %k6, %1                            \n\t"
133         "cmpl $63, %k6                          \n\t"
134         " jb 2b                                 \n\t"
135         "mov %2, %0                             \n\t"
136         "movl %k6, (%0)                         \n\t"
137         "4:                                     \n\t"
138         "addl %8, %k0                           \n\t"
139         "shr $2, %k0                            \n\t"
140         :"=&q"(coeff_count),"+m"(last), "+m"(index), "+&r"(c->low), "=&r"(bit),
141          "+&r"(c->range), "=&r"(state)
142         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_coeff_ctx_base),
143          "i"(offsetof(CABACContext, bytestream))
144         : "%"REG_c, "memory"
145     );
146     return coeff_count;
147 }
148 #endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
149
150 #endif /* AVCODEC_X86_H264_I386_H */