h264: use one table instead of several for cabac functions
[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 Libav.
6  *
7  * Libav 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  * Libav 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 Libav; 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 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
49     __asm__ volatile(
50         "3:                                     \n\t"
51
52         BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
53                              "%5", "%k0", "%b0",
54                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15")
55
56         "test $1, %4                            \n\t"
57         " jz 4f                                 \n\t"
58         "add  %10, %1                           \n\t"
59
60         BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
61                              "%5", "%k0", "%b0",
62                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15")
63
64         "sub  %10, %1                           \n\t"
65         "mov  %2, %0                            \n\t"
66         "movl %7, %%ecx                         \n\t"
67         "add  %1, %%"REG_c"                     \n\t"
68         "movl %%ecx, (%0)                       \n\t"
69
70         "test $1, %4                            \n\t"
71         " jnz 5f                                \n\t"
72
73         "add"OPSIZE"  $4, %2                    \n\t"
74
75         "4:                                     \n\t"
76         "add  $1, %1                            \n\t"
77         "cmp  %8, %1                            \n\t"
78         " jb 3b                                 \n\t"
79         "mov  %2, %0                            \n\t"
80         "movl %7, %%ecx                         \n\t"
81         "add  %1, %%"REG_c"                     \n\t"
82         "movl %%ecx, (%0)                       \n\t"
83         "5:                                     \n\t"
84         "add  %9, %k0                           \n\t"
85         "shr $2, %k0                            \n\t"
86         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
87           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
88         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
89           "i"(offsetof(CABACContext, bytestream)),
90           "i"(offsetof(CABACContext, bytestream_end)),
91           "i"(H264_NORM_SHIFT_OFFSET),
92           "i"(H264_LPS_RANGE_OFFSET),
93           "i"(H264_MLPS_STATE_OFFSET)
94         : "%"REG_c, "memory"
95     );
96     return coeff_count;
97 }
98
99 static int decode_significance_8x8_x86(CABACContext *c,
100                                        uint8_t *significant_coeff_ctx_base,
101                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
102     int minusindex= 4-(intptr_t)index;
103     int bit;
104     x86_reg coeff_count;
105     x86_reg last=0;
106     x86_reg state;
107
108     __asm__ volatile(
109         "mov %1, %6                             \n\t"
110         "3:                                     \n\t"
111
112         "mov %10, %0                            \n\t"
113         "movzbl (%0, %6), %k6                   \n\t"
114         "add %9, %6                             \n\t"
115
116         BRANCHLESS_GET_CABAC("%4", "(%6)", "%3", "%w3",
117                              "%5", "%k0", "%b0",
118                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16")
119
120         "mov %1, %k6                            \n\t"
121         "test $1, %4                            \n\t"
122         " jz 4f                                 \n\t"
123
124         "movzbl "MANGLE(ff_h264_cabac_tables)"+%a17(%k6), %k6\n\t"
125
126         "add %11, %6                            \n\t"
127
128         BRANCHLESS_GET_CABAC("%4", "(%6)", "%3", "%w3",
129                              "%5", "%k0", "%b0",
130                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16")
131
132         "mov %2, %0                             \n\t"
133         "mov %1, %k6                            \n\t"
134         "movl %k6, (%0)                         \n\t"
135
136         "test $1, %4                            \n\t"
137         " jnz 5f                                \n\t"
138
139         "add"OPSIZE"  $4, %2                    \n\t"
140
141         "4:                                     \n\t"
142         "addl $1, %k6                           \n\t"
143         "mov %k6, %1                            \n\t"
144         "cmpl $63, %k6                          \n\t"
145         " jb 3b                                 \n\t"
146         "mov %2, %0                             \n\t"
147         "movl %k6, (%0)                         \n\t"
148         "5:                                     \n\t"
149         "addl %8, %k0                           \n\t"
150         "shr $2, %k0                            \n\t"
151         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
152           "=&r"(bit), "+&r"(c->range), "=&r"(state)
153         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
154           "m"(sig_off), "m"(last_coeff_ctx_base),
155           "i"(offsetof(CABACContext, bytestream)),
156           "i"(offsetof(CABACContext, bytestream_end)),
157           "i"(H264_NORM_SHIFT_OFFSET),
158           "i"(H264_LPS_RANGE_OFFSET),
159           "i"(H264_MLPS_STATE_OFFSET),
160           "i"(H264_LAST_COEFF_FLAG_OFFSET_8x8_OFFSET)
161         : "%"REG_c, "memory"
162     );
163     return coeff_count;
164 }
165 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
166
167 #endif /* AVCODEC_X86_H264_I386_H */