Merge commit 'a7a07cc98ac548297b5b0628cb81280e11952e3f'
[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 #if HAVE_INLINE_ASM
38
39 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
40 //as that would make optimization work hard)
41 #if HAVE_7REGS
42 #define decode_significance decode_significance_x86
43 static int decode_significance_x86(CABACContext *c, int max_coeff,
44                                    uint8_t *significant_coeff_ctx_base,
45                                    int *index, x86_reg last_off){
46     void *end= significant_coeff_ctx_base + max_coeff - 1;
47     int minusstart= -(intptr_t)significant_coeff_ctx_base;
48     int minusindex= 4-(intptr_t)index;
49     int bit;
50     x86_reg coeff_count;
51
52 #ifdef BROKEN_RELOCATIONS
53     void *tables;
54
55     __asm__ volatile(
56         "lea   "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
57         : "=&r"(tables)
58     );
59 #endif
60
61     __asm__ volatile(
62         "3:                                     \n\t"
63
64         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
65                              "%5", "%q5", "%k0", "%b0",
66                              "%c11(%6)", "%c12(%6)",
67                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
68                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
69                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
70                              "%13")
71
72         "test $1, %4                            \n\t"
73         " jz 4f                                 \n\t"
74         "add  %10, %1                           \n\t"
75
76         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
77                              "%5", "%q5", "%k0", "%b0",
78                              "%c11(%6)", "%c12(%6)",
79                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
80                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
81                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
82                              "%13")
83
84         "sub  %10, %1                           \n\t"
85         "mov  %2, %0                            \n\t"
86         "movl %7, %%ecx                         \n\t"
87         "add  %1, %%"REG_c"                     \n\t"
88         "movl %%ecx, (%0)                       \n\t"
89
90         "test $1, %4                            \n\t"
91         " jnz 5f                                \n\t"
92
93         "add"OPSIZE"  $4, %2                    \n\t"
94
95         "4:                                     \n\t"
96         "add  $1, %1                            \n\t"
97         "cmp  %8, %1                            \n\t"
98         " jb 3b                                 \n\t"
99         "mov  %2, %0                            \n\t"
100         "movl %7, %%ecx                         \n\t"
101         "add  %1, %%"REG_c"                     \n\t"
102         "movl %%ecx, (%0)                       \n\t"
103         "5:                                     \n\t"
104         "add  %9, %k0                           \n\t"
105         "shr $2, %k0                            \n\t"
106         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
107           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
108         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
109           "i"(offsetof(CABACContext, bytestream)),
110           "i"(offsetof(CABACContext, bytestream_end))
111           TABLES_ARG
112         : "%"REG_c, "memory"
113     );
114     return coeff_count;
115 }
116
117 #define decode_significance_8x8 decode_significance_8x8_x86
118 static int decode_significance_8x8_x86(CABACContext *c,
119                                        uint8_t *significant_coeff_ctx_base,
120                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
121     int minusindex= 4-(intptr_t)index;
122     int bit;
123     x86_reg coeff_count;
124     x86_reg last=0;
125     x86_reg state;
126
127 #ifdef BROKEN_RELOCATIONS
128     void *tables;
129
130     __asm__ volatile(
131         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
132         : "=&r"(tables)
133     );
134 #endif
135
136     __asm__ volatile(
137         "mov %1, %6                             \n\t"
138         "3:                                     \n\t"
139
140         "mov %10, %0                            \n\t"
141         "movzbl (%0, %6), %k6                   \n\t"
142         "add %9, %6                             \n\t"
143
144         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
145                              "%5", "%q5", "%k0", "%b0",
146                              "%c12(%7)", "%c13(%7)",
147                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
148                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
149                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
150                              "%15")
151
152         "mov %1, %k6                            \n\t"
153         "test $1, %4                            \n\t"
154         " jz 4f                                 \n\t"
155
156 #ifdef BROKEN_RELOCATIONS
157         "movzbl %c14(%15, %q6), %k6\n\t"
158 #else
159         "movzbl "MANGLE(ff_h264_cabac_tables)"+%c14(%k6), %k6\n\t"
160 #endif
161         "add %11, %6                            \n\t"
162
163         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
164                              "%5", "%q5", "%k0", "%b0",
165                              "%c12(%7)", "%c13(%7)",
166                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
167                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
168                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
169                              "%15")
170
171         "mov %2, %0                             \n\t"
172         "mov %1, %k6                            \n\t"
173         "movl %k6, (%0)                         \n\t"
174
175         "test $1, %4                            \n\t"
176         " jnz 5f                                \n\t"
177
178         "add"OPSIZE"  $4, %2                    \n\t"
179
180         "4:                                     \n\t"
181         "addl $1, %k6                           \n\t"
182         "mov %k6, %1                            \n\t"
183         "cmpl $63, %k6                          \n\t"
184         " jb 3b                                 \n\t"
185         "mov %2, %0                             \n\t"
186         "movl %k6, (%0)                         \n\t"
187         "5:                                     \n\t"
188         "addl %8, %k0                           \n\t"
189         "shr $2, %k0                            \n\t"
190         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
191           "=&r"(bit), "+&r"(c->range), "=&r"(state)
192         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
193           "m"(sig_off), "m"(last_coeff_ctx_base),
194           "i"(offsetof(CABACContext, bytestream)),
195           "i"(offsetof(CABACContext, bytestream_end)),
196           "i"(H264_LAST_COEFF_FLAG_OFFSET_8x8_OFFSET) TABLES_ARG
197         : "%"REG_c, "memory"
198     );
199     return coeff_count;
200 }
201 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
202
203 #endif /* HAVE_INLINE_ASM */
204 #endif /* AVCODEC_X86_H264_I386_H */