fft: port FFT/IMDCT 3dnow functions to yasm, and disable on x86-64.
[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 #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                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15", "%16")
67
68         "test $1, %4                            \n\t"
69         " jz 4f                                 \n\t"
70         "add  %10, %1                           \n\t"
71
72         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
73                              "%5", "%q5", "%k0", "%b0",
74                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15", "%16")
75
76         "sub  %10, %1                           \n\t"
77         "mov  %2, %0                            \n\t"
78         "movl %7, %%ecx                         \n\t"
79         "add  %1, %%"REG_c"                     \n\t"
80         "movl %%ecx, (%0)                       \n\t"
81
82         "test $1, %4                            \n\t"
83         " jnz 5f                                \n\t"
84
85         "add"OPSIZE"  $4, %2                    \n\t"
86
87         "4:                                     \n\t"
88         "add  $1, %1                            \n\t"
89         "cmp  %8, %1                            \n\t"
90         " jb 3b                                 \n\t"
91         "mov  %2, %0                            \n\t"
92         "movl %7, %%ecx                         \n\t"
93         "add  %1, %%"REG_c"                     \n\t"
94         "movl %%ecx, (%0)                       \n\t"
95         "5:                                     \n\t"
96         "add  %9, %k0                           \n\t"
97         "shr $2, %k0                            \n\t"
98         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
99           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
100         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
101           "i"(offsetof(CABACContext, bytestream)),
102           "i"(offsetof(CABACContext, bytestream_end)),
103           "i"(H264_NORM_SHIFT_OFFSET),
104           "i"(H264_LPS_RANGE_OFFSET),
105           "i"(H264_MLPS_STATE_OFFSET) TABLES_ARG
106         : "%"REG_c, "memory"
107     );
108     return coeff_count;
109 }
110
111 #define decode_significance_8x8 decode_significance_8x8_x86
112 static int decode_significance_8x8_x86(CABACContext *c,
113                                        uint8_t *significant_coeff_ctx_base,
114                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
115     int minusindex= 4-(intptr_t)index;
116     int bit;
117     x86_reg coeff_count;
118     x86_reg last=0;
119     x86_reg state;
120
121 #ifdef BROKEN_RELOCATIONS
122     void *tables;
123
124     __asm__ volatile(
125         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
126         : "=&r"(tables)
127     );
128 #endif
129
130     __asm__ volatile(
131         "mov %1, %6                             \n\t"
132         "3:                                     \n\t"
133
134         "mov %10, %0                            \n\t"
135         "movzbl (%0, %6), %k6                   \n\t"
136         "add %9, %6                             \n\t"
137
138         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
139                              "%5", "%q5", "%k0", "%b0",
140                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16", "%18")
141
142         "mov %1, %k6                            \n\t"
143         "test $1, %4                            \n\t"
144         " jz 4f                                 \n\t"
145
146 #ifdef BROKEN_RELOCATIONS
147         "movzbl %a17(%18, %q6), %k6\n\t"
148 #else
149         "movzbl "MANGLE(ff_h264_cabac_tables)"+%a17(%k6), %k6\n\t"
150 #endif
151         "add %11, %6                            \n\t"
152
153         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
154                              "%5", "%q5", "%k0", "%b0",
155                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16", "%18")
156
157         "mov %2, %0                             \n\t"
158         "mov %1, %k6                            \n\t"
159         "movl %k6, (%0)                         \n\t"
160
161         "test $1, %4                            \n\t"
162         " jnz 5f                                \n\t"
163
164         "add"OPSIZE"  $4, %2                    \n\t"
165
166         "4:                                     \n\t"
167         "addl $1, %k6                           \n\t"
168         "mov %k6, %1                            \n\t"
169         "cmpl $63, %k6                          \n\t"
170         " jb 3b                                 \n\t"
171         "mov %2, %0                             \n\t"
172         "movl %k6, (%0)                         \n\t"
173         "5:                                     \n\t"
174         "addl %8, %k0                           \n\t"
175         "shr $2, %k0                            \n\t"
176         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
177           "=&r"(bit), "+&r"(c->range), "=&r"(state)
178         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
179           "m"(sig_off), "m"(last_coeff_ctx_base),
180           "i"(offsetof(CABACContext, bytestream)),
181           "i"(offsetof(CABACContext, bytestream_end)),
182           "i"(H264_NORM_SHIFT_OFFSET),
183           "i"(H264_LPS_RANGE_OFFSET),
184           "i"(H264_MLPS_STATE_OFFSET),
185           "i"(H264_LAST_COEFF_FLAG_OFFSET_8x8_OFFSET) TABLES_ARG
186         : "%"REG_c, "memory"
187     );
188     return coeff_count;
189 }
190 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
191
192 #endif /* HAVE_INLINE_ASM */
193 #endif /* AVCODEC_X86_H264_I386_H */