H.264: optimize CABAC x86 asm for Atom
[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 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     int low;
49     int range;
50     __asm__ volatile(
51         "movl %a11(%6), %5                      \n\t"
52         "movl %a12(%6), %3                      \n\t"
53
54         "2:                                     \n\t"
55
56         BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
57                              "%w3", "%5", "%k0", "%b0", "%a13")
58
59         "test $1, %4                            \n\t"
60         " jz 3f                                 \n\t"
61         "add  %10, %1                           \n\t"
62
63         BRANCHLESS_GET_CABAC("%4", "%6", "(%1)", "%3",
64                              "%w3", "%5", "%k0", "%b0", "%a13")
65
66         "sub  %10, %1                           \n\t"
67         "mov  %2, %0                            \n\t"
68         "movl %7, %%ecx                         \n\t"
69         "add  %1, %%"REG_c"                     \n\t"
70         "movl %%ecx, (%0)                       \n\t"
71
72         "test $1, %4                            \n\t"
73         " jnz 4f                                \n\t"
74
75         "add  $4, %2                            \n\t"
76
77         "3:                                     \n\t"
78         "add  $1, %1                            \n\t"
79         "cmp  %8, %1                            \n\t"
80         " jb 2b                                 \n\t"
81         "mov  %2, %0                            \n\t"
82         "movl %7, %%ecx                         \n\t"
83         "add  %1, %%"REG_c"                     \n\t"
84         "movl %%ecx, (%0)                       \n\t"
85         "4:                                     \n\t"
86         "add  %9, %k0                           \n\t"
87         "shr $2, %k0                            \n\t"
88
89         "movl %5, %a11(%6)                      \n\t"
90         "movl %3, %a12(%6)                      \n\t"
91         :"=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
92          "=&r"(low), "=&r"(bit), "=&r"(range)
93         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
94          "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
95          "i"(offsetof(CABACContext, bytestream))
96         : "%"REG_c, "memory"
97     );
98     return coeff_count;
99 }
100
101 static int decode_significance_8x8_x86(CABACContext *c,
102                                        uint8_t *significant_coeff_ctx_base,
103                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
104     int minusindex= 4-(intptr_t)index;
105     int bit;
106     x86_reg coeff_count;
107     int low;
108     int range;
109     x86_reg last=0;
110     x86_reg state;
111     __asm__ volatile(
112         "movl %a12(%7), %5                      \n\t"
113         "movl %a13(%7), %3                      \n\t"
114
115         "mov %1, %6                             \n\t"
116         "2:                                     \n\t"
117
118         "mov %10, %0                            \n\t"
119         "movzbl (%0, %6), %k6                   \n\t"
120         "add %9, %6                             \n\t"
121
122         BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
123                              "%w3", "%5", "%k0", "%b0", "%a14")
124
125         "mov %1, %k6                            \n\t"
126         "test $1, %4                            \n\t"
127         " jz 3f                                 \n\t"
128
129         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%k6), %k6\n\t"
130         "add %11, %6                            \n\t"
131
132         BRANCHLESS_GET_CABAC("%4", "%7", "(%6)", "%3",
133                              "%w3", "%5", "%k0", "%b0", "%a14")
134
135         "mov %2, %0                             \n\t"
136         "mov %1, %k6                            \n\t"
137         "movl %k6, (%0)                         \n\t"
138
139         "test $1, %4                            \n\t"
140         " jnz 4f                                \n\t"
141
142         "add  $4, %2                            \n\t"
143
144         "3:                                     \n\t"
145         "addl $1, %k6                           \n\t"
146         "mov %k6, %1                            \n\t"
147         "cmpl $63, %k6                          \n\t"
148         " jb 2b                                 \n\t"
149         "mov %2, %0                             \n\t"
150         "movl %k6, (%0)                         \n\t"
151         "4:                                     \n\t"
152         "addl %8, %k0                           \n\t"
153         "shr $2, %k0                            \n\t"
154
155         "movl %5, %a12(%7)                      \n\t"
156         "movl %3, %a13(%7)                      \n\t"
157         :"=&q"(coeff_count),"+m"(last), "+m"(index), "=&r"(low), "=&r"(bit),
158          "=&r"(range), "=&r"(state)
159         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_coeff_ctx_base),
160          "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)),
161          "i"(offsetof(CABACContext, bytestream))
162         : "%"REG_c, "memory"
163     );
164     return coeff_count;
165 }
166 #endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
167
168 #endif /* AVCODEC_X86_H264_I386_H */