yading@10
|
1 /*
|
yading@10
|
2 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
|
yading@10
|
3 *
|
yading@10
|
4 * This file is part of Libav.
|
yading@10
|
5 *
|
yading@10
|
6 * Libav is free software; you can redistribute it and/or
|
yading@10
|
7 * modify it under the terms of the GNU Lesser General Public
|
yading@10
|
8 * License as published by the Free Software Foundation; either
|
yading@10
|
9 * version 2.1 of the License, or (at your option) any later version.
|
yading@10
|
10 *
|
yading@10
|
11 * Libav is distributed in the hope that it will be useful,
|
yading@10
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
yading@10
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
yading@10
|
14 * Lesser General Public License for more details.
|
yading@10
|
15 *
|
yading@10
|
16 * You should have received a copy of the GNU Lesser General Public
|
yading@10
|
17 * License along with Libav; if not, write to the Free Software
|
yading@10
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
yading@10
|
19 */
|
yading@10
|
20
|
yading@10
|
21 #ifndef AVCODEC_X86_CABAC_H
|
yading@10
|
22 #define AVCODEC_X86_CABAC_H
|
yading@10
|
23
|
yading@10
|
24 #include "libavcodec/cabac.h"
|
yading@10
|
25 #include "libavutil/attributes.h"
|
yading@10
|
26 #include "libavutil/x86/asm.h"
|
yading@10
|
27 #include "libavutil/internal.h"
|
yading@10
|
28 #include "config.h"
|
yading@10
|
29
|
yading@10
|
30 #if HAVE_INLINE_ASM
|
yading@10
|
31
|
yading@10
|
32 #ifdef BROKEN_RELOCATIONS
|
yading@10
|
33 #define TABLES_ARG , "r"(tables)
|
yading@10
|
34
|
yading@10
|
35 #if HAVE_FAST_CMOV
|
yading@10
|
36 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
|
yading@10
|
37 "cmp "low" , "tmp" \n\t"\
|
yading@10
|
38 "cmova %%ecx , "range" \n\t"\
|
yading@10
|
39 "sbb %%rcx , %%rcx \n\t"\
|
yading@10
|
40 "and %%ecx , "tmp" \n\t"\
|
yading@10
|
41 "xor %%rcx , "retq" \n\t"\
|
yading@10
|
42 "sub "tmp" , "low" \n\t"
|
yading@10
|
43 #else /* HAVE_FAST_CMOV */
|
yading@10
|
44 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
|
yading@10
|
45 /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \
|
yading@10
|
46 "sub "low" , "tmp" \n\t"\
|
yading@10
|
47 "sar $31 , "tmp" \n\t"\
|
yading@10
|
48 "sub %%ecx , "range" \n\t"\
|
yading@10
|
49 "and "tmp" , "range" \n\t"\
|
yading@10
|
50 "add %%ecx , "range" \n\t"\
|
yading@10
|
51 "shl $17 , %%ecx \n\t"\
|
yading@10
|
52 "and "tmp" , %%ecx \n\t"\
|
yading@10
|
53 "sub %%ecx , "low" \n\t"\
|
yading@10
|
54 "xor "tmp" , "ret" \n\t"\
|
yading@10
|
55 "movslq "ret" , "retq" \n\t"
|
yading@10
|
56 #endif /* HAVE_FAST_CMOV */
|
yading@10
|
57
|
yading@10
|
58 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
|
yading@10
|
59 "movzbl "statep" , "ret" \n\t"\
|
yading@10
|
60 "mov "range" , "tmp" \n\t"\
|
yading@10
|
61 "and $0xC0 , "range" \n\t"\
|
yading@10
|
62 "lea ("ret", "range", 2), %%ecx \n\t"\
|
yading@10
|
63 "movzbl "lps_off"("tables", %%rcx), "range" \n\t"\
|
yading@10
|
64 "sub "range" , "tmp" \n\t"\
|
yading@10
|
65 "mov "tmp" , %%ecx \n\t"\
|
yading@10
|
66 "shl $17 , "tmp" \n\t"\
|
yading@10
|
67 BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
|
yading@10
|
68 "movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t"\
|
yading@10
|
69 "shl %%cl , "range" \n\t"\
|
yading@10
|
70 "movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\
|
yading@10
|
71 "shl %%cl , "low" \n\t"\
|
yading@10
|
72 "mov "tmpbyte" , "statep" \n\t"\
|
yading@10
|
73 "test "lowword" , "lowword" \n\t"\
|
yading@10
|
74 "jnz 2f \n\t"\
|
yading@10
|
75 "mov "byte" , %%"REG_c" \n\t"\
|
yading@10
|
76 "add"OPSIZE" $2 , "byte" \n\t"\
|
yading@10
|
77 "movzwl (%%"REG_c") , "tmp" \n\t"\
|
yading@10
|
78 "lea -1("low") , %%ecx \n\t"\
|
yading@10
|
79 "xor "low" , %%ecx \n\t"\
|
yading@10
|
80 "shr $15 , %%ecx \n\t"\
|
yading@10
|
81 "bswap "tmp" \n\t"\
|
yading@10
|
82 "shr $15 , "tmp" \n\t"\
|
yading@10
|
83 "movzbl "norm_off"("tables", %%rcx), %%ecx \n\t"\
|
yading@10
|
84 "sub $0xFFFF , "tmp" \n\t"\
|
yading@10
|
85 "neg %%ecx \n\t"\
|
yading@10
|
86 "add $7 , %%ecx \n\t"\
|
yading@10
|
87 "shl %%cl , "tmp" \n\t"\
|
yading@10
|
88 "add "tmp" , "low" \n\t"\
|
yading@10
|
89 "2: \n\t"
|
yading@10
|
90
|
yading@10
|
91 #else /* BROKEN_RELOCATIONS */
|
yading@10
|
92 #define TABLES_ARG
|
yading@10
|
93 #define RIP_ARG
|
yading@10
|
94
|
yading@10
|
95 #if HAVE_FAST_CMOV
|
yading@10
|
96 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
|
yading@10
|
97 "mov "tmp" , %%ecx \n\t"\
|
yading@10
|
98 "shl $17 , "tmp" \n\t"\
|
yading@10
|
99 "cmp "low" , "tmp" \n\t"\
|
yading@10
|
100 "cmova %%ecx , "range" \n\t"\
|
yading@10
|
101 "sbb %%ecx , %%ecx \n\t"\
|
yading@10
|
102 "and %%ecx , "tmp" \n\t"\
|
yading@10
|
103 "xor %%ecx , "ret" \n\t"\
|
yading@10
|
104 "sub "tmp" , "low" \n\t"
|
yading@10
|
105 #else /* HAVE_FAST_CMOV */
|
yading@10
|
106 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
|
yading@10
|
107 "mov "tmp" , %%ecx \n\t"\
|
yading@10
|
108 "shl $17 , "tmp" \n\t"\
|
yading@10
|
109 "sub "low" , "tmp" \n\t"\
|
yading@10
|
110 "sar $31 , "tmp" \n\t" /*lps_mask*/\
|
yading@10
|
111 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\
|
yading@10
|
112 "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\
|
yading@10
|
113 "add %%ecx , "range" \n\t" /*new range*/\
|
yading@10
|
114 "shl $17 , %%ecx \n\t"\
|
yading@10
|
115 "and "tmp" , %%ecx \n\t"\
|
yading@10
|
116 "sub %%ecx , "low" \n\t"\
|
yading@10
|
117 "xor "tmp" , "ret" \n\t"
|
yading@10
|
118 #endif /* HAVE_FAST_CMOV */
|
yading@10
|
119
|
yading@10
|
120 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
|
yading@10
|
121 "movzbl "statep" , "ret" \n\t"\
|
yading@10
|
122 "mov "range" , "tmp" \n\t"\
|
yading@10
|
123 "and $0xC0 , "range" \n\t"\
|
yading@10
|
124 "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
|
yading@10
|
125 "sub "range" , "tmp" \n\t"\
|
yading@10
|
126 BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \
|
yading@10
|
127 "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t"\
|
yading@10
|
128 "shl %%cl , "range" \n\t"\
|
yading@10
|
129 "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\
|
yading@10
|
130 "shl %%cl , "low" \n\t"\
|
yading@10
|
131 "mov "tmpbyte" , "statep" \n\t"\
|
yading@10
|
132 "test "lowword" , "lowword" \n\t"\
|
yading@10
|
133 " jnz 2f \n\t"\
|
yading@10
|
134 "mov "byte" , %%"REG_c" \n\t"\
|
yading@10
|
135 "add"OPSIZE" $2 , "byte" \n\t"\
|
yading@10
|
136 "movzwl (%%"REG_c") , "tmp" \n\t"\
|
yading@10
|
137 "lea -1("low") , %%ecx \n\t"\
|
yading@10
|
138 "xor "low" , %%ecx \n\t"\
|
yading@10
|
139 "shr $15 , %%ecx \n\t"\
|
yading@10
|
140 "bswap "tmp" \n\t"\
|
yading@10
|
141 "shr $15 , "tmp" \n\t"\
|
yading@10
|
142 "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
|
yading@10
|
143 "sub $0xFFFF , "tmp" \n\t"\
|
yading@10
|
144 "neg %%ecx \n\t"\
|
yading@10
|
145 "add $7 , %%ecx \n\t"\
|
yading@10
|
146 "shl %%cl , "tmp" \n\t"\
|
yading@10
|
147 "add "tmp" , "low" \n\t"\
|
yading@10
|
148 "2: \n\t"
|
yading@10
|
149
|
yading@10
|
150 #endif /* BROKEN_RELOCATIONS */
|
yading@10
|
151
|
yading@10
|
152
|
yading@10
|
153 #if HAVE_7REGS && !(defined(__i386) && defined(__clang__) && (__clang_major__<2 || (__clang_major__==2 && __clang_minor__<10)))\
|
yading@10
|
154 && !( !defined(__clang__) && defined(__llvm__) && __GNUC__==4 && __GNUC_MINOR__==2 && __GNUC_PATCHLEVEL__<=1)
|
yading@10
|
155 #define get_cabac_inline get_cabac_inline_x86
|
yading@10
|
156 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
|
yading@10
|
157 uint8_t *const state)
|
yading@10
|
158 {
|
yading@10
|
159 int bit, tmp;
|
yading@10
|
160 #ifdef BROKEN_RELOCATIONS
|
yading@10
|
161 void *tables;
|
yading@10
|
162
|
yading@10
|
163 __asm__ volatile(
|
yading@10
|
164 "lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t"
|
yading@10
|
165 : "=&r"(tables)
|
yading@10
|
166 );
|
yading@10
|
167 #endif
|
yading@10
|
168
|
yading@10
|
169 __asm__ volatile(
|
yading@10
|
170 BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
|
yading@10
|
171 "%2", "%q2", "%3", "%b3",
|
yading@10
|
172 "%c6(%5)", "%c7(%5)",
|
yading@10
|
173 AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
|
yading@10
|
174 AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
|
yading@10
|
175 AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
|
yading@10
|
176 "%8")
|
yading@10
|
177 : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp)
|
yading@10
|
178 : "r"(state), "r"(c),
|
yading@10
|
179 "i"(offsetof(CABACContext, bytestream)),
|
yading@10
|
180 "i"(offsetof(CABACContext, bytestream_end))
|
yading@10
|
181 TABLES_ARG
|
yading@10
|
182 ,"1"(c->low), "2"(c->range)
|
yading@10
|
183 : "%"REG_c, "memory"
|
yading@10
|
184 );
|
yading@10
|
185 return bit & 1;
|
yading@10
|
186 }
|
yading@10
|
187 #endif /* HAVE_7REGS */
|
yading@10
|
188
|
yading@10
|
189 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
|
yading@10
|
190 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
|
yading@10
|
191 {
|
yading@10
|
192 x86_reg tmp;
|
yading@10
|
193 __asm__ volatile(
|
yading@10
|
194 "movl %c6(%2), %k1 \n\t"
|
yading@10
|
195 "movl %c3(%2), %%eax \n\t"
|
yading@10
|
196 "shl $17, %k1 \n\t"
|
yading@10
|
197 "add %%eax, %%eax \n\t"
|
yading@10
|
198 "sub %k1, %%eax \n\t"
|
yading@10
|
199 "cltd \n\t"
|
yading@10
|
200 "and %%edx, %k1 \n\t"
|
yading@10
|
201 "add %k1, %%eax \n\t"
|
yading@10
|
202 "xor %%edx, %%ecx \n\t"
|
yading@10
|
203 "sub %%edx, %%ecx \n\t"
|
yading@10
|
204 "test %%ax, %%ax \n\t"
|
yading@10
|
205 "jnz 1f \n\t"
|
yading@10
|
206 "mov %c4(%2), %1 \n\t"
|
yading@10
|
207 "subl $0xFFFF, %%eax \n\t"
|
yading@10
|
208 "movzwl (%1), %%edx \n\t"
|
yading@10
|
209 "bswap %%edx \n\t"
|
yading@10
|
210 "shrl $15, %%edx \n\t"
|
yading@10
|
211 "add $2, %1 \n\t"
|
yading@10
|
212 "addl %%edx, %%eax \n\t"
|
yading@10
|
213 "mov %1, %c4(%2) \n\t"
|
yading@10
|
214 "1: \n\t"
|
yading@10
|
215 "movl %%eax, %c3(%2) \n\t"
|
yading@10
|
216
|
yading@10
|
217 : "+c"(val), "=&r"(tmp)
|
yading@10
|
218 : "r"(c),
|
yading@10
|
219 "i"(offsetof(CABACContext, low)),
|
yading@10
|
220 "i"(offsetof(CABACContext, bytestream)),
|
yading@10
|
221 "i"(offsetof(CABACContext, bytestream_end)),
|
yading@10
|
222 "i"(offsetof(CABACContext, range))
|
yading@10
|
223 : "%eax", "%edx", "memory"
|
yading@10
|
224 );
|
yading@10
|
225 return val;
|
yading@10
|
226 }
|
yading@10
|
227
|
yading@10
|
228 #endif /* HAVE_INLINE_ASM */
|
yading@10
|
229 #endif /* AVCODEC_X86_CABAC_H */
|